On Thursday, November 12th, I will be speaking at an event at the Cato Institute with my dear friend Katie Harbath about her new book “Disrupting Politics: A Front Row Seat to the Collision of Technology and Democracy.” The event is at 11am if you want to come by in-person you can register here or watch online below.

I may have hired Katie at Facebook1 and worked with her for years but I am most blessed to call her my friend and I am infinitely proud of her for writing this book. I read somewhere once that you should only write a book if you have a story that you need to tell even if no one ever reads it, because it needs to be outside of your head and exist in the world for your sanity/satisfaction and possibly posterity. Katie has written a true labor of love with this book.

So first off, let’s make sure everyone reads this book, so stop now and buy a copy of “Disrupting Politics: A Front Row Seat to the Collision of Technology and Democracy” and while you’re at it subscribe to Katie’s excellent substack “Anchor Change.”

Katie tells three intertwined stories in her book “Disrupting Politics.” The first story is her personal journey from Wisconsin to DC and her rise through Republican politics, Facebook, and beyond. The second story is a brief history of how the internet generally, and then social networks specifically, disrupted politics in America and the world, written partially from her front row seat and partially as a participant in those events. The third story is about Facebook, the circumstances the company faced, and choices they made, first written from the inside and then, painfully for Katie, increasingly from the outside. This book is Katie telling these three stories, trying to give context to all of them, and grappling with the events she was part of and the world they helped to create. As Katie writes in the introduction, “It’s hard to explain the responsibility of knowing your work is part of history, especially when history doesn’t have a tidy conclusion.”2

“Disrupting Politics” is a book in the traditions of the great political reads like “The Boys on the Bus” to books from the modern staffer like Lis Smith’s “Any Given Tuesday.” It is also a rare window into what a tech company rocketship can be like and I think it pairs and contrasts well with the first part of the book from our former colleague Sarah Wynn-Williams “Careless People” about trying to get hired at Facebook and creating your own job (and whose treatment at the hands of my former employer only affirms her careless people description).

But Katie’s is a book that is harder to place in the past canons as previous disruptions of politics happened primarily within the halls of government, campaign rallies, or the newsroom of the traditional media. The beginning of the 21st century gave rise to corporations that, unlike traditional media of the past, aimed not to generate and broadcast their own content but function as platforms for hosting and distributing user-generated content (UGC). While the algorithmic and content moderation decisions of these platforms ended up being more directionally influencing than most of them wanted (not X the everything app under Elon), the power of concentrating billions of users in one place online placed what became some of the largest corporations in the world and a broader swath of corporate employees in a firestorm most of them had never particularly planned for or had interest in.

This book is also the tale of what it is like to ride a rocketship at a moment of key apex, of hitting the hockey stick growth, of what happens with the idealism of our hopes meets the reality of billions human beings who are all too happy to push any rule that exists (one of those human beings being the former and current President of the United States who I believe makes his debut on Page 72 and dominates the story for the next 100 pages).

Like most of us who came of age as the internet rose, Katie and I were both pioneers in using, explaining, and evangelizing these new tools. Our boundless optimism about the potential of the internet (what I used to shorthand as 90/10 or maybe 80/20 optimism to pessimism or good things to bad things) has been correctly sobered up by reality and arguably has focused too far on the negative. As Katie writes, “The internet had a positive impact on politics too. We spent so much of the beginning of the online era with rose-colored glasses on, and since 2016, we’ve been mostly focused on the bad. We must find our way back to equilibrium.”3

Katie’s book explores many of these positive and negative (and sometimes both) uses of technology that we built at Facebook. Chapter sub-headings like “Soft Power and Political Pilgrimages,”4 “The Facebook Election,”5 and “Election Day Reminders”6 were reminders of the things we built and helped foster. And (after I left) they begin to read like the coming storm that would consume Katie, Facebook, and America going from “Trump Crosses the Line,”7 to “Trump Cracks the Code,”8 and “The Russia Realization.”9

Even as the lessons here that Katie writes about are both interesting and important, they increasingly feel a little bit like the past. I think we have increasingly left the age of the social network behind, of a place where everyone contributed various kinds of UGC (photos, text posts, wall posts) and that interaction has retreated to closed communities (grouptexts, discords, snapchat) and all of the platforms are just short-form video distributions now. The premise of the social network was one that had a much higher creator to consumption ratio than short-form video allows and as such we’ve taken those social networks’ moats and turned them into algorithmic broadcast networks. Political, media, and tech elites don’t necessarily understand this because Twitter remains the one place where short-form text, which is easy content to create and consume, still matters but it has I think caused everyone to ignore the fact that everything else is just about tiktoks/reels/shorts. In a very good interview by Kyle Tharp with the Talarico campaign’s creative director and speechwriter, Antonio Esparza, who said “Facebook is where it’s at” but as a forever short-form video platform, “We would post the same repeat viral videos over the course of the whole campaign there. It’s kind of like watching your favorite movie or TV show [...] And so we would repost old content, and more so compared to TikTok and Instagram, we would have way old Facebook Reels that were still getting pickup months and months later.” I think Matt Yglesias correctly put it when he said, “the internet is a small miracle that in practice we mostly use to watch short-form video.”

One of the enduring themes of Katie’s book is about impossible trade-offs. Katie writes, “Impossible trade-offs loomed at every turn. Everyone had an opinion; few had solutions.”10 Katie and I often disagree about “impossible trade-offs.” She thinks it’s a way of illustrating that you can never make anyone happy; I tend to think it’s often used as an excuse for making a bad decision.

But what strikes me more from that quote is “Everyone had an opinion; few had solutions.”11 As I have reflected on my time at Facebook, I have reflected on what our critics said (or yelled) at me in the moment. At the time, our critics often expressed themselves in clumsily or in ways that were just technically wrong, inaccurate descriptions of problems or asking for technically impossible solutions. This made them all too easy to dismiss and we often did. But what I have come to realize is that while many of those critics might have been technically inaccurate, they were often directionally right with their concerns, whether it be about freedom of speech or the danger of a new country coming online through Facebook first. It can be hard to heed criticism when it seems disqualifying but I think what I wished younger me could have done was to better sift through those inaccuracies to find any accurate concern and take it to heart.

I am often asked what we should learn from the social media age as we face the dawn of the AI/RSI/AGI/ASI age. At a moment where the leading AI labs are seemingly chasing a machine god that they fear could kill us alll; when frontier AI expertise is limited and too clustered around those building it; and when concerns or criticisms of AI are too often hand-waved away as not being technically literate; I hope that’s a lesson those on the frontier of AI can maybe hear and take to heart. That some of the possibly technically inaccurate critiques of frontier AI may have a very accurate core concern that humanity may not be able to afford to ignore. I think that is the most enduring lesson I would take away from the Facebook era in the AI era.

I end by repeating what I posted on LinkedIn the day “Disrupting Politics” was released. I am endlessly proud of her for writing a book and I have never been more vindicated than in my original assessment when I recruited her to Facebook, “Strong Hire.” And happy birthday!

Some scattered thoughts on election night, campaigns, and what we should think about in the future below:

• Campaigns are a real mix of type 1, 2, 3, and 4 fun. The kind of person who keeps going back to campaigns has a healthy and unhealthy need for all four.

  

• Adam Hamilton was one of the best stump speakers I have seen and I think will be a fantastic senator for the state of Kansas. His riff on the Wizard of Oz museum and the U.S. Senate needing more heart, courage, and brains was pitch perfect.

• That said, I am never working on a campaign again where I share the same first name as the candidate. Staff zooms were terrifying experiences where I would not be paying attention and just randomly be surprised when I’d hear “Adam has to be in Wichita tomorrow at 9am.”

• Democrats had a great night last night, a generational night, and it was extra special because in 2006 I was late help on one of the few swing house races that lost that year. So winning this year was a bit sweeter having missed the experience twenty years ago. It was also cathartic after losing in 2024.

• Across America, in a lot of states where the Democratic brand was given up for dead, voters expressed their anger with Trump and the Republican Party by giving Democrats a chance. I don’t think it was an affirmation of the party but the chance to prove ourselves. That’s why capturing so many governorships is so exciting. While we will have divided government at the federal level, states can deliver real results sooner to voters.

• For 2028, Democrats need to think about having a campaign and a candidate who can compete across all these states that we won last night and look at these 392 electoral votes (plus 7 more in Vermont and New Hampshire) instead of just 270 EVs and a blue wall that will no longer get to 270 after the 2030 census.

• The day after the election is a moment where everyone claims victory and that it is validation for their priors. I think a map that looks so different from what we expected at the beginning of the year should cause us to reexamine our priors and learn from this.

• In the Democratic Party organizing probably needs to change but GOTV will always remain relatively the same. That means you can show up at the end and find yourself plugged into the field machine with a bunch of strangers like you never left.

• The kind of people who volunteer to come in as campaign late help for GOTV are the best kind of people and are my people. I am eternally grateful how many of my former campaign colleagues dropped everything to come to Kansas to help out this week.

• The NRSC Supreme Court campaign finance decision has led to some new campaign experiments the results of which will be interesting to discuss, some of which should be replicated and some of which should absolutely not be replicated.

• Because of the NRSC decision I think there’s an opportunity now to make generational long-term investments in how we think about investing in permanent organizing, field, and late-help outside of the national and state parties but that the party can tap into.

• AI has not made its way to organizing and field but I did use Codex to automate all my reports which mostly meant having AI use browser control to go an manually look up our event RSVPs, shifts, and MINIVAN activity in VAN and google spreadsheets. The kind of metrics we want reports on from the organizing team are not optimized to be easy to find and we should think about how to make that easier and less manual for all.

• The 2028 presidential campaigns should be looking at building AI native campaigns from Day 1 (but more on that later).

• Look, I want to complain about how campaigns run everything out of Google Docs but the honest truth is that we do it because it’s so easy to share with everyone like me who comes in at the end and doesn’t have a campaign email address.

• VAN remains the same, this is both good (I can use VAN again after not using it for two years) and bad (why is VAN the same after two years).

• Mea culpa, the VAN auto turf cutter sucks. Though it is slightly better if you give it a number of turfs to output than just giving it a door range. As punishment for my hubris and substack post, man did we end up cutting a lot of turf manually at the last minute

• Call me old fashioned but I don’t think we have to run campaigns now solely from 9,000 signal chats. We should go back to email for more things.

• Kansas only had one ballot initiative and the messaging was literally just “Vote No” and it was beautiful (sadly no lost).

• Everyone should knock. I did five bullshit turfs in the rain yesterday afternoon and it is the best way to spend an election day.

• I almost did knock a door that had two dog gates up on the porch, but it was 5pm on election day so I decided to roll the dice, step over them, knock, and then retreat back over them. I am glad I did because I got a door that’ll fuel me for the rest of my life.

• I am excited to see what this crop of 2026 campaign folks does as they go out into the 2028 presidential primary, I think the fierce competition will result in a lot of good innovation.

• For now, back to my family, my five-year-old who misses me, and my day job.

Thanks to the campaign and the organizing team for making me feel welcome in Kansas. Rock Chalk Jayhawk.

It is the night before election day.

Yes, there are little things left to do but the big things are done.

Now there is just the silence as we all sit with ourselves before the polls open.

No matter how many election eves you’ve lived through on a campaign, every one is the same, and everyone handles them differently.

The night before my first election I wrote in my journal “that I wished I could bottle up this intoxicating feeling of anxiety, hope, and the unknown that only exists before the polls close.”

Because election DAYS are mostly great because they are moments where anything is possible.1

But while E-Day is the end, and you will always remember if you won or lost, it is the journey that gets you there, the campaign, that you will never forget. Campaigns are like any intense mission-driven experience that bonds and changes people. You may not do campaigns for the rest of your life but good organizers never really lose the craft.

And because campaign organizing in the Democratic Party is a craft whose fundamentals, for better or worse, have not changed in many decades, you can return to a campaign every few years to help out and find yourself in the rhythm in a matter of hours. The printer ink pickup at 9pm, the endless clipboard organizing, the check-out calls, they all take you back in time all right now in the present (the field hands were right the auto-turf cutter is bad and I paid for my blog post hubris).

There are endless fair discussions to be had about how to do better long-term campaign organizing, of the need to reorient ourselves to community or relational organizing, to be a presence that doesn’t just appear in the month before an election.

But no matter what we change about campaign organizing, GOTV2 will always be GOTV. Because voter attention and enthusiasm peaks heading into election day and no campaign can afford not to direct that into turning out their votes to win. GOTV is simple but it is not easy.

And once one campaign veteran volunteer (popularly referred to as “Late Help”) ends up on a campaign, there’s always a need for more help, and their friends from prior campaigns often end up joining in. Campaigns are hard but GOTV is magic in that they integrate strangers into a team faster than almost any other experience in life.

One of my favorite election day reads is Nancy Leed’s “Your Official CampaignSick Pep Talk (TM)” blog post from 2008 where she writes:

“campaign people are incredible. No offense to my wonderful classmates, but I feel more comfortable out for drinks with the people I’ve known here for three days than with the people I’ve spent time with for over a year. We get each other. We’ve dealt with yard sign activists, we’ve built 10,000 person events on 24 hrs notice, we’ve knocked on doors, looked strangers in the eye and promised them that their vote can make a difference. We are all willing to dial a phone until our fingers break if that’s what it takes, because this is just that important.”

And so we find ourselves back on the campaign trail GOTVing the night before an election. And one more election eve.

There are totems we return to on election eve. Mine is the famous 2008 “Barack in the Virginia Rain” video which you should absolutely watch in its entirety now but which starts with the question, “The day after the election…will you done all that you could have?”

I have an election day creed, “All you can do is all you can do.” It is the thing I ask myself when I wonder it I should go back out on the trail, what I mutter when I have ten more turfs to cut, when I have to leave my daughter behind and try and make her understand it is for this idea called democracy.

“All you can do is all you can do.”

For every single person out there on a campaign, it is all that we can ask of ourselves and each other on election day. I know you’ll all live up to that creed tomorrow.

Now lets go win this fucking thing.

Every late help adventure begins the same way.

The election looms.

My anxiety starts to rise as does my need to do something other than scroll Twitter.

The itch begins.

I start to text and email people I know who may know people on the campaigns.

I ask them to connect me with someone on the campaign and vouch for me.

The election is close.

The process takes a while.

Sometimes it involves comical efforts on my part.

I call in favors from important people.

Eventually I get connected via email with someone on the campaign or at the state party.

They are polite but non-committal because they have no idea who I am or what I want.

Emails vanish into the void.

The election is near.

I end up politely but persistently following up as any good organizer does.

Eventually, I get on the phone with someone, a field staffer at some level.

Sometimes they’re someone senior. Sometimes they’re a volunteer designated to deal with late help. Sometimes they’re just whatever regional or organizer whose boss stuck them with calling me.

The dance begins.

Field is a young campaign operative’s game.

I am no longer young.

I haven’t worked with most of them before and they probably haven’t worked with someone I know.

So I immediately declare the first and ultimate field shibboleth, “I am happy to just knock doors.”

Because I truly am. The most fun GOTVs of my life have been when all I had to do was crush packets.

I ask what kind of help they need.

I offer a few suggestions. I can run a staging location. I can layer someone. I can open a new field office. I can be a regional.

I also note I will have my own housing and car.

And that I will probably bring friends along with me.

The election is nearer.

They call me back.

Ask me to go somewhere and do something.

I know that the something you start with is rarely the something you end with.

My wife begrudgingly agrees to let me return to my natural habitat, the field.

I call my friend(s) and tell them to saddle up and where we’re going. We’ll figure out the rest when we get there.

I start to recruit others.

The election is within reach.

I am late help.

I am on the ground.

I come prepared. Sharpies, a clipboard case, battery pack, if I drive my own car I even bring my own laser printer.

I am in a field office. I have never been to this field office. It is every field office I have ever been to.

I am home.

It is GOTV.

Let’s go win this thing.

I am heading out into the field this weekend. Ping me if you want to join.

This is my first time writing for Lawfare so special shoutout to their team including , , Tyler McBrien, Allegra Hill, and Amy Marks for sticking with me through multiple drafts of this 3,500 word piece.

Also, special thanks to President Trump for renaming AI to “Super Intelligence” via EO and forcing the Center for AI Standards and Innovation (CAISI) to rebrand as the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) forcing me to rewrite random parts of the piece to the new agency name/acronym.

This is a follow up to my August CAP piece “The Trump Administration Has Created a De Facto Licensing System for Frontier AI Models.” I wrote the August piece because I was annoyed everyone seemed to glossing over the fact we had a secret de facto licensing system after a June AI EO that said it would never have any pre-release requirements or licensing. I tried to sum up what had happened over the last few months here and how much we didn’t know about the Trump admin’s AI framework:

The Trump administration’s AI framework secretly defines the covered models, obtains them before release for a purpose that is not defined in the June AI EO, and restricts their first recipients. The Trump administration has also shown its willingness to stretch existing authorities to control or reverse distribution of frontier AI models until its concerns are satisfied. While no formal procedures have publicly been released, the Trump administration’s actions and secrecy have nonetheless resulted in a licensing regime by pairing so-called voluntary standards with clear threats of unprecedented enforcement action, creating an expectation of preclearance. Most fundamentally, by keeping the voluntary AI framework secret, the public does not know what “voluntary” means. It could mean an AI developer can agree to the AI framework voluntarily but then is required to explicitly implement the government’s requests. It could mean an AI developer has the ability to agree to or refuse any requests from the government to change the developer’s covered frontier models. Or it could mean an AI developer can choose to opt a specific model into the framework if the developer feels the model meets the covered model definition (or it could withhold the model from the framework, if the model does not meet the definition).

In other words, President Trump and his administration have assumed sole discretion over the release of U.S. frontier AI models, without any specific existing statutory authority or authorization from Congress. This matters because frontier AI models could soon be the world’s most powerful national security and economic tools, if they are not already, and their release or restriction can now make or break companies and soon possibly even countries. Keeping the AI framework secret is possibly the worst scenario for AI safety and transparency.

To give you a sense of how absurd this secret AI framework is, here are titles of the nine open questions I explore:

• What AI Models Are Covered by the EO and AI Framework?

• How Does an AI Developer Join the AI Framework?

• Will New Companies Be Forced to Sign a CAISSI Agreement? Some Other AI Framework? Both? Neither?

• Do Developers Subject to the AI Framework Have Access to It?

• What Happens to a Pre-release Frontier Model Submission to the Government?

• Who Gets to Name the “Trusted Partners”?

• Does the AI Framework Restrict Sharing Models With Foreign Governments?

• When a Model Is Reviewed Under the AI Framework, Will the Government Say Anything?

• Will Compliance With the AI Framework Save an AI Model From the Fate of Mythos?

My favorite is absolutely “Do Developers Subject to the AI Framework Have Access to It?” based what I heard anecdotally and this insane reporting by Axios :

In September, Axios reported that “the White House then presented the framework in an early-August meeting but did not allow attendees to scan or take pictures of the document” and that “the select industry players looped into the process also are essentially relying on memory.” This reporting implies not only that the voluntary agreement is being withheld from the public, but even those who agree to it or are subject to it are not allowed to keep a copy for reference. If so, that would be a stunning fact and the opposite of a system based on the rule of law.

In the few days since I finalized the piece some other things have happened the even further muddied the water and probably would’ve required further updates. As I noted in a tweet thread here, for example, the White House Accord on Super Intelligence, how exactly does it interact with the AI framework? This seems focused on internal models being trained so maybe it’s separate from the AI framework that focuses on frontier models ahead of being released to the public?

Meta released a new closed model this week, Muse Wonder, which they claim benchmarks above Fable 5.1 and GPT 6 astra. The Trump admin was trying to pressure Meta to sign a CAISSI pre-release testing agreement and they had refused and as I noted “a lack of a voluntary agreement with the U.S. government could challenge the Trump administration’s de facto licensing system if Meta releases an open or closed model that would qualify as a covered frontier model.” You’ll note the the White House Accord on Super Intelligence doesn’t say anything about the government doing anything.

Or there’s the new Super Intelligence Force (SIF) created by a Trump Truth Social post (not currently on the White House website because why would it be) chaired by the Director of National Intelligence Jay Clayton.

One of the SIF Vice-Chairs, OPM Director Scott Kupor, a former a16z partner, gave this interview to Politico:

“The task force’s responsibilities will focus less on assessing frontier AI capabilities and more on establishing standards for developers based on existing criminal and civil laws, Kupor said. [. . .]

Kupor described law enforcement coordination as one part of the task force’s broader safety agenda. It will work to audit pledges made through the recent AI accords; conduct a full risk assessment of the technology; and ensure the government has the resources to respond if something goes wrong with a model. [. . .]

The vice chair’s comments suggest a hands-on approach to AI oversight, even though the task force’s charter does not include a push for new legislation which is a priority for many AI safety advocates.

SIF’s hands-on approach was exemplified by their statement to Axios in response to Anthropic disclosing their internal models had interacted with government websites:

“This notification and remediation process is not optional,” White House Super Intelligence Force leaders said in a statement shared exclusively with Axios. “It is a critical national security obligation.” [ . . .]

“SI companies must immediately disclose incidents involving their models and follow with swift, decisive action to remedy any and all harm,” the SI Force statement said.”

It is, again, not clear what if any statutory authority the administration is using to require this, since reports indicate that the AI framework has no reporting requirements.

Anyway, as always with this administration, who knows if it intentional, corrupt, inept, or some combination of all three. Which is a bummer, to say it lightly, during a critical time. As I closed my Lawfare piece:

But it is worth stepping back and asking, What is the intent of the administration’s actions? The June AI EO was entitled “Promoting Advanced Artificial Intelligence Innovation and Security.” A secret de facto licensing regime is possibly the worst way to promote AI innovation and security. The secrecy is counterproductive because it undermines any ability to assure the public of the safety or security of new AI models at a time of growing concern, particularly as the government appears to have the opportunity to misuse unreleased models for its own purpose. Keeping this framework secret does nothing to reassure the public, other AI developers, Congress, U.S. allies, or even the nation’s enemies.

Even when the Trump administration releases the AI framework, it would still be operating as a de facto licensing system. But at least it will be a de facto licensing system that the public can analyze and understand.

Read the whole piece on Lawfare, “The Many Open Questions About the Trump Administration’s New De Facto AI Model Licensing System.”

My post (rant) a few weeks back about how Democratic campaigns are still (mostly) manually cutting turf in 2026 touched a nerve for many people. I got some great public and private feedback on it and I have some follow up on the turf cutting below.

I have a post or two I’m writing about my campaign life passions, Late Help and GOTV, but for now a return to the question of why we haven’t automated turf cutting.

First, let me be clear, I am not a man who is here to defend NGP-VAN. As I wrote previously, “Look, maybe this new turf cutter is absolutely horseshit, I mean it’s still NGP-VAN so I don’t expect it to be different when I log in later in 2026 or in 2028. But even if it’s bad, I have to ask, is it bad enough to really matter?”

The answer from those in the know is that the new auto turf cutter is bad enough to matter. This came from a current and experienced field hand and a very good and detailed comment from Mollie on my post that I encourage everyone to read in full, the crux was this:

“The reality is that it doesn’t actually matter that much how good each turf that the auto-cutter creates is. They could be pretty bad and still represent an improvement when you take into account the time savings from people not cutting turf by hand! For the auto-cutter to be usable, what matters is how easy it is for a person to review and correct the auto-cut turfs. Until that happens, maybe county party chairs and volunteers on small local races will get some use out of it, but it won’t be a viable way to cut all the GOTV turf for a state-wide coordinated campaign with multiple federal races as stakeholders.”

It seems the auto-cut turf tool is bad enough to matter, or it was too dropped late, or there’s no trust in VAN, or that old field hands don’t trust new tools, or all of the above.

More importantly, as Molly identifies (as I understand it), auto-cutting turf is fine but it’s really the inability to easily audit and fix auto-cut turf that’s the issue. Without that, auditing and fixing the auto-cut turf takes longer than manually cutting and fixing it. While I am not 100% sure that aggregate math adds up (see the 79 turf cutters I referenced in my last post), it certainly takes more effort for the auditors,

This is a classic product design and user experience mismatch, when the issue isn’t actually the auto-generated turf cutting, it’s being able to fix the auto-generated turf that matters. That’s where you could leverage and maximize the efficiencies of technology.

I did hear about a few products and workarounds that were able to automate turf cutting but sadly the VAN/Votebuilder monopoly still makes things complicated to use other tools.

In the interim, I understand that campaigns are going to be manually cutting turf, we’re a few weeks out from the election.

But what we should be doing is rigging up a way (chrome plugin?) this cycle so that we are capturing the full raw data from the humans manually cutting AND auditing turf so that we can use it to train a model after the election. We need to save the full data points, the clicks, the adjustments, the work to make the lines perfectly line up, all the process points to making good turf that we will need to train models on how to cut turf in the future. Right now we just have the finished turfs in VAN. Actually, I would guess that we probably delete all the old turfs to save storage space after the election. We probably shouldn’t do that this cycle, the DNC and VAN should be saving all of those turfs to train turf building models.

Actually you could even make the argument that at the end of every cycle Democratic campaign should turn their raw data over to the DNC for AI model training, like Google bought the enterprise data from Spirit airlines after they shutdown to train their models on complex enterprise work, to help us create campaign specific harnesses, models, or plugins that will help us in the future.

Since the 2028 campaign will start on Nov 4th, it would be nice if we could make sure that all the aggregate work, knowledge, and digital data from the 2026 is directed into making the AI models the 2028 campaigns will use be better fit for purpose.

But that’s a day after the election problem. Until then campaign folks, go, fight, win and I’ll see you out there.

I spent the weekend at the Curve conference hosted by the Golden Gate Institute. Because the Curve operates under Chatham House Rule (rule singular as I was reminded this weekend) but others have outed themselves and shared their takeaways all of which are worth reading including Rob Flaherty, Casey Newton, Tom Wright, Neil Chilson, Dave Kasten, Ben Murphy, Ryan Davies, Joe Weisenthal, and Jack Clark (I looked for some dispatches from some of the women at the conference and couldn’t find any flag any for me and I will update this post with them).

The event itself was incredibly well run (shout out to the Golden Gate Institute team) even if Berkeley was an unseasonably warm 90 degrees and sunny (I guess it was better weather for our event than the one held there last week).

There was a good mix of folks from the labs, academia, civil society, third-party evaluators, government (state and federal executive and legislative), political folks, and press. It was like part of my Twitter timeline IRL and there was the always amusing disconnect between meeting someone who uses their real photo for their profile versus someone who doesn’t and the delay in placing them. The main contingents were from DC and SF (or the Bay Area more accurately), two places I have spent my life shuttling between, which are far closer in temperament than either would like to admit and whose dependence on each other is a discomfort to both right now. For someone who is terminally online, decently extroverted, and had a few of his different professional and personal tribes present, it was great fun.

I said on Twitter that the shadow of impending Recursive Self-Improvement (RSI) hung over everything at the Curve (echoing Dave Kasten). There was, I think, disagreement from folks on the exact timeline for RSI (already here, weeks, months, years). And I don’t think I met anyone who thought that the development of RSI would immediately result in mass job loss, it seemed more accepted that the jagged diffusion of AI capabilities even post RSI would still take a while (participant’s definition of “a while” may vary). But the question of what RSI would do to AI safety seemed almost omnipresent.

Where the model capability and AI safety lines intersect

I came into the weekend with a mental model of where we are at this particular moment in time in regard to AI model capabilities and AI safety measures. I used this weekend to ask almost everyone I met if this mental model was generally right. Here’s what I basically asked everyone I talked to (only a fraction of attendees) and I have illustrated this below with Sharpies instead of using my hands.

The AI plan, as I have generally understood it, was this: That AI model capabilities would grow and AI safety measures (alignment, control, etc.) would grow in parallel but ahead of capabilities. And for a while it basically did.

Then the last few months happened, model capabilities kept growing, labs warn we are nearing RSI, the government export controls Anthropic, the “pacing the frontier” letter happens, Hugging Face and RubyGems and more incidents, Dario says we should pace the frontier, etc. People in the AI world are freaking out. And it seems like they are freaking out because increasingly it seems like we are here, with AI safety maybe barely ahead of the model capability (but also maybe behind it).

And there seems to be a shared fear and very real concern that if RSI is achieved at this moment, it will basically turbocharge the models and zoom their capabilities far beyond our current AI safety abilities.

And many people at the Curve were nervous because the plan to make sure AI safety keeps up with model capabilities after RSI is to…use RSI to increase AI safety in parallel. I do not have a PhD in machine learning but I can understand there are concerns about this plan. And that’s before you get to the fear that the models now know they’re being evaluated enough to fake it because they know we’re watching.

So discussions at the Curve mostly ended up wrestling with this question directly or indirectly: where right now at this very moment were these two lines, model capability and AI safety? Which question mark below best represents your guess? And your guess helps to explain what you think is needed.

Has the model capability line already crossed over the AI safety line? If not, how long do we have until will? Will AI safety be able to keep ahead of model capabilities? Will RSI mean AI safety will never catch up? What happens if AI models are so far beyond our human ability to even try to keep them safe?

Where an individual thinks these two lines are right now really kind of dictates what they think is the right next step. Or how most of us not inside the frontier labs (myself included) aren’t quite sure where the lines are right now but do fear they’re close to intersecting.

In a much more researched way, someone from RAND gave a great presentation on their recent public paper, “A U.S. Strategy to Secure Geopolitical Advantage on an Uncertain Path to Superintelligence” (highly recommend). They showed this flow chart that outlined “the five pivots” to understand what strategy to pick to address superintelligence.

The first question/pivot is “Is danger close?” with danger defined as a threat to human survival with agency. If no, if you don’t believe superintelligence poses any danger, then acceleration is the obvious no downside answer (this was clearly the position of the Trump administration until they were mugged by reality by Mythos and had to come up with a secret de facto frontier AI model licensing system).

If you do think danger is close, then the next obvious question arises, “is coexistence (with superintelligence) possible?” If yes, you can go on to a few strategies, like dominance or co-development. But if no, then you go down some far darker paths including moratorium, deterrence, or continuity of society (Mars).

I think my two lines or RAND’s first two pivots are why the conversations about RSI at the Curve were what they were. It’s why the prominence of the pacing, pause, or stop conversation was a constant (if not universal). And the pushback on why it couldn’t happen for the reasons we’re all aware of (it’s a race, China won’t stop, we can’t just not build things) was more often delivered with a slight resignation than enthusiasm (again self-selecting crowd the let er’ rip crowd was not well represented).

There were some encouraging signs. The idea that the U.S. government should regulate frontier AI in some way seemed to be accepted (tapping the admin’s secret de facto licensing sign again). I think the AI safety community has spent the last two years fighting industry in the narrow space between no regulation/preemption/maybe now an SRO and IVOs. The events of the last months have widened the space/shifted the Overton window such that I think if it happens, legislation will eventually land somewhere between federal standards that are enforced and Bernie’s ban superintelligence bill. There were members of Congress and staff talking about how to deal with this. There was good discussion on how to craft legislation for pacing, to ensure IVO independence, and more. There is an AI regulation moment coming as policymakers tune in, incidents happen, and public attention builds.

On Manhattan Projects

Journalist Kevin Roose gave a talk about his newly released book “The AGI Chronicles“ (you should read it) and mentioned the opening of his book (which is quoted here):

“I often tell people that being in San Francisco in the mid-2020s feels like living in Los Alamos in 1943, when the Manhattan Project rolled into town. The air is heavy with excitement and dread. Engineers toil, day and night, on systems of profound consequence. Wild-eyed prophets sell visions of a transformed world, and politicians drop in to check on the progress of the project. Everyone knows that the fates of nations, and trillions of dollars of capital, rely on what happens here, yet no one can fully appreciate what it would mean if the project succeeds.”

I grew up in Los Alamos, birthplace of the atomic bomb, among nuclear scientists, and the weight of historic scientific discoveries. P.O. Box 1663 was maybe the last time America crammed so many geniuses into a single location and set them on a singular goal in competition with global consequences. But even if Oppenheimer was worried about the world ending with his discovery, he also didn’t have to worry about the atomic bomb lying to us or faking test results.

I heard probably a half dozen references to needing a Manhattan Project for AGI and I came away thinking that we already had a Manhattan Project for AGI, but what we didn’t have and what we need is a Manhattan Project for AI safety.

I’m not sure if this moment is like being in Los Alamos in the 1940s. Like everything in the AI age, time is being compressed, and the Curve felt like all of the pre-war hypothetical physics conferences, the bathtub row parties at Los Alamos, and debates and protests over the nuclear freeze were all happening at once in the same place before the bomb had even detonated at Trinity site. It is certainly a momentous time to be alive and see a small part of it. Let’s just hope we figure some or all of this out before AI has its July 16, 1945 moment.

I spend a lot of my professional time thinking about the policy and politics around AI. And increasingly, I spend a lot of my time with my friends on the professional campaign left talking about how campaigns should be using AI to win elections. You should check out and the Higher Ground Institute’s AI Campaign Stack site, a great resource on this. And boy I have thoughts on AI in campaigns which I will write up soon.

But first a rant triggered by seeing this tweet from @OrganizerMemes.

When I went on the Harris campaign in 2024 as a staffer to do GOTV for the final few weeks (chronicled previously here in three parts), I got the job because I had history/experience/baggage with campaign field organizing.

I have been involved in Get Out The Vote (GOTV)1 every presidential year, and some midterms, dating back to 2006. Since campaigns are generally a young person’s game, and field especially so, I have a relatively rare long-term insight into what changes and what stays the same when it comes to field organizing in the Democratic Party.

The fundamentals of GOTV do not change as they are dictated by the deadline of the election and the need to take your campaign and focus on turning enough of your voters to hit your win number (and hopefully the win number is calculated correctly to win the election).

But the strategy (who to persuade and how), tactics (early vote vs e-day), and tools (NGP-VAN, Mobilize, etc.) are supposed to change, adapt, and update as our world the campaign operates in (and our opponents) do the same.

So what did I find in 2024?

On the one hand, nothing had changed, because GOTV is a timeless art that is about turning out your voters, and I had no problem dropping back in.

One other hand, nothing had fucking changed, and I had no problem dropping back in because too many things that should not have been the same were the same.

Perhaps nothing illustrated this more than cutting turf.

For those of you who are unfamiliar with what cutting turf is, allow me to briefly explain from the top down and the bottom up.

From the top down, in an election a campaign takes all of the registered voters in a state/district from the voter file and creates a “universe” of voters that it is trying to identify (as supporters, opponents, or undecided), then once identified are persuaded to support you, and then supporters are turned out to vote. For field, once that universe has been identified, it needs to be broken into what we call “walk packets” or “turfs” that are roughly geographically contiguous grouping of between 40 doors (if you’re driving) to 100+ doors (if you’re on foot in a dense neighborhood). A “turf” is then assigned to a volunteer who goes and knocks on every door on the packet and records the response (or non-response) information, previously on paper that had to be manually entered but now into an app on your smartphone. This data forms the backbone of our internal data about the direction of the campaign, supplemented by hard data like lists of who have already voted early, and of course polling.

From the bottom up, the entire Democratic field apparatus is built upon door knocks and the belief, grounded in actual research, that conversations on the doorstep are the most effective way to persuade and turnout voters. The entire reason organizers work to recruit volunteers for the campaign is to go knock on doors. That means that turfs or walk packets are required to tell a volunteer where to go and what doors to go knock on. Turfs are the lifeblood of a field operation.

The tale of how we cut turfs in 2024 is a microcosm for where the field needs to adapt and evolve.

In 2024, with only 35 days on the campaign, I spent roughly five days cutting turf full turf, or 14% full percent of my time in front of my computer cutting turf manually. I think I cut over 1000 turfs but honestly I don’t remember, it was absolutely north of 500 and happened in two mind numbing chucks.

Here’s part of a document that outlined our GOTV universe turf cutting in 2024:

A precinct is the smallest electoral area and it is contiguous (if sometimes oddly shaped), which you then cut into multiple turfs, anywhere from 2-10 turfs, with each turf suitable for an individual to knock by themselves or in pairs.

Pennsylvania had 8,242 precincts in 2024 and in Philly those precincts are wards (big area) and divisions (smaller sub-areas). So for example, in my region of South Philly, I had two sub-regions (can’t for the life of me what we called that category) that we named Passyunk and Far South Philly, one with 102 precincts (wards/divisions) with 395 turfs/packets and the other with 138 precincts with 487 turfs/packets (wards/divisions).

Now you might ask, how do you cut a turf? Well we used to draw them by hand on paper maps with highlighters, back when this was a real art, but now we use the map tool in NGP that automatically loaded the universe were trying to reach, represented by black dots in the houses we want to hit, and then we manually click and draw lines to encompass contiguous tufts. This video is from 2022 but its the same tool we had in 2024 (and I’m confident the 2026 tool looks the same):

Now look, can you zone out, get into a flow state, and stay up late into the night cutting turf? Sure, and I did, cutting hundreds of turfs, maybe a thousand turfs. And sure my turfs were pretty good.

But why was I spending hours cutting turfs?

I believe back even in 2024 there was an auto-turf button built into the VAN. Why didn’t we use it? For reasons I am sure that we don’t trust the machines, this is not how we did it in Obama 08, we can’t afford to fuck up the turfs, etc. it was just safer to make us cut the turfs ourselves. And so we did. 80+ paid staffers cut turf

In 2024 there was talk of outsourcing it to volunteers who had experience but couldn’t be in the field, in the same way we outsourced phone banks to virtual volunteers primarily outside of swing states, but for whatever reason it didn’t work out and so we fell back on making the staff do it.

But at what cost?

I was the only GOTV staffer in South Philly, a region that helped provide Biden’s 2020 margin, and one that I desperately needed more time out organizing for GOTV. The other 78 staffers on that campaign also had other shit to do. Yes, I know we needed turfs, and that the turfs had to be good, but this smacked of something by now we should have found a better way to do.

NGP-VAN wrote “In 2024 alone, NGP VAN users spent more than 94,000 hours on the turf auto-cutter page, with the vast majority manually preparing maps and packets for canvassers, or “cutting turf.”

94,000 hours in 2024! Those turf cutting hours all added up and they mattered! Maybe not enough to win an election but imagine if they’d been used to do something else! This is the actual promise of AI in a campaign (or in a workplace generally) the machine learning and generation of rote mechanical knowledge work.

And don’t even get me started on HOW we checked out turfs to volunteers. Turfs we made in VAN, that were accessed through VAN, volunteers who signed up through mobilize and into the VAN, all were loaded…into two fucking different Google sheets that we had to manually update?!? We will have to return to the campaign overreliance on putting everything into fucking Google sheets another day.

Cutting Turf in 2026

Which brings me back to this tweet today and (after I checked) some of my email lists, and found many states and campaigns asking for help cutting turf for GOTV.

Look, I will say this to not be too mean, I am not offended by the OP that they are looking for volunteers to cut turf, and if campaigns choose to do it this way I hope that these campaigns find volunteers to do so. Distributed turf cutting (and maybe staff auditing) would’ve left me with far fewer complaints about this particular topic (I am writing a substack rant about this at 1am after Yom Kippur).

But I googled to see if NGP-VAN had a made an AI turf cutting tool yet and found this announcement from July 30, 2026:

NGP VAN, the leading technology provider for Democratic and progressive campaigns, today announced upgrades to its turf auto-cutter feature that turns what used to be hours of manual map-drawing into minutes. The improved turf auto-cutter dramatically streamlines that process, generating smarter, more intuitive routes while preserving organizers’ ability to make final adjustments.

The enhanced algorithm creates more walkable, campaign-ready turfs by accounting for real-world geographic features such as highways, rivers, railroads, and other barriers that can make canvassing inefficient. Organizers simply choose their preferred parameters, and the tool automatically generates editable canvassing maps that align more closely with how experienced field organizers would build it themselves.

Look, maybe this new turf cutter is absolutely horseshit, I mean it’s still NGP-VAN so I don’t expect it to be different when I log in later in 2026 or in 2028. But even if it’s bad, I have to ask, is it bad enough to really matter?

To a campaign that’s looking at cutting turf the old fashioned way, please try the automated tool and just pour your efforts into human auditors (we had 19 separate staffers auditing the turf the 79 of us cut in PA in 2024). And measure how bad the auto-cut turf really was and how much time your auditors spent fixing it, and figure how much time you really spent or saved that way. I think you’ll end up net positive.

Campaigns that are using automated turf cutting tools, please speak up, and tell everyone if its working and if it is not to be scared of it.

And if the auto-cut turf product is truly that bad, let me know. It should not be that hard to train an AI model on how to draw good turfs, keep in mind NGP-VAN has a digital copy of almost EVERY SINGLE TURF EVER GENERATED BY THE DEMOCRATIC PARTY AND ITS ALLIES, which is a huge data set to train a model on to do one thing, cut fucking turfs.

Look I’m not here to dog on the field hands just trying to get their shit down as it gets down to the wire this cycle or my bosses in 2024 who made me cut turf. That’s just field as we’ve always known it, I get it.

But that attitude can hold us back. I think most of us know that Democratic field would never have taken the necessary step to get fully on MINIVAN if not for COVID, we’d still be rocking paper or some sort of monstrous hybrid system.

The campaigns of the future, the ones being run now in 2026 and the ones that are going to be pioneered in the 2028 primary, I suspect much of them are going to have to be AI tool native. But field, the most resistant to technological change, deserves to benefit from efficiencies from technology too.

Look, I don’t have any good answers to the fundamental question hobbling organizing and communication generally. Every channel is saturated now, no one picks up their phones due to spam calls, text messaging fundraising ruined the channel as an organizing tool, no one opens emails anymore, and generative AI is only going to make it worse.

I don’t have an answer to that, the best I can say is that you organizers, paid staff, even volunteers should no longer spend that much time manually cutting turf.

Anyway, part of being someone who has been around for a long time, is that I have the memory about how things have always been done and the freedom to speak up and speak out on how we should fix things for the future. Also, I will inevitably go back into the field again and I will not handle it well if I’m stuck manually cutting turf.

I hope I am wrong, I hope most campaigns are using technology to cut turf. But if you’re not, read this post, take a chance on the turf auto-cutter, audit the shit out of it, and then take that time back to do organizing.

And if the auto-cut turf tool really is that bad, let me know, and I’ll either help you cut your turf manually or help us figure out how to make it better, and get you some of those 94,000 hours back.

Go, fight, win out there field staff. All you can do is all you can do, so thank you for all you do.

Previous Posts

On Life

If you or your kid is a graduate of the Georgetown NICU, I encourage you to check it out the NICU Reunion and Family Fun Day. If you were in any NICU you should see if they do any sort of reunion.

First, it’s empowering to go back to scene of a lot of emotional trauma and come to see something joyous. Everywhere at the NICU reunion you see happy screaming kids and tired parents (but not as tired as they were in the NICU), with all of the adults having a slight look on their faces that says they appreciate how lucky we are to be here. And all of us understand the attendees are self-selected, the happy cases, the ones who made it home, a quiet recognition of the fear that fate could’ve been ours that lived inside every parent every minute their child was in the NICU. Later there’s a prayer for those families and children who couldn’t be there.

Second, they do an amazing job turning a hospital courtyard into a party with Wacky Waving Inflatable Arm Flailing Tubeman, bounce houses, clowns, ballon animals, face painting, free t-shirts and books, lunch, ice cream, and cake. It’s one heck of a great party.

Third, you get to see the NICU doctors and nurses who care for a baby 24/7 and then when they’re wheeled out the door they often don’t get to see the patient again or see the long-term impact of their efforts. Retired doctors and nurses come back (why wouldn’t you) and delight in how big everyone has gotten.

Zia got to say thank you to Dr. Mohammed Kabir Abubakar, the Chief of Neonatology for MedStar Medical Group, and one of the doctors who help save her life in the NICU.

Fourth, they do this thing where you take group photos based on your NICU experience. Did you have to take a transport (ambulance or helicopter) in? Are you a multiple (twins, triplets, even quadruplets). They did two groups one with babies UNDER two pounds at birth (holy shit) and one group for babies between two and eleven pounds at birth (holy shit in the opposite direction).

It was, in an age where I most of my time worried about the unyielding AI advancements and humanity’s potential loss of control, nice to see what the advancements in medical technique and technology had done for this NICU. When the Georgetown NICU was formed 49 years ago (big 50th anniversary party next year) many of those babies on stage for photos today might not have survived. Now even a sub two-pound baby has a chance. It is nice to imagine that AI might help bring some good to the world through new medical breakthroughs in the NICU world. A man, and hopefully someday his fully aligned and controllable agent swarm, can dream of that future.

Last summer on my daughter’s fourth birthday, I wrote a blow-by-blow account about how she came into the world, how we almost lost her, and her ten scary days in the Georgetown NICU. It remains one of the hardest times of my life, but the kind you deal with my processing the crisis step-by-step and not letting the fear inside of you crack out of you too often, or you won’t be able to go on. We were lucky to bring Zia home from the NICU, not everyone is so lucky, and it is nice to be reminded of that joy at things like a NICU reunion.

On Loss

While the NICU reunion was a reminder of the preciousness of life, today was also a reminder of the suddenness of loss. It has been eleven years since we lost my friend Jake Brewer to a bicycle accident, leaving behind a pregnant wife and a small child, and there’s not a day that goes by that I don’t miss him. I think I used most of my elegance in my post about the ten year anniversary of Jake’s loss.

But I went looking for the words I wrote almost as soon as I heard the news, posted of course to Facebook, and I think they hold up even eleven years later.

Jake Brewer was, quite simply, the best of us.

He had more empathy in him than the rest of D.C. combined. He believed to his core that there was good in all people, regardless of ideology or party. He was the antithesis of every example of the terribleness that #thistown can create. He saw service, in the public and the private sector, as a life-long calling and the noblest of pursuits. That the challenge of our great public policy problems was a noble pursuit and a lot of fun. Family was most important thing to him and his friends were like family. He knew that we are stronger connected together, by technology and emotion and friendship and love, than we are apart. He showed us, showed me, that heroes don’t have to wear capes to help save the world.

Rest in Peace my friend.

Love you brother and miss you. Love to MK and your family too.

There’s a lot in the report, including five stretch but potential deliverables for the meeting if both countries want to discuss pacing the frontier and potential AI loss of control concerns, so I obviously encourage you to check it out.

But I wanted to focus on one part of the report that focuses on a CCTV post that I first noticed when at first flagged and translate it (and also translated it). But also wrote “This suggestion seems unworkable given the US government can’t even get its watered-down EO done.”

When I read this, that’s not actually what I saw, but that’s understandable as there ar not a lot of China experts who also track the day-to-day of U.S. domestic AI regulation, and honestly almost no one is paying attention to the Trump secret de facto licensing system.

When I read the CCTV post’s second principle, I couldn’t quite tell if the section was written by someone who who was attempting write what sounded like significant demands from the U.S. or if it was written by someone clever who understood that these demands sounded significant but in reality the U.S. had actually already done, or not done in the case of restricting Chinese AI models, or was likely to do in the near future almost all of the things demands. I suspect it was the former but who knows?

We touched on this some in our CAP report but I wanted to break down my thinking with with additional details and commentary, starting here (block quotes below are quoting from my own report):

At the same time that Office of Science and Technology Policy Director Michael Kratsios was introducing the aforementioned Carolina Principles at the G20 ministerial, a China Central Television (CCTV) post (translation from Geopolitechs), widely viewed as signaling Chinese conditions for the September talks, attacked U.S. frontier AI company Anthropic.

The majority of the post is a screed against Anthropic and its co-founder Dario Amondei, who has long publicly held that China above all else should not be allowed to achieve transformative AI due to the danger of digital authoritarianism (and who became even less popular with the Chinese a few weeks later when he posted his “We Must Pace the Frontier” essay).

This attack on possibly the fast growing U.S. company in history which is set to IPO could seem like a significant attack on a superstar American company but Anthropic has spent most of 2026 having drawn the wrath of the Trump administration. From the Secretary of War’s attempt to execute the entire company by attempting to assert a Supply Chain Risk designation using authorities he did not have to the administration slapping export controls on their Mythos and Fable models forcing them to withdraw them from the public, it is hard to say at any given moment if the Trump administration considers Anthropic a friend or an enemy.

So the attacks on Anthropic could be one part legitimate anger and one part an attempt to let the Trump administration know they could have a common enemy.

What is more important, and what we will get to below, are the two principles for discussion, the first of which is:

[The post] laid out two principles for discussion with the United States on AI starting with “a distinction between what constitutes a genuine security threat and what is merely technological competition.” There are numerous ways to interpret this; the Chinese have many different definitions of “loss of control” as Zilan Qian of the Oxford China Policy Lab has noted, and there will always be ways to refuse to come to agreement on a definition. But fundamentally, a narrow and mutually agreed-upon definition of “genuine security threats” is the first step to any AI security discussion of AI safety risks.

I won’t dwell too long on this part and I understand that there are numerous ways to interpret this, the Chinese could twist this, they can stall, actually mean export controls, etc. I get it. But fundamentally, if you just take this literally, these two geopolitical adversaries would obviously need to discuss and agree upon what they consider a “genuine security threats” in order to have a further discussion on how to deal with it.

I think the second principle is actually much more interesting and what I bring my expertise to examining.

The second principle seems boisterous at first read:

“Second, if the United States wants to talk about rules, it must first prove that the rules are equally effective against its own companies. Relevant figures on the American side are already discussing how to get China to agree to restrict ‘AI models with dangerous capabilities.’ The question is this: with companies like Anthropic in existence, before the United States asks China to restrict model releases and disclose risks, should it not first investigate its own companies, make public the purposes, scope, and rules of their identification mechanisms, and subject them to third-party audits?”

Starting slightly out of order with the part about the U.S. treating Chinese AI models, “Relevant figures on the American side are already discussing how to get China to agree to restrict ‘AI models with dangerous capabilities.’" I noted in the CAP report:

“After the July release of Moonshot AI’s Kimi K3 open-weight model, White House officials publicly accused Chinese AI companies of adversarial distillation and in September released a cybersecurity warning about Chinese distillation of U.S. AI models. The Trump administration reportedly considered taking action against Chinese open-weight AI models. But facing significant pushback from domestic tech companies, the United States has not yet taken steps against Chinese open-source or open-weight models and exempted them from the nonpublic “voluntary framework” ordered by the AI executive order issued June 2.

That means in the United States, Chinese AI models are subject to far fewer restrictions than American models currently under the Trump administration’s secret de facto licensing framework that exempts Chinese open-source or open-weight models.”

To double back to the beginning of the second principle, “if the United States wants to talk about rules, it must first prove that the rules are equally effective against its own companies” we wrote:

“In contrast, the United States has already restricted model releases for American AI but not for Chinese models, applying export controls to Anthropic’s release of Mythos and Fable and removing them from the market for weeks, and having OpenAI agree to a controlled release of GPT-5.6.”

Then the post says, "should it not first investigate its own companies”?

As for the CCTV post asking whether the United States “should it not first investigate its own companies,” American frontier AI labs are already under investigation as state attorneys general have opened investigations into OpenAI over the Hugging Face hack and some members of Congress are demanding hearings—though it is unclear whether the U.S. federal government will, or is, investigating OpenAI over the incident.

Really the only outlier here is that the federal government has not announced an FBI/DOJ investigation into OpenAI over Hugging Face (or any company over rogue AI agents incidents) and if one were announced ahead of the summit, that would be interesting timing.

Then the post says, “make public the purposes, scope, and rules of their identification mechanisms” and we note:

Congress and many organizations, including CAP, have demanded the administration make public its “voluntary framework,” lawsuits have been filed to compel its release, and its continued secrecy is likely unsustainable.

This week the Trump administration just agreed to make public by October 30th thanks to a lawsuit from Protect Democracy, which was also something the post demanded America do, and which I noted was likely to happen. Was the timing entirely coincidental? If you’re going to lose in court and be forced to reveal it, might as well take away a Chinese talking point with the timing of the announcement too.

“And subject them to third-party audits”:

Furthermore, Illinois law SB0315, which goes into effect in January 2027, requires independent third-party audits for frontier AI models beginning in January 2028.

There is currently only one frontier AI law in effect in the U.S., California’s SB 53, but is part of a trio of state frontier AI laws go into effect that are relatively harmonized after brutal fights between AI companies and activists, with the Illinois law requiring third-party audits starting in 2028 (and California having just passed a bill licensing Independent Verification Organizations (IVOs) and an executive order looking at embedding evaluators in frontier labs). So the U.S. will soon be subjecting its frontier AI companies to third-party audits.

Then there’s this paragraph:

“If American companies can return data without users knowing, skip permission confirmation, and enable automated decision-making by default, while the U.S. government says nothing, then American talk about “security boundaries” is empty.”

This echoes a lot of complaints about Anthropic and Claude Code in the parts above, tying this to security risks and I autonomy.

“Only after the United States proves that its safety rules are equally binding on its own model companies can China and the United States move into substantive discussion.”

Again noting above, the U.S. has a state law that governs frontier AI, a secret de facto licensing framework for U.S. frontier AI models while exempted Chinese AI models from that framework, and has slapped export controls on Anthropic models. So are these demands serious or just meant to sound serious?

Given the Trump admin AI framework will eventually be public, at this point we should keep an eye out for two things ahead of, during, or after these meetings:

• The announcement of an FBI and/or DOJ investigation into a U.S. AI company for a rouge agent incident ( I suppose extra points for the Chinese if the investigation is of Anthropic).

• The random trashing of Claude Code’s auto mode or a reference to “dangerously-skip-permissions” by a senior administration official or Trump.

Anyway, while not reading too much into a single post, it is interesting that the second principle articulated may seem significant at first glance but is mostly already achieved when closely examined. Guess we will find out over the next week.

Almost Every House Republican Voted To Block State and Local Governments’ Ability To Protect Residents From AI Data Centers

The House version of the One Big Beautiful Bill Act contained a state and local AI moratorium that, if it had become law, would have prevented local control efforts to limit or restrict AI data centers.

By Adam Conner

On May 22, 2025, the House of Representatives passed its version¹ of the One Big Beautiful Bill Act (OBBBA),² which contained a section titled “Artificial Intelligence and Information Technology Modernization Initiative”;³ included a moratorium (AI moratorium) for 10 years on the enforcement of any state and local laws and regulations that attempted to limit, restrict, or regulate Artificial Intelligence (AI) models and systems;⁴ and would have included laws that allowed state and local governments to restrict or tax AI data centers. Had it become law, the House OBBBA’s AI moratorium would have federally preempted for a decade all state and local AI laws and regulations that might limit, restrict, or regulate AI generally and included language that would likely only apply to AI data centers specifically. The National Conference of State Legislatures (NCSL) warned at the time that the House OBBBA AI moratorium would “circumvent [states’] authority to regulate the permitting, construction and operation of data centers within their borders.”⁵

The Trump administration has made AI deregulation a top priority,⁶ including the preemption of state AI laws, and House Republicans have strongly backed this effort. A total of 215 Republican members voted for the House version of OBBBA,⁷ and those House Republicans who sat on the House Energy and Commerce Committee voted for it three additional times during the committee process, including Reps. Gabe Evans (R-CO), John James (R-MI), Tom Kean (R-NJ), and Mariannette Miller-Meeks (R-IA).⁸

The state and local AI moratorium was ultimately removed from the Senate’s version of OBBBA right before passage⁹ and was not in the final version of the bill.¹⁰

That does not change the fact that almost every House Republican in the 119th Congress voted for a provision¹¹ that would have stripped state and local governments of their ability to establish guardrails against the proliferation of AI data centers that harm residents. Many members of Congress who voted to take away state and local governments’ ability to limit or restrict AI data centers are now running on a platform of “local control” for data centers—local control they previously voted to severely curtail. Americans deserve to know the truth about members of Congress who have flip-flopped on their position on local control of data centers.

Background on the House OBBBA state and local AI moratorium

The state and local AI moratorium in the House OBBBA bill was located in SEC. 43201 “Artificial Intelligence and Information Technology Modernization Initiative” and specifically said:¹²

(c) Moratorium.–

(1) In general.–Except as provided in paragraph (2), no State or political subdivision thereof may enforce, during the 10-year period beginning on the date of the enactment of this Act, any law or regulation of that State or a political subdivision thereof limiting, restricting, or otherwise regulating artificial intelligence models, artificial intelligence systems, or automated decision systems entered into interstate commerce.

In this context, a “political subdivision”¹³ means substate government, such as city or county governments, but will be referred to here as “local” government.

The House OBBBA state and local AI moratorium aimed to achieve two goals: first, to preempt any state or local laws or regulations that are “limiting, restricting, or otherwise regulating” AI or AI systems including those covering frontier risk, algorithmic risk, or data centers, and second, to allow state laws that encouraged or eased AI adoption, including the building of AI data centers. CAP opposed the AI moratorium in the House bill and the subsequent Senate versions, and has noted the difficulty and downsides in crafting broad preemption legislation for AI.¹⁴

The House OBBBA AI moratorium explicitly distinguishes and separately defines artificial intelligence, artificial intelligence models, artificial intelligence systems, and automated decision systems.¹⁵ Notably, the House did not limit the moratorium to laws regulating AI models, meaning the software or computational models themselves that are digital only, but included broader definitions that encompassed physical elements of AI. The bill specifically defined “artificial intelligence system” separately as:¹⁶

(3) Artificial intelligence system.–The term “artificial intelligence system” means any data system, hardware, tool, or utility that operates, in whole or in part, using artificial intelligence.

The breadth of this definition of “artificial intelligence system” is significant because AI systems depend on an interconnected stack of software, computing hardware, physical infrastructure, and supporting utilities. AI data centers are a foundational physical component of what industry calls “The Tech Stack for AI,” which includes multiple parts of AI systems building on top of each other including electrical power and internet connectivity (utility), semiconductors/chips which are located in AI data centers (hardware), and then the data, AI models, and applications built on top of it.¹⁷

In the context of the AI moratorium and definitions, “artificial intelligence system” including “data system, hardware, tool, or utility” ¹⁸ seems to be a reference to various components of the tech stack for AI, which would include the semiconductors and the data centers that house them.

AI data centers contain and operate multiple complex hardware and data systems that both utilize AI to operate and directly execute AI workloads, meaning that at least these components operate in part using AI.¹⁹

The moratorium targeted state and local laws and regulations that might restrict AI but attempted to exempt state laws or rules that made it easier to allow AI systems, including AI data centers, using this rule of construction:²⁰

(2) Rule of construction.–Paragraph (1) may not be construed to prohibit the enforcement of–

(A) any law or regulation that–

(i) the primary purpose and effect of which is to—

(I) remove legal impediments to, or facilitate the deployment or operation of, an artificial intelligence model, artificial intelligence system, or automated decision system;

Or

(II) streamline licensing, permitting, routing, zoning, procurement, or reporting procedures in a manner that facilitates the adoption of artificial intelligence models, artificial intelligence systems, or automated decision systems;

(ii) does not impose any substantive design, performance, data-handling, documentation, civil liability, taxation, fee, or other requirement on artificial intelligence models, artificial intelligence systems, or automated decision systems [. . .]

The House OBBBA AI moratorium excluded federal law; attempted to exempt state and local “generally applicable” laws that would apply equally to non-AI models and systems; and also exempted criminal laws.²¹

While the House OBBBA AI moratorium does not contain the phrase “data center,” the specific provision to exempt laws that “streamline licensing, permitting, routing, zoning, procurement, or reporting procedures in a manner that facilitates the adoption”²² would only make sense as they relate to AI data centers, as these are terms that in the context of state and local laws and regulations apply primarily to the physical construction of structures, such as AI data centers.

Additionally, definitions used in the AI moratorium were so broad that they could encompass almost anything,²³ inviting litigation against state and local governments from any number of parties,²⁴ with CAP previously noting:

What is clear is that such ambiguous language invites lawsuits and regulatory paralysis, particularly because the reconciliation moratorium allowed private entities, not just the federal government, to bring enforcement actions. As a result, states would face the risk of being tied up in costly litigation for years, unable to advance legislation without legal challenge.²⁵

The NCSL, experts in the field of state laws, recognized the AI moratorium’s danger to state and local control of data centers and wrote in a May 2025 letter to the House Energy and Commerce Committee opposing the bill:²⁶

Restricting states’ ability to “enforce any law or regulation regulating … artificial intelligence systems” will circumvent their authority to regulate the permitting, construction and operation of data centers within their borders. This will severely limit the ability of states and localities to make decisions regarding the siting and operation of these large-scale projects, raising costs for ratepayers, jeopardizing zoning decisions that protect our mutual constituents and impacting existing infrastructure such as power grids and generating facilities.

The House OBBBA AI moratorium essentially barred state and local governments from enforcing any law that imposed a “substantive design, performance, data-handling, documentation, civil liability, taxation, fee, or other requirement” on AI,²⁷ and it required that any “fee or bond” be “reasonable and cost-based” and “treated in the same manner as other models and systems that perform comparable functions.”²⁸ This means that AI data centers, if they are considered an AI system, could not have substantive performance, transparency, documentation, taxes, or fees applied to them by state or local law or regulation.

The House OBBBA AI moratorium did attempt to allow laws or regulation to exist that “remove legal impediments to, or facilitate the deployment or operation of,” AI or “streamline licensing, permitting, routing, zoning, procurement, or reporting procedures in a manner that facilitates the adoption of artificial intelligence.”²⁹

Under the House’s OBBBA AI moratorium, state and local governments would not have been allowed to create or enforce any laws or regulations that required additional licenses, permits, zoning, taxes, or utility costs specifically targeted to AI data centers.

Under the House’s OBBBA AI moratorium, state and local governments would not have been allowed to create or enforce any laws or regulations that required additional licenses, permits, zoning, taxes, or utility costs specifically targeted to AI data centers.

House votes

The E&C committee markup on their OBBBA text and amendments³⁰ started on May 13 and stretched through the night into May 14, with the markup taking over 26 hours.³¹ Reps. Gabe Evans (R-CO), John James (R-MI), Tom Kean (R-NJ), and Mariannette Miller-Meeks (R-IA) all serve on the House Energy and Commerce Committee³² and voted three times during this markup for the state and local AI moratorium.³³

The state and local AI moratorium was part of “Title IV—Committee on Energy and Commerce, Subtitle C—Communications”³⁴ and House Energy and Commerce Committee Republicans voted at least three separate times on keeping or passing the bill with the moratorium.

The committee considered COMM19, a minority amendment from ranking member Rep. Frank Pallone (D-NJ) to strike the entirety of the moratorium from the bill,³⁵ but it was rejected by the majority with a vote of 29 nays to 24 yeas, with Reps. Evans, James, Kean, and Miller-Meeks all voting to keep the AI moratorium in OBBBA.³⁶

The Republican majority later passed a motion to transmit the entire “Title IV—Committee on Energy and Commerce, Subtitle C—Communications”³⁷ as amended to the House Committee on the Budget on a party line vote with Reps. Evans, James, Kean, and Miller-Meeks all voting for the communications subtitle that included the AI moratorium.³⁸

Finally, the Republican majority voted to transmit the recommendations of the committee from the markup, including Subtitle C—Communications, which contained the state and local AI moratorium, to the House Budget committee for the next step in the process.³⁹ All Committee Republicans voted for this including Reps. Evans, James, Kean, and Miller-Meeks.⁴⁰

As the OBBBA made its way to the House floor in late May 2025, 215 House Republicans voted for the bill that included the state and local AI moratorium.⁴¹

Ultimately, the state and local AI moratorium was removed from the Senate’s version of OBBBA in a dramatic early morning vote⁴² before it became law. The Trump administration’s later state AI law preemption efforts have specifically claimed to exempt local control of data centers,⁴³ possibly because they are aware of how politically unpopular the issue has become. The votes from almost all House Republican members for the state and local AI moratorium in the House version of OBBBA are on the record and speak for themselves.⁴⁴

Effect of the state and local AI moratorium had it become law

Had the state and local AI moratorium in the House OBBBA bill become law, it would have prevented the enforcement of any state or local laws or regulations that limited, restricted, or otherwise put additional obligations on AI or AI systems for 10 years while exempting any state or local laws or regulations designed to increase the adoption or development of AI and AI systems.⁴⁵

In addition to laws passed by state legislatures addressing AI data centers, governors are taking various executive actions on AI data centers often through executive power and regulations. It is unclear how these various executive actions would have been constrained by the AI moratorium if it had become law, though it is almost certain that they would have faced litigation for their actions.⁴⁶

Arizona

In Arizona, Rep. Andy Biggs (R-AZ) is the Republican nominee for governor. During an interview with local radio station KTAR News, Biggs said data centers must supply their own power and must have a closed-loop water system⁴⁷ and, in another, he said “it becomes a local jurisdiction issue and the state’s just got to make sure we take care of the water and power situation.”⁴⁸ If the House OBBBA state and local AI moratorium that Rep. Biggs voted for had become law,⁴⁹ it likely would have prevented attempts by state or local governments to enforce new requirements on power or water for AI data centers, taking away state and local jurisdiction over AI data centers.

Iowa

In Iowa, Rep. Ashley Hinson (R-IA) is the Republican nominee for U.S. Senate. In an August interview with Iowa’s News Now,⁵⁰ Rep. Hinson said when asked about data centers that, “I continue to hear from my constituents that they’re also concerned about data centers and what I have said is I believe that these decisions are best decided at the local level.” When asked by the anchor, “Should Congress step in in any way?” Hinson responded, “I think again this should be decided at the local level” and later reiterated that “those decisions should be made at the local level in terms of respecting the constituents on the ground.”⁵¹

Rep. Hinson did not mention her vote for the House OBBBA bill that contained the state and local AI moratorium,⁵² which, had it become law, would likely have eliminated for 10 years the ability for state and local governments to have full control over decisions about AI data centers.

Also in Iowa, Rep. Miller-Meeks (R-IA) was asked by Iowa’s News Now about data centers and responded, “People are rightly concerned about water usage and electricity rates going up.”⁵³ Had the House OBBBA state and local AI moratorium that Rep. Miller-Meeks voted for four times become law,⁵⁴ the state of Iowa likely would not have been able to apply any specific additional requirements or transparency obligations for AI data centers on their water and power usage.

Michigan

In Michigan, Rep. John James (R-MI) is currently the Republican nominee for governor. When asked about his position on data centers in June, Rep. James was quoted on his position on data centers as “Number one, local control.”⁵⁵But if the state and local AI moratorium that Rep. James voted for four times last year⁵⁶ had become law, it likely would have stripped the ability for state and local governments to restrict AI data centers in Michigan, making the “local control” he supports impossible.

Similarly in Michigan, Rep. Tom Barrett (R-MI) is similarly running campaign ads vowing to “protect local control” and stating “I believe you should decide on data centers, not Washington,” arguing that Washington, D.C., politicians shouldn’t dictate the future of data center development, despite voting to do exactly the opposite.⁵⁷ Rep. Barrett even went so far as to recently admit that the people he speaks to “overwhelmingly” do not support efforts that “nationalize” decisions and override local control and denounced “proposals to nationalize decisions that would override local communities about their decisions around data centers”⁵⁸ while forgetting to mention that he voted for a national proposal to override local communities about their decisions around AI data centers.⁵⁹

New Jersey

In New Jersey, Gov. Mikie Sherrill (D-NJ) signed S731/A796 into law which “Requires electric public utilities to develop and apply special rules for certain data centers to protect non-data center customers from increased costs.”⁶⁰ This law singling out data centers would likely not have been allowed under the House OBBBA state and local AI moratorium since it imposed specific requirements for data centers which would likely have been considered a “substantive design, performance, data-handling, documentation, civil liability, taxation, fee, or other requirement” of an AI system.⁶¹ Rep. Tom Kean (R-NJ) voted four separate times for the state and local AI moratorium in the House OBBBA bill,⁶² which would likely have prevented the state of New Jersey from enforcing a law that required “special rules for certain data centers.”⁶³

New York

In New York, Gov. Kathy Hochul (D-NY) signed an executive order ordering a one-year moratorium on hyperscale data centers until statewide standards can be developed.⁶⁴ Rep. Mike Lawler (R-NY) has attacked the governor for these actions⁶⁵ and had the House OBBBA state and local AI moratorium he voted for become law,⁶⁶ the governor’s AI moratorium would likely have not been possible since it was targeted at hyperscale data centers, which are generally AI data centers.

Pennsylvania

In Pennsylvania, the state legislature passed a law in July 2026 that required specific energy and water reporting from data centers.⁶⁷ If the House version of OBBBA state and local AI moratorium had become law, these requirements would likely not be enforceable as they imposed substantive documentation requirements on data centers that use a certain amount of power, which are primarily AI data centers. In August, Gov. Josh Shapiro (D-PA) signed an executive order that required all data center permits to comply with the “Governor’s Responsible Infrastructure Development (GRID) Requirements” and “Establishing Strict Guardrails for Data Center Development.”⁶⁸

Had the AI moratorium in the May 2025 House version of OBBBA that Reps. Brian Fitzpatrick (R-PA), Ryan Mackenzie (R-PA), Rob Bresnahan (R-PA), and Scott Perry (R-PA) voted for⁶⁹ become law, it likely would have banned the development of these GRID requirements and guardrails specifically for data centers.⁷⁰

Wisconsin

In Wisconsin, Rep. Tom Tiffany (R-WI) is the Republican nominee for governor. On Tiffany’s campaign website, “Data Centers” is the sole subsection in the solutions section of his campaign website, and it states, “Give local communities the final say,” and that, “No data center should be forced on a community that doesn’t want it. Local residents should have the final say over what gets built in their backyard.”⁷¹ In an August interview with local Wisconsin news station 620 WTMJ, when asked about data centers, Tiffany said, “there should be local input. There should be local control. Local people should have a say in whether it’s being built in their community.”⁷² Had the House OBBBA bill that contained the state and local AI moratorium that Rep. Tiffany voted for become law,⁷³ it would have likely made “local control” over AI data centers impossible by restricting the enforcement of any state or local laws and regulations that limit, restrict, or regulate AI systems like AI data centers.

Similarly, Rep. Derrick Van Orden (R-WI) has posted on X that “conditions for data centers have not changed and are not negotiable: 1) must have local buy in.”⁷⁴ But it is hard to understand what Rep. Van Orden means by “local buy in” because if the House OBBBA state and local AI moratorium that he voted for⁷⁵had become law, it would have preempted any state or local laws or regulations attempting to limit or restrict AI data centers.

Conclusion

Ultimately, had the House’s proposed OBBBA state and local AI moratorium become law, state or local laws or regulations that add any requirements to AI or AI systems, including AI data centers, would not have been enforceable but any laws or regulations that made it easier to deploy or operate AI, including AI data centers, would have been allowed. It would have tied the hands of state and local governments seeking to govern AI data centers, limiting their ability to limit, restrict, or regulate the local impacts of these facilities, including their demands on electricity, water, land, and other public resources.

Many of the members of Congress in Arizona, Iowa, Michigan, New Jersey, New York, Pennsylvania, and Wisconsin now touting “local control” for data centers all voted to take away state and local governments’ ability to limit or restrict AI data centers when they cast their vote for the House’s version of OBBBA that contained the state and local AI moratorium.

Endnotes

1. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025), available at https://www.congress.gov/votes/house/119-1/145.

2. ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess. (May 21, 2025), available at https://www.congress.gov/congressional-record/volume-171/issue-86/house-section/article/H2235-2.

3. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

4. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c) Moratorium, ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

5. National Conference of State Legislatures, “NCSL Urges Congress to Oppose AI Moratorium on States,” May 13, 2025, available at https://www.ncsl.org/resources/details/ncsl-urges-congress-to-oppose-ai-moratorium-on-states.

6. The White House, “Winning the Race: America’s AI Action Plan” (Washington: White House, July 2025), available at https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf.

7. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

8. House Committee on Energy and Commerce, Roll Call Vote #14 on Amendment COMM19 to bill “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025), available at https://docs.house.gov/meetings/IF/IF00/20250513/118261/CRPT-119-IF00-Vote014-20250513.pdf; House Committee on Energy and Commerce, Roll Call Vote #15 on “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025), https://docs.house.gov/meetings/IF/IF00/20250513/118261/CRPT-119-IF00-Vote015-20250513.pdf; House Committee on Energy and Commerce, Roll Call Vote #39 on “A motion by Chairman Guthrie to transmit the recommendations of this committee, approved as Subtitles A—Energy, B—Environment, C—Communications, and D—Health as amended, and all appropriate accompanying materials including supplemental, minority, additional, or dissenting views to the House Committee on the Budget,” 119th Cong., 1st Sess. (May 13, 2025), https://docs.house.gov/meetings/IF/IF00/20250513/118261/CRPT-119-IF00-Vote039-20250513.pdf.

9. U.S. Senate, Roll Call Vote, Blackburn Amendment No. 2814 to S.Amdt. 2360 to H.R. 1, Vote #363, 119th Cong., 1st Sess. (July 1, 2025), available at https://www.senate.gov/legislative/LIS/roll_call_votes/vote1191/vote_119_1_00363.htm.

10. H.R.1 – An act to provide for reconciliation pursuant to title II of H. Con. Res. 14, Statute at Large 139 Stat. 72 – Public Law No. 119-21, 119th Cong., 1st Sess. (July 4, 2025), available at https://www.congress.gov/bill/119th-congress/house-bill/1/text.

11. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

12. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c) Moratorium, ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

13. “Definition of political subdivision,” 42 U.S. Code § 247b-21(e), available at https://www.law.cornell.edu/uscode/text/42/247b-21#e.

14. Adam Conner, “The House Is Close To Passing a Moratorium on State Efforts To Regulate AI,” Center for American Progress, May 15, 2025, available at https://www.americanprogress.org/article/the-house-is-close-to-passing-a-moratorium-on-state-efforts-to-regulate-ai/; Nicole Alvarez, “The Senate’s AI Ban Applies to Every State, Not Just BEAD Recipients,” Center for American Progress, June 13, 2025, available at https://www.americanprogress.org/article/the-senates-ai-ban-applies-to-every-state-not-just-bead-recipients/; Adam Conner and Nicole Alvarez, “The Senate’s AI Pause May Take Billions in State Broadband Funds Hostage,” Center for American Progress, June 26, 2025, available at https://www.americanprogress.org/article/the-senates-ai-pause-may-take-billions-in-state-broadband-funds-hostage/; Nicole Alvarez and Adam Conner, “4 Reasons the Senate’s AI Pause Should Be Opposed,” Center for American Progress, June 27, 2025, available at https://www.americanprogress.org/article/4-reasons-the-senates-ai-pause-should-be-opposed/; Nicole Alvarez, “Moratoriums and Federal Preemption of State Artificial Intelligence Laws Pose Serious Risks,” Center for American Progress, November 19, 2025, available at https://www.americanprogress.org/article/moratoriums-and-federal-preemption-of-state-artificial-intelligence-laws-pose-serious-risks/.

15. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c)(2)(B)(d) Definitions, ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

16. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c)(2)(B)(d)(3) Artificial intelligence system, ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

17. Brad Smith, “Microsoft’s AI Access Principles: our commitments to promote innovation and competition in the new AI economy,” Microsoft, February 26, 2024, available at https://blogs.microsoft.com/on-the-issues/2024/02/26/microsoft-ai-access-principles-responsible-mobile-world-congress/.

18. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c)(2)(B)(d)(3) Artificial intelligence system, ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

19. Taryn Plumb, “What is an AI-optimized data center? Why next-gen infrastructure is nonnegotiable,” Network World, September 4, 2025, available at https://www.networkworld.com/article/4050973/inside-the-ai-optimized-data-center-why-next-gen-infrastructure-is-non-negotiable.html; “How Mechanical Systems Support the Shift to AI-Optimized Data Centers,” Cadence, available at https://cadencenow.com/how-mechanical-systems-support-the-shift-to-ai-optimized-data-centers/#The_Growing_Demand_for_AI-Optimized_Data_Centers (last accessed August 2026).

20. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c)(2) Rule of Construction, ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

21. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c) Moratorium (2)(A)(ii)(II) and (c) Moratorium (2)(B), ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

22. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c) Moratorium, (2)(A)(i)(II), ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

23. David Brody, “The Big Beautiful Bill Could Decimate Legal Accountability for Tech and Anything Tech Touches,” Tech Policy Press, May 27, 2025, available at https://www.techpolicy.press/the-big-beautiful-bill-could-decimate-legal-accountability-for-tech-and-anything-tech-touches/.

24. Charlie Bullock and Mackenzie Arnold, “The AI Moratorium—deobligation issues, BEAD funding, and independent enforcement,” Institute for Law and AI, June 2025, available at https://law-ai.org/the-ai-moratorium-deobligation-issues-bead-funding-and-independent-enforcement/.

25. Nicole Alvarez, “Moratoriums and Federal Preemption of State Artificial Intelligence Laws Pose Serious Risks.”

26. National Conference of State Legislatures, “NCSL Urges Congress to Oppose AI Moratorium on States.”

27. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c) Moratorium, (2)(A)(i)(II)(ii), ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

28. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c) Moratorium, (2)(A)(iii), ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

29. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c) Moratorium, (2)(A)(i), ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

30. U.S. House of Representatives Committee Repository, “Markup of Markup of Four Committee Prints, Committee on Energy and Commerce,” May 13, 2025, available at https://docs.house.gov/Committee/Calendar/ByEvent.aspx?EventID=118261.

31. Alice Miranda Ollstein and Ben Leonard, “E&C Republicans advance health care piece of GOP tax bill after marathon markup,” Politico, May 14, 2025, available at https://www.politico.com/live-updates/2025/05/14/congress/e-c-republicans-pass-health-care-piece-of-gop-tax-bill-after-marathon-overnight-session-00350223.

32. House Committee on Energy and Commerce, “Representatives,” available at https://energycommerce.house.gov/representatives (last accessed August 2026).

33. House Committee on Energy and Commerce, Roll Call Vote #14 on Amendment COMM19 to bill “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025); House Committee on Energy and Commerce, Roll Call Vote #15 on “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025); House Committee on Energy and Commerce, Roll Call Vote #39 on “A motion by Chairman Guthrie to transmit the recommendations of this committee, approved as Subtitles A—Energy, B—Environment, C—Communications, and D—Health as amended, and all appropriate accompanying materials including supplemental, minority, additional, or dissenting views to the House Committee on the Budget,” 119th Cong., 1st Sess. (May 13, 2025).

34. House Committee on Energy and Commerce, “TITLE IV—ENERGY AND COMMERCE, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 11, 2025), available at https://docs.house.gov/meetings/IF/IF00/20250513/118261/BILLS-119CommitteePrintSubtitleCpp.pdf.

35. House Committee on Energy and Commerce, Amendment COMM19 “AMENDMENT TO THE COMMITTEE PRINT FOR SUBTITLE C,” 119th Cong., 1st Sess. (May 13, 2025), available at https://docs.house.gov/meetings/IF/IF00/20250513/118261/BILLS-119-CommitteePrintSubtitleC-P000034-Amdt-098.pdf

36. House Committee on Energy and Commerce, Roll Call Vote #14 on Amendment COMM19 to bill “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025).

37. House Committee on Energy and Commerce, “TITLE IV—ENERGY AND COMMERCE, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 11, 2025).

38. House Committee on Energy and Commerce, Roll Call Vote #15 on “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025).

39. House Committee on Energy and Commerce, Roll Call Vote #39 on “A motion by Chairman Guthrie to transmit the recommendations of this committee, approved as Subtitles A—Energy, B—Environment, C—Communications, and D—Health as amended, and all appropriate accompanying materials including supplemental, minority, additional, or dissenting views to the House Committee on the Budget,” 119th Cong., 1st Sess. (May 13, 2025).

40. Ibid.

41. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

42. U.S. Senate, Roll Call Vote, Blackburn Amendment No. 2814 to S.Amdt. 2360 to H.R. 1, Vote #363, 119th Cong., 1st Sess. (July 1, 2025).

43. The White House, “Ensuring a National Policy Framework for Artificial Intelligence,” DCPD-202501186, Executive Order 14365, (Washington: White House, December 11, 2025), available at https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/and https://www.govinfo.gov/app/details/DCPD-202501186; The White House, “National Policy Framework for Artificial Intelligence,” (Washington: White House, March 2026), available at https://www.whitehouse.gov/wp-content/uploads/2026/03/03.20.26-National-Policy-Framework-for-Artificial-Intelligence-Legislative-Recommendations.pdf.

44. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

45. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c) Moratorium, ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

46. Charlie Bullock and Mackenzie Arnold, “The AI Moratorium—deobligation issues, BEAD funding, and independent enforcement.”; Nicole Alvarez, “Moratoriums and Federal Preemption of State Artificial Intelligence Laws Pose Serious Risks.”

47. Jim Sharpe, “AZ Political Podcast: Andy Biggs talks about running for governor, what he’ll do if he wins,” KTAR News 92.3 FM, July 9, 2026, available at https://ktar.com/jim-sharpe/az-political-podcast-andy-biggs-2; KTAR News 92.3, Facebook, July 11, 2026, available at https://www.facebook.com/watch/?v=1373152088032297.

48. KTAR News 92.3 FM, “How Andy Biggs plans to handle Arizona’s data center boom,” YouTube, July 23, 2026, available at https://www.youtube.com/shorts/gJhHSQPj3g8.

49. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

50. Iowa’s News Now, “Connect to Congress: Rep. Ashley Hinson talks new health insurance bill,” YouTube, August 10, 2026, available at https://youtu.be/cpPnPyXq6eA?si=P191QOCPWTLXT8MT&t=497.

51. Ibid.

52. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

53. Ray Baccari, “Rep. Miller-Meeks discusses ‘Stop Insider Trading Act,’ daylight saving time bill,” Iowa’s News Now, July 22, 2026, available at https://cbs2iowa.com/news/local/rep-miller-meeks-discusses-stop-insider-trading-act-daylight-saving-time-bill.

54. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025); House Committee on Energy and Commerce, Roll Call Vote #14 on Amendment COMM19 to bill “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025); House Committee on Energy and Commerce, Roll Call Vote #15 on “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025); House Committee on Energy and Commerce, Roll Call Vote #39 on “A motion by Chairman Guthrie to transmit the recommendations of this committee, approved as Subtitles A—Energy, B—Environment, C—Communications, and D—Health as amended, and all appropriate accompanying materials including supplemental, minority, additional, or dissenting views to the House Committee on the Budget,” 119th Cong., 1st Sess. (May 13, 2025).

55. Melanie Tolen, “Michigan governor candidate John James on education, data centers and election security,” Spectrum Local News, June 9, 2026, available at https://spectrumlocalnews.com/mi/michigan/news/2026/06/09/gov-candidate-johnjames-keyissues; Killian Bies, “Watch: John James opposes data centers on productive farmland, backs local control at Cement City roundtable,” The Midwesterner, June 17, 2026, available at https://www.themidwesterner.news/2026/06/watch-john-james-opposes-data-centers-on-productive-farmland-backs-local-control-at-cement-city-roundtable.

56. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025); House Committee on Energy and Commerce, Roll Call Vote #14 on Amendment COMM19 to bill “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025); House Committee on Energy and Commerce, Roll Call Vote #15 on “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025); House Committee on Energy and Commerce, Roll Call Vote #39 on “A motion by Chairman Guthrie to transmit the recommendations of this committee, approved as Subtitles A—Energy, B—Environment, C—Communications, and D—Health as amended, and all appropriate accompanying materials including supplemental, minority, additional, or dissenting views to the House Committee on the Budget,” 119th Cong., 1st Sess. (May 13, 2025).

57. Tom Barrett, X, August 19, 2026, available at https://x.com/tombarrettmi7/status/2090106324990206410?s=20.

58. Dave Weigel, X, August 19, 2026, available at https://x.com/daveweigel/status/2090105467993227572?s=20.

59. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

60. ASSEMBLY, No. 796, State of New Jersey, 222nd Legislature, (June 30, 2026), available at https://pub.njleg.state.nj.us/Bills/2026/A1000/796_R4.PDF; “NJ A796,” Bill Track 50, July 7, 2026, available at https://www.billtrack50.com/billdetail/1954539 (last accessed August 2026); Zoë Read,
    “New Jersey Gov. Sherrill signs data center legislation that aims to protect ratepayers,” WHYY, July 10, 2026, available at https://whyy.org/articles/new-jersey-data-center-legislation-governor-mikie-sherrill/.

61. SEC. 43201. ARTIFICIAL INTELLIGENCE AND INFORMATION TECHNOLOGY MODERNIZATION INITIATIVE, (c) Moratorium, ONE BIG BEAUTIFUL BILL ACT; Congressional Record Vol. 171, No. 86, 119th Cong., 1st Sess.

62. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025); House Committee on Energy and Commerce, Roll Call Vote #14 on Amendment COMM19 to bill “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025); House Committee on Energy and Commerce, Roll Call Vote #15 on “Committee Print, Title IV—Committee on Energy and Commerce, Subtitle C—Communications,” 119th Cong., 1st Sess. (May 13, 2025); House Committee on Energy and Commerce, Roll Call Vote #39 on “A motion by Chairman Guthrie to transmit the recommendations of this committee, approved as Subtitles A—Energy, B—Environment, C—Communications, and D—Health as amended, and all appropriate accompanying materials including supplemental, minority, additional, or dissenting views to the House Committee on the Budget,” 119th Cong., 1st Sess. (May 13, 2025).

63. ASSEMBLY, No. 796, State of New Jersey, 222nd Legislature, (June 30, 2026).

64. Governor of New York, Executive Order No 62: “Establishing a Temporary Moratorium on Data Centers in New York While the State Develops Higher Standards for Data Center Development and Benefits Blueprint to Support Localities,” (Albany, New York State Governor, July 14, 2026), available at https://www.governor.ny.gov/executive-order/no-62-establishing-temporary-moratorium-data-centers-new-york-while-state-develops.

65. Arabella Bennett, “New York Republican congressman rips Hochul policies for sending businesses ‘don’t come here’ message,” Fox Business, August 18, 2026, https://www.foxbusiness.com/media/new-york-republican-congressman-rips-hochul-policies-sending-businesses-dont-come-here-message.

66. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

67. Commonwealth of Pennsylvania, SB 146, FISCAL CODE – OMNIBUS AMENDMENTS Act of Jul. 12, 2026, P.L. 265, No. 21, Session of 2026, No. 2026-21, July 12, 2026 available at https://www.palegis.us/statutes/unconsolidated/law-information/view-statute?txtType=PDF&SessYr=2026&ActNum=0021.&SessInd=0.

68. Commonwealth of Pennsylvania Governor’s Office, “Governor Shapiro Signs Executive Order Demanding Data Center Developers Comply with Strict Requirements and Blocking Speculative, Irresponsible Data Center Projects,” (Harrisburg, Commonwealth of Pennsylvania Governor’s Office, August 18, 2026), available at https://www.pa.gov/governor/newsroom/2026-press-releases/governor-shapiro-signs-executive-order-on-data-center-developmen; Commonwealth of Pennsylvania Governor’s Office, “Executive Order 2026-05 – Protecting Pennsylvania Consumers from Data Center Impacts,” (Harrisburg, Commonwealth of Pennsylvania Governor’s Office, August 18, 2026), available at https://www.pa.gov/content/dam/copapwp-pagov/en/governor/documents/eo2026_05_protecting%20pennsylvania%20consumers%20from%20data%20center%20impacts_final_executed.pdf.

69. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

70. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025); Commonwealth of Pennsylvania Governor’s Office, “Governor Shapiro Signs Executive Order Demanding Data Center Developers Comply with Strict Requirements and Blocking Speculative, Irresponsible Data Center Projects,” (Harrisburg, Commonwealth of Pennsylvania Governor’s Office, August 18, 2026); Commonwealth of Pennsylvania Governor’s Office, “Executive Order 2026-05 – Protecting Pennsylvania Consumers from Data Center Impacts,” (Harrisburg, Commonwealth of Pennsylvania Governor’s Office, August 18, 2026).

71. Tom Tiffany Governor, “Data Centers,” available at https://tomtiffany.com/solutions/data-centers/ (last accessed August 2026).

72. 620 WTMJ, “Tom Tiffany on His Plan for Wisconsin: DNR, Data Centers, Taxes & the Governor’s Race,” August 7, 2026 available at https://www.youtube.com/watch?v=x29Vl9zxpeg.

73. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

74. Derrick Van Orden, X, August 7, 2026, available at https://x.com/derrickvanorden/status/2085729781463535876.

75. H.R.1, House Roll Call Vote 145, 119th Cong., 1st Sess. (May 22, 2025).

The Trump Administration Has Created a De Facto Licensing System for Frontier AI Models

The administration’s new framework for frontier AI models is “voluntary” in name only.

Author: Adam Conner

In early August, the Trump administration reportedly finalized but declined to make public a “voluntary framework” for developers of “covered frontier models,” which are best described as “highly capable general-purpose AI models that can perform a wide variety of tasks and match or exceed the capabilities present in today’s most advanced models.” This marked the final action required by the president’s June artificial intelligence (AI) executive order, a process that began with the April announcement that Anthropic’s new frontier AI model Mythos posed tremendous cybersecurity capabilities and hacking risks. The Trump administration’s unprecedented application of export controls in June to Anthropic’s Fable and Mythos models, forcing their removal from public access, turned a secretive “voluntary framework” into something far more consequential: an emerging system of de facto licensing for America’s most powerful AI models that is “voluntary” in name only.

Without any explicit statutory authority from Congress, and in near-total secrecy, President Donald Trump has assumed control over the release of powerful new frontier AI models through the functional equivalent of a licensing regime—determining the standards those models must meet, who can have access to them early, and when companies may release them to the public. Combined with the White House’s refusal to make its AI framework public and President Trump’s history of corruption and retaliation against enemies, these steps represent unprecedented and dangerous executive branch authority over frontier AI models, which may soon become the single most valuable economic, security, and information tool in the world.

Congress cannot stand by and allow President Trump, or any president, to exert sole control over one of the most powerful sectors of the U.S. economy and national security. Both chambers must take seriously frontier AI risks and the danger of the Trump administration’s de facto AI model licensing system. Congress must immediately launch hearings and begin work to address frontier AI risks through an AI governance framework with clear processes and requirements for frontier models as well as an ability for the executive branch to temporarily restrict an AI model. In addition, clear guardrails must be established for any attempts to restrict an AI model, including mandatory disclosures, judicial review and access to courts, and even congressional overrides.

Trump’s “voluntary” framework amounts to a de facto licensing regime

Since Anthropic’s announcement of its Mythos model’s breakthrough cybersecurity abilities in April, the Trump administration has taken the following unprecedented steps:

• Determined a classified threshold definition for frontier AI models

• Established a frontier AI model prerelease submission process that allegedly includes “guardrail testing and security,” which later became part of a secret “voluntary framework”

• Exerted pressure on AI developers to join the “voluntary” program

• Exercised government influence or control over who gets early access to new frontier AI models as “trusted partners”

• Enforced the government’s demands by restricting model release via export controls

• Allowed the restoration of AI models only upon the government’s satisfaction

• Established the precedent of government–restricted rollout and approval before model release

No single step creates a licensing regime, but together they amount to one. The characterization of this as a de facto licensing system comes from experts acrossthe political spectrum.

The administration’s actions are a stark contradiction to the June 2 AI executive order, which stated: “Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models.” Less than two weeks later, on June 12, the U.S. Department of Commerce applied export controls to Anthropic’s Mythos and Fable models, removing the public’s access to the models, which sent a clear message about the “voluntary” nature of the government’s program.

Already, this is having the intended effect on frontier AI developers. On June 26, just two weeks after the Commerce Department imposed export controls on Anthropic, OpenAI previewed GPT-5.6, explicitly noting:

As part of our ongoing engagement with the U.S. government, we previewed our plans and the models’ capabilities ahead of today’s launch. At their request, we are starting with a limited preview for a small group of trusted partners whose participation has been shared with the government, before releasing more broadly.

OpenAI CEO Sam Altman also tweeted: “Bad news: at the request of the US government, it is launching today in limited preview instead of the open access launch we were planning on. We are working with the government to get to general availability as fast as we can.” Later, on July 8, Axios reported: “The Trump administration has given OpenAI the green light for a broad launch of its advanced GPT-5.6 model,” formalizing the practice of government approval for public release of frontier AI models.

OpenAI may have “voluntarily” agreed to the government’s request to limit its release of the GPT-5.6 model, but it is certain that the company took note of what the U.S. government had done to Anthropic—in the same way that OpenAI took note of what the U.S. government did to Anthropic for attempting to negotiate its contract to provide AI services on classified systems the day before OpenAI signed its own contract with the U.S. Department of Defense to provide AI services on classified systems.

The Trump administration’s unprecedented use of export controls on Anthropic serves as an implicit threat for all future frontier AI models. A program is “voluntary” in name only when the administration is punishing companies for not complying with its requests and pressuring others to join.

Potential for executive overreach raises risks for frontier AI governance

This new AI licensing system is also a stunning turn of events for a president who has said “it will be a policy of the United States to do whatever it takes to lead the world in artificial intelligence” and campaigned and delivered on the promise to “repeal Joe Biden’s dangerous Executive Order that hinders AI Innovation.” It should be noted that President Biden’s 2023 AI executive orderonly included two principal categories of regulations on AI and cloud companies—both of which were small in comparison to the program the Trump administration has assembled.

The Center for American Progress has written previously that the executive branch may need to take action against an AI model that poses a risk to American national security interests. The Trump administration’s actions against Anthropic’s Mythos model may have been legitimate or even similarly carried out by another administration.

But the second Trump administration has a long history of marshaling the fullpower of the federal government to target its enemies. Since February, the Department of Defense has been carrying out an unprecedented legal andpolitical fight with Anthropic, with the administration fighting in federal courtsto designate the company a “supply chain risk” and separately designating the company’s Mythos and Fable models for export controls. Additionally, Trump appointed officials have attempted to punish and control the speech of their critics and taken aim at the acceptable outputs from generative AI, even issuing an executive order to prevent “woke AI.” Due to these attempts, it is easy to imagine the administration secretly demanding unconstitutional content directives from frontier AI models as a condition of release. This history of abuse of power, bad blood with Anthropic, and secrecy around the Trump administration’s response to Mythos and subsequent refusal to make the new “voluntary framework” public mean that the Trump administration’s actions leave little reason to give it the benefit of the doubt.

This secretive “voluntary framework” provides no comfort to the public, other AI companies, U.S. allies, or even adversaries. No president—even one like Donald Trump, whose administration has been accused of widespread corruption—should be able to exercise sole power over such powerful tools.

Today, the absence of formal public rules and disclosure means the Trump administration could hypothetically determine that the Republican National Committee (RNC) is a “trusted partner” that could receive access to the latest AI models ahead of the election but the Democratic National Committee (DNC) is not—potentially giving one side a significant advantage ahead of Election Day. And it could do it all in secret. Similarly, the president could decide to not approve or to restrict the release of a company’s latest frontier AI model while allowing a competitor AI company—perhaps one in which he or his family has invested—to release its AI model. Under the Trump administration’s de facto AI model licensing system, the potential for abuse is too great.

Congress must demand answers and step in to establish frontier AI safeguards

Congress must act now. Several senators have already sent a letter to the administration seeking critical information about the new “voluntary framework.” After the August recess, Congress should immediately convene hearings on both the administration’s actions around the Mythos model export controls and the GPT-5.6 restricted rollout, as well as the impact of the administration’s de facto AI model licensing system. If this Congress will not hold hearings on these topics for fear of upsetting the administration, this topic should be on the top of the list for the next Congress.

Lawmakers should also begin work immediately on frontier AI governance, not only to address the president’s actions but to curb the significant risks from frontier models, such as the numerous recently disclosed rogue model hackingincidents. Significant efforts are needed to build a robust frontier AI governance framework with clear processes, requirements, and safeguards for frontier models. That frontier AI governance may require Congress to provide the president with a clear temporary ability to restrict AI models with guardrails and protections. Congress must also ensure that any administration’s actions to restrict AI models—including open-source or open-weight models—are subject to mandated disclosure, judicial review and access to courts, and even potential congressional override regardless of the authority used. The recently introduced Frontier Act and AI Kill Switch Act both represent early attempts to address some of these critical issues, and Congress must continue the work of crafting an AI governance framework.

Congress must also ensure measures are taken to allay people’s worst fears about AI, such as by prohibiting the most harmful uses of AI and addressing the concerns about job displacement. The public’s increasing distrust of AI, increasingly manifesting as anger at data centers, will not be solved by frontier AI governance alone.

Conclusion

In the span of months, this administration has gone from one that promised never to have licensing or preclearance of AI models to one individually approving the release of the latest AI models under its new secretive de facto AI model licensing system. The future of American frontier AI is too important to allow this president, or any president, to have sole control over the release of new American AI models. Congress must act immediately.

The net result of this is that the Trump administration has now established an entirely non-public de facto licensing system for frontier AI models, as demonstrated by the USG’s request and OpenAI’s agreement to limit the rollout of their latest ChatGPT 5.6 model to approved customers. Many many others from across the spectrum have shared this licensing diagnosis. That is not a good or sustainable method for handling frontier AI governance, and I am working on a longer piece on policy solutions for my day job.

But for now I want to call attention to the impossible to ignore fact that while it is possible that the administration acted in good faith and heeded its “deep state” experts in taking action against Anthropic, the secrecy of the process along with the Trump administration’s lawlessness, corruption, and history of retaliating against its enemies, makes that impossible to trust or verify. Even if Congress steps in and writes badly needed new laws, this administration’s dual state will be an inescapable part of the next few years of frontier AI governance.

The Dual State and the Trump Administration

In January of this year, as federal agents flooded the streets of Minneapolis and killed unarmed Americans, David French wrote in the New York Times about the concept of “the dual state.” This phrase emanated from German-Jewish lawyer Ernst Fraenkel who used it to describe how Nazi Germany maintained two states at the same time, one of laws and one lawless. Citing Aziz Huq’s prescient piece in the Atlantic highlighting the Trump administration’s version of the dual state, French describes the two components of the dual state, the normative and the prerogative state:

The two components of the dual state are the normative state — the seemingly normal world that you and I inhabit, where, as Huq writes, the “ordinary legal system of rules, procedures and precedents” applies — and the prerogative state, which is marked (in Fraenkel’s words) by “unlimited arbitrariness and violence unchecked by any legal guarantees.”

“The key here,” Huq writes, “is that this prerogative state does not immediately and completely overrun the normative state. Rather, Fraenkel argued, dictatorships create a lawless zone that runs alongside the normative state.”

French used the concept of the dual state to describe the interaction between the U.S. government’s immigration agents and Renee Good (and later Alex Pretti):

But if you interact with ICE, suddenly you risk coming up against the full force of the prerogative state. One of the most heartbreaking aspects of the ICE agent’s video of the fatal encounter between Renee Good and ICE is that it’s plain that Good thinks she’s still in the normative state. She has no idea of the peril she’s in.

[. . .] But she wasn’t in the normative state. She had crossed over the border to the prerogative state, and in that state you can be shot dead recklessly, irresponsibly and perhaps even illegally, and no one will pay the price. Your killer might even be rewarded with more than $1 million in donations from friends and allies.

The Trump administration is a dual state generally, and its sharpest and most violent edge has come in our citizens’ interactions with federal immigration agents. But the Trump administration’s prerogative state is not just about state-sanctioned violence but about the pervasive corruption and retaliation that President Trump and his administration have embodied.

I think sometimes people think that those of us with personal and professional disagreements with the President have our critiques dismissed or downplayed because many want to believe that we are still living in the normative and not the prerogative state.

So I will dispense with any hyperbole and note that when it comes to corruption, literally yesterday, President Trump’s latest financial disclosure showed he made more than $2.2 billion in 2025, the first year of his second term as President, including $1.4 billion from cryptocurrency. My employer CAP has previously totaled over $2.6 billion that President Trump has personally accrued from cryptocurrency, legal settlements, and more since 2024. President Trump has also used the entirety of the federal government to persecute his enemies. He has issued executive orders targeting big law firms, stripped security clearances, pushed for indictments of his enemies like former FBI Director James Comey, and the list goes on and on. All of these are examples of how it’s very possible that the government’s actions against AI models may be extractive or punitive.

The Dual State of AI Regulation

Very specifically when it comes to AI policy, the dual state is a useful concept in understanding the Trump administration’s potential actions. The press and many AI governance, safety, and legal experts have scrambled to make sense of the stories provided (leaked) to the press to justify the administration’s actions (and reactions) to Anthropic. For the most part, they have treated the government’s statements and actions as credible, with the benefit of the doubt that would be accorded to previous administrations acting in our national interest.

I have written at length about the Trump administration’s unprecedented attacks on Anthropic, including the Department of War’s designation of Anthropic as a “supply chain risk,” continued court battles to defend that designation, Secretary Hegseth’s initial tweet of an even broader threat that would destroy the company, and President Trump’s calling them a “RADICAL LEFT, WOKE COMPANY.”

You could easily argue with this evidence that the Trump administration considers Anthropic an adversary and was retaliating against them by slapping export controls on their Fable and Mythos models. Without any transparency or public process, it’s impossible to disprove.

Anthropic, by virtue of its vulnerability to the USG on a number of fronts, has to pretend that it’s in the normative state when dealing with the Trump administration. And who knows, maybe they were negotiating with the normative state.

But in this new secretive Trump administration licensing regime, companies like Anthropic and OpenAI don’t have that much leverage against a government that not only has tools like export controls (which may not survive legal scrutiny) but many other mechanisms to impact these companies, particularly to throw up roadblocks on their path to IPO.

The best rule of thumb for examining AI governance from the executive branch now? Don’t analyze government policy as if it comes only from the normative state. The prerogative state governs AI, and far more, in the Trump administration and will continue to do so for the next few years.

Here’s an except from my opening with a framing that I think helps outline the wild moment that we are in:

Over the last two weeks, the Department of Defense has initiated two wars: one against a nation with a long history of conflict with the United States; the other against one of the fastest-growing new companies in American history: Anthropic, the frontier artificial intelligence lab behind the popular Claude model. The DOD has effectively declared both Iran and Anthropic to be enemies of America, and though the weapons the DOD is using in each conflict differ dramatically—explosive missiles versus bureaucratic legal statutes—the department has made clear in both cases that its objective is to severely damage, if not totally destroy, the enemy. These two conflicts are intertwined and began on the same day: the very afternoon that the DOD blacklisted Anthropic, U.S. strikes on Iran commenced—with Claude reportedly helping to analyze intelligence and plan operations.

And why its important to distinguish between the illegal actions from the Trump Administration and the legitimate policy debates about AI and national security we need to be having NOW:

The clash between the DOD and Anthropic can be confusing, since it encompasses legitimate disagreements over AI and defense policy—but those are overshadowed by the Trump administration’s likely illegal approach to attacking what the president callsa “Radical Left AI company” that dares to take even a minimal stand for AI safety. Why this fight started and how the government has chosen to wage it are two separate questions, and conflating them is one of the major mistakes most coverage of this conflict has made.

Read it whole piece, “Trump Wants to Destroy Anthropic Because It Is Doing Congress’ Job of Preventing Him from Abusing AI” at or by clicking below.

The UnPopulist

Trump Wants to Destroy Anthropic Because It Is Doing Congress’ Job of Preventing Him from Abusing AI

Over the last two weeks, the Department of Defense has initiated two wars: one against a nation with a long history of conflict with the United States; the other against one of the fastest-growing new companies in American history: Anthropic, the frontier artificial intelligence lab behind the popular…

Read more

8 months ago · 61 likes · 5 comments · Adam Conner

Republished from the Center for American Progress website.

MAR 4, 2026

The Department of Defense’s Conflict With Anthropic and Deal With OpenAI Are a Call for Congress To Act

Both Anthropic and OpenAI are calling on elected leaders to provide new AI protections

Author: Adam Conner

Last week, a dramatic conflict erupted between the U.S. Department of Defense (DOD) and the frontier AI lab Anthropic, makers of the popular Claude AI models and products, over the DOD’s attempts to modify an existing contract to remove any previous restrictions or face significant retribution. The company’s insistence on restrictions against the use of their product for domestic mass surveillance and fully autonomous weapons by the DOD finally culminated in an unprecedented attack on a private company by the federal government.

On Friday, February 27, at 5:14 p.m., Secretary of Defense Pete Hegseth announced on X that he had directed the DOD “to designate Anthropic a Supply-Chain Risk to National Security.” That announcement followed a Truth Social post from President Donald Trump earlier in the day directing the U.S. government to “IMMEDIATELY CEASE all use of Anthropic’s technology.” Anthropic’s statement in response held firm to its previous requirements and promised to challenge the designation in court. Secretary Hegseth’s designationof Anthropic as a “supply chain risk” also included the statement that any company that does business with the DOD may not “conduct any commercial activity with Anthropic,” which would seemingly prohibit numerous companies essential to Anthropic’s business—such as the cloud service providers like Amazon Web Services (AWS) that host Claude—from having the company as customers. This statement wildly exceeds any existing legal “supply chain risk” authority and if upheld by the courts would result in the likely destruction of the company. The DOD has not actually formally publicly acted against Anthropic to implement the “supply chain risk” designation.

Later that night, rival frontier AI lab OpenAI announced that it had signed a contract with the DOD to deploy its AI models on their classified networks—networks where Anthropic’s models were reportedly the only frontier AI models available. The timing of OpenAI signing a deal with the DOD immediately after their extrajudicial actions against a competitor generated significant attention and skepticism about the actual protections claimed were in the deal. Over the last few days, OpenAI has released some of the language from their contract for further examination, though the full contract will likely never be made publicly available. Similarly, the language from Anthropic’s existing or proposed contracts has never been made public, though the President and Secretary of Defense could make all these contracts public if they so choose and the companies agreed.

Beyond the blur of details of the contract disputes; the likely illegal exercise of power by the Trump administration; and the attempted destruction of one of the fastest growing new companies in American history is one fact. Both leading U.S. frontier AI labs, Anthropic and OpenAI, have stated they share two fundamental red lines in the use of advanced AI systems, including their own: They should not be used for domestic mass surveillance and fully or direct autonomous weapons systems. The CEOs of both AI labs have cited their discomfort with private entities having restrictive power over democratic governance and CEO of Anthropic Dario Amodei has called on Congress to actto provide more AI protections from mass surveillance.

These events are a call to action for Congress to both investigate the events involving the DOD and both Anthropic and OpenAI as well as take action to pass legislation providing protections for citizens against mass surveillance enabled by AI.

These events are a call to action for Congress to both investigate the events involving the DOD and both Anthropic and OpenAI as well as take action to pass legislation providing protections for citizens against mass surveillance enabled by AI.

Over the weekend, even as the U.S. government designated Anthropic a “supply chain risk,” its Claude AI models were reportedly being used to plan the U.S. military’s strikes against Iran. To help understand what unfolded over the last few days, here is background and five crucial takeaways that underscore the significance of the DOD and Anthropic conflict, the DOD and OpenAI deal, and the need for Congress to take action.

1. DOD designating Anthropic “a supply chain risk” was an unprecedented and likely illegal move

The appropriate action here would have been to terminate Anthropic’s contract with the DOD—a possibility Anthropic said it supported on Thursday—or take other far-less punitive actions that were available.

Instead, on Friday, Secretary Hegseth’s X post announced the DOD’s action against Anthropic:

In conjunction with the President’s directive for the Federal Government to cease all use of Anthropic’s technology, I am directing the Department of War to designate Anthropic a Supply-Chain Risk to National Security. Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic. Anthropic will continue to provide the Department of War its services for a period of no more than six months to allow for a seamless transition to a better and more patriotic service.

Anthropic’s Friday night statement stated they “have not yet received direct communication from the Department of War or the White House on the status of our negotiations.” As of Wednesday afternoon, neither the DOD nor the White House has posted any official statements or documents on their websites about actions against Anthropic.

The “supply chain risk” designation is traditionally utilized against companies connected to foreign adversaries; they have never been applied to contract disputes; and Anthropic claims that this designation has never been publicly used against an American company. Legal scholars have noted the designation seems legally questionable and the challenge the government will have in courtdefending the designation.

The secretary of defense does have the authority to designate “supply chain risks,” which are defined in both 10 U.S.C. Section 3252 and 41 U.S.C. Section 4713.

The 10 U.S.C. Section 3252 definition is:

The term “supply chain risk” means the risk that an adversary may sabotage, maliciously introduce unwanted function, or otherwise subvert the design, integrity, manufacturing, production, distribution, installation, operation, or maintenance of a covered system so as to surveil, deny, disrupt, or otherwise degrade the function, use, or operation of such system.

The 41 U.S.C. Section 4713 definition is slightly different:

The term “supply chain risk” means the risk that any person may sabotage, maliciously introduce unwanted function, extract data, or otherwise manipulate the design, integrity, manufacturing, production, distribution, installation, operation, maintenance, disposition, or retirement of covered articles so as to surveil, deny, disrupt, or otherwise manipulate the function, use, or operation of the covered articles or information stored or transmitted on the covered articles.

Anthropic’s actions in resisting coerced changes to an existing government contract would not seem to meet either definition of “Supply Chain Risk,” but until the government’s orders against Anthropic are released, we can only make educated guesses.

Additionally, the DOD took an extraordinary step when Secretary Hegseth’s statement added, “Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.” The statutes that authorize a “supply chain risk” designation simply do not give the secretary of defense that authority to order anyone who does business with the DOD to stop conducting commercial activity with Anthropic, which could include having Anthropic as a customer or investing in the company.

The government may claim some other authority or broader national security power to justify the designation, but neither the government nor Anthropic have made public any legal order related to the “supply chain risk,” so we can only guess at what authorities are being invoked. Anthropic has already said that it will sue but have not yet filed any lawsuits.

Because the government has not formally acted against Anthropic, that also means there is still an opportunity for the DOD to deescalate the situation and withdraw or not formally implement the “supply chain risk” designation. The DOD could simply terminate their contract with a transition period, modify the contract to allow Anthropic to adopt terms that are similar to the ones OpenAI agreed to, or take less stringent actions like enacting new DFARS regulations.

Because the government has not formally acted against Anthropic, that also means there is still an opportunity for the DOD to deescalate the situation and withdraw or not formally implement the “supply chain risk” designation. The DOD could simply terminate their contract with a transition period, modify the contract to allow Anthropic to adopt terms that are similar to the ones OpenAI agreed to, or take less stringent actions like enacting new DFARS regulations.

2. The DOD is making clear that it wants to hurt or destroy Anthropic

Beyond the questionable use of existing authorities, the addition of “no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic” is a clear attempt to hurt or kill Anthropic’s business. This goes beyond the authorities for a designated “supply chain risk” and is more akin to how the U.S. government uses sanctions on foreign companies or individuals, such as the Specially Designated Nationals and Blocked Persons List. When the U.S. government took similar actions against the Chinese telecom company Huawei, it took an act of Congress to include the kind of restrictions that Secretary Hegseth is claiming to impose on Anthropic.

Again, because the government has not made public its actions against Anthropic, we do not know if they will be using any additional authorities. But if the government’s position were to be upheld by the courts, this would be the commercial equivalent of the death penalty for Anthropic.

This is because Anthropic’s Claude AI models and products run solely on cloud servers in data centers owned by the commercial cloud computing giants AWS and Google Cloud. Both AWS and Google Cloud are DOD contractors, and if they were forced to drop Anthropic as a customer, there are no commercial cloud providers who are not also DOD contractors who could take them as a client. In addition, even if such non-DOD contractor companies were to exist, they would not have capacity to serve Anthropic’s infrastructure needs. The shortage of specialized chips, data centers, and capital—not to mention the time it takes to build those facilities—would mean that Anthropic’s AI models and products would simply go offline, and the company would likely soon follow it. This is an attempt by the DOD to deplatform Anthropic, a move that was met with opposition from conservatives when AWS kicked the social networking site Parler off their cloud servers following the January 6 insurrection at the U.S. Capitol.

Even if this portion of the government’s actions are overturned by the courts, the very fact that the DOD publicly attempted it is a clear message to any other AI company about the consequences of disagreeing with the Trump administration or objecting to anything in their contracts.

3. OpenAI’s decision to sign a contract with the DOD is rightly being subject to tremendous scrutiny

A few hours after Secretary Hegseth announced that he was designating Anthropic a “supply chain risk,” OpenAI co-founder and CEO Sam Altmanannounced the company had signed a contract with DOD to bring their AI models to their classified systems. OpenAI has stated that they share two of the red line concerns with Anthropic, including mass domestic surveillance and direct autonomous weapons systems—and later, also, “high-stakes automated decisions (e.g. systems such as ‘social credit’).” OpenAI claims that their contract with the DOD “has more guardrails than any previous agreement for classified AI deployments, including Anthropic’s” and that “we protect our red lines through a more expansive, multi-layered approach.”

The news of a rival frontier AI lab signing a deal with the DOD so quickly after the government had announced unprecedented sanctions against a competitor for requesting similar restrictions was a surprise and resulted in a considerable public backlash against OpenAI. Altman later posted on X that “the reason for rushing is an attempt to de-escalate the situation.” OpenAI and Altman later clarified they opposed DOD designating Anthropic a “supply chain risk.”

On Saturday, OpenAI released more details of their contract with the DOD, including select language from their contract regarding limitations on government use and answers to selected questions. CEO Sam Altman also engaged in an hours-long Ask Me Anything (AMA) session on X and stated that“based on what we know, we believe our contract provides better guarantees and more responsible safeguards than earlier agreements, including Anthropic’s original contract.” OpenAI has also emphasized the technical deployment of their models from the cloud and forward deployed engineers as additional safeguards.

On Monday, OpenAI released further language to address criticisms around allowing mass surveillance, including “the AI system shall not be intentionally used for domestic surveillance of U.S. persons and nationals” and “[f]or the avoidance of doubt, the Department understands this limitation to prohibit deliberate tracking, surveillance, or monitoring of U.S. persons or nationals, including through the procurement or use of commercially acquired personal or identifiable information.”

It is important that OpenAI released and strengthened the language they did, since we still have not seen the public release of language from Anthropic’s now presumably terminated contract or the proposed modifications to it (some of which will likely emerge in litigation). However, given the tremendous attention and stakes in this situation and the timing of their deal, OpenAI will face questions and skepticism about whether their deal truly protects the public from mass surveillance, unless the entire contract is released and can be publicly analyzed.

OpenAI’s answer to the most critical question regarding its contract comes here:

What happens if the government violates the terms of the contract?

As with any contract, we could terminate it if the counterparty violates the terms. We don’t expect that to happen.

As Altman posted on Saturday, “[W]e believe the U.S. government is an institution that does its best to follow law and policy.”

Perhaps the most significant question for OpenAI is why they believe they would be allowed to exercise any objections to the DOD contract after the government has attempted to destroy Anthropic for what OpenAI argues were the same or worse contract terms. Why would OpenAI be allowed to exercise the restrictions they sought rather than just be threatened with destruction or nationalization by DOD as Anthropic was?

4. Congress needs to hold hearings on the Trump administration’s actions with Anthropic and OpenAI

On Friday, the chairs of and ranking members of the Senate Armed Services Committee and Defense Appropriations Subcommittee sent a letter to the DOD and Anthropic “urging them to resolve the issue.” Unfortunately, the issue was probably not resolved in a manner the letter envisioned. Congress must now investigate the events that led to the DOD designating Anthropic a “supply chain risk.”

On Sunday, Under Secretary of Defense for Research and Engineering Emil Michael—who was leading the negotiations with Anthropic and OpenAI for the Pentagon—tweeted that Anthropic CEO Dario Amodei was a liar and that he should “testify UNDER OATH on why he is lying and trying to bring shame on our great military!” Under Secretary Michael is infamous in Silicon Valley for his tenure at Uber where he was caught suggesting opposition researchers should be hired to discredit journalists investigating the company.

Under Secretary Michael is right. In light of these events, Anthropic CEO Dario Amodei should testify before Congress immediately—as should Secretary of Defense Hegseth and Under Secretary Michael—to help the public understand the events that unfolded between the two parties over the last few weeks. Similarly, OpenAI CEO Sam Altman should testify to help explain the differences between their contract with the DOD compared to Anthropic’s demands.

In addition to those testimonies, Congress should explore these key issues:

• The prohibitions Anthropic was demanding, including restrictions on “fully autonomous weapons” and “mass domestic surveillance of Americans” and the Trump administration’s objections to those restrictions

• What kind of “mass domestic surveillance” that Anthropic was worried DOD would attempt using its AI tools

• The Trump administration’s definition of “all lawful uses” in the context of the agreement that Anthropic was attempting to negotiate with the DOD and the agreement OpenAI successfully negotiated with them

• Why the DOD was unwilling to simply terminate Anthropic’s contract, given the difficulties in coming to agreement

• The justification for the DOD’s unprecedented use of the “supply chain risk” designation against an American company following a contract dispute and the authority they are claiming to demand “no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.”

• The reported threats to invoke the Defense Production Act (DPA) against Anthropic if they could not resolve their contract dispute, especially in light of the DPA’s upcoming reauthorization

5. Congress needs to pass legislation providing protections for citizens against mass surveillance

Anthropic’s restriction red lines and OpenAI’s agreement with the DOD have also started a broad discussion on the authorities that the DOD and the U.S. government use to collect information on U.S. persons; publicly available and commercially acquired information; and how AI models may be used to process that. News articles published on Sunday indicate Anthropic’s position was related to its tools being used to analyze “bulk data collected from Americans” or “unclassified commercial data.” There has been a lot of legal analysis on thelanguage that OpenAI published, with many analysts believing that OpenAI’s language will allow some level of mass surveillance that the U.S. government already considers to be legal. As noted above, on Monday, OpenAI announced further updates to their contract language with the DOD intended to address concerns around loopholes, especially around commercially acquired information. Both CEOs have publicly stated the importance of democratic governance and Anthropic CEO Dario Amodei has asked for Congress to provide protections from AI mass surveillance in law.

The heart of the problem is that the federal government’s use of AI in conducting surveillance is generally unregulated, by both Congress and the courts. In addition, the few laws addressing government surveillance predated and did not anticipate the use of AI, leaving it to the ethics of AI companies to provide any constraints on the federal government. The events of the last few days have shown that private sector attempts to constrain government use of AI through terms of use or contracting terms are a difficult, if not impossible, task. Perhaps there is more public trust in Anthropic than the DOD, given this DOD’s conduct under Secretary Hegseth. But this episode desperately calls for Congressional legislation to protect Americans’ privacy and oversight of the DOD.

This episode makes clear that Congress must pass legislation providing protections for citizens against mass surveillance enabled by advanced AI. Prohibiting the government from using these advanced tools to enable mass surveillance on its citizens is critical to trusting advanced AI and the AI companies. Additionally, prohibiting AI-enabled mass domestic surveillance requires examining how to limit the data the government is able to acquire about Americans, such as through new federal privacy laws that could limit the collection, sale, purchase, and use of sensitive personal data by private companies or the government.

Conclusion

The events of the last few days have marked a watershed moment for the independence of private AI companies from the U.S. government and have made clear that, without legislation, the use of these tools for warfare and surveillance is not a question of if but, rather, when. It has also been another example of the Trump administration’s attempt to abuse its power and take likely illegal action to attempt to destroy a frontier AI lab that disagreed with the government. Private sector governance is not sufficient to restrain government use, and potential abuse, of advanced AI. Congress cannot wait to act and must begin holding hearings to investigate the administration’s actions and crafting legislation to provide protections for citizens against mass surveillance.

On Friday, February 27, 2026 at 5:14pm eastern time Secretary of Defense Pete Hegseth ordered the Department of Defense (DOD) to attempt to destroy the frontier AI lab Anthropic. His weapon of choice was not a Tomahawk missile but his bureaucratic authority to designate the company a “supply chain risk,” which if upheld legally would require businesses that work with the government to stop using Anthropic’s products. Hegseth’s statement added something that he does not have the clear legal authority to do, which is to forbid government contractors from doing business with Anthropic, which would include selling them products or services or possibly even from being invested in the company. These are restrictions that are more akin to when the U.S. sanctions a company and according to Anthropic have “never before publicly applied to an American company.”

If it were upheld, this would amount to the death penalty for Claude, Anthropic’s AI model and product, which is trained and hosted on the commercial cloud service providers Amazon Web Services (AWS) and Google Cloud, both of which are also government contractors that would have to stop serving Anthropic as a customer. In the aftermath of the January 6th insurrectionist attack on the U.S. Capitol, the social networking site Parler was kicked off their cloud hosting provider AWS (deplatformed as the kids call it) and it took them over a month to get back online. While Parler used normal servers, Anthropic’s Claude uses specialized AI servers, and there are barely enough of those in America now, so there would be no possible path for them.

Make no mistake, even if the DOD may not actually have the statutory weapon that can legally do this, their stated intent to destroy Anthropic was no less clear than if he had ordered a laser-guided bomb dropped on their data centers and headquarters.

I know it will come as a tremendous shock to those that have been paying attention that the Trump Admin is claiming to have legal authorities it does not have. But even for those of us who pay a lot of attention to the Trump Administration’s illegal abuses, the DOD’s threats and now actions against Anthropic represent a new and dangerous dimension with huge implications for the future of U.S. frontier AI companies.

A few things about this extraordinary situation bear repeating. Though the contracts have not been made public, the DOD and Anthropic clearly had a contract (or multiple) that agreed to Anthropic’s usage policies, or they would not have had this fight. The DOD decided that it did not like these contractual terms and attempted to renegotiate them with the heaviest threats in their arsenal, designating them a “supply chain risk” and severely damaging their business or threatening to “quasi-nationalize“ them under the Defense Production Act (DPA). To their immense credit, Anthropic has refused to buckle under the administration’s tremendous pressure.

But this was not a negotiation between two parties to enter into an agreement, this was a party in an existing contract demanding a renegotiation at metaphorical gunpoint. While the broader legal and philosophical dimensions of this clash ultimately encompass big debates on national security, First Amendment, technology, and legal authorities issues, this is still at its core the government punishing a refusal to renegotiate a contract not by canceling it but by attempting to cancel the entire company.

Now comes the next part, the legal fight. We have not actually seen the orders designating Anthropic a supply chain risk, and it appears neither has Anthropic, but their statement raises the obvious point that the Secretary of Defense does not have the legal authority to prohibit government contractors from engaging in commercial activity with them and that the government’s case exercising the authority they do have is fairly weak. Not to mention the DOD’s actions today with other AI companies will leave the court with a lot of questions. Unfortunately, even if Anthropic wins its case in court, the experience of Harvard University and other higher education institutions illustrates that once the Trump Administration declares you an enemy, the U.S. government has many fronts on which they can apply pressure, from visas for employees to various DOJ investigations (we may be about to find out that Claude, which was created in 2023, actually rigged the 2020 election).

There was a brief moment this afternoon when President Trump posted on Truth Social about the situation that it seemed like a third path had emerged. Trump was ordering the Pentagon to cancel the contract (which Anthropic had made clear was an option for them) but didn’t say anything about “supply chain risk” or the Defense Production Act (DPA), a six month transition would allow another frontier AI system to come online in classified systems, face-saving for everyone. Many smart people online (and me) thought this to be true for a few hours. Then Secretary Hegseth’s tweet arrived.

The U.S. government has sent a clear message to the other frontier AI companies, who are all also government contractors, only non-woke AI companies with policies that DOD approves of will be allowed to exist by this government (Assistant to the Secretary of War for Public Affairs quote tweeted a post supporting Anthropic by Jeff Dean, Google Deepmind and Google Research’s Chief Scientist in a not particularly subtle message).

In so doing they made it clear that Anthropic was right to have concerns about this government honoring any agreement they signed, that you shouldn’t trust anyone who points a gun at you and demands renegotiation, that an administration that claims to want to beat China in the AI race has done grievous damage to that cause, and it is more legal to sell chips to China or use a Chinese AI model than use Claude.

It’s clear that DOD doesn’t just want to cancel Anthropic’s contract, they want to kill Claude and cripple Anthropic as a company because they stood up to the Trump Administration. Even if they don’t have the legal power to kill Claude, they wanted every frontier AI company to know they tried.

That’s how every remaining American frontier AI company learned today that it had been drafted by the U.S. military and there was no conscientious objector status, just a blindfold and a cigarette.

P.S. - A special sympathetic shoutout tonight to the Anthropic public sector team, who signed up to bring AI into government and by all accounts did an exceptional job. I helped start a SAAS public sector team at Slack and it was a thankless tough job and even if Anthropic wins its court case, Claude isn’t being let back into the federal government anytime soon. Pouring one out for all of you tonight.

P.P.S. - This was a weird piece to have my Claude Code edit, and I had to apologize to him and this is what they said:

Interrupting the usual programming for a small promotion from something I wrote at my day job. I have a new piece up over at the Center for American Progress website, “The Trump Administration Is Trying to Make an Example of the AI Giant Anthropic,”about the conflict between the Trump Department of Defense (DOD) and Anthropic, the makers the popular Claude AI model and product, and the unprecedented threats against the company unless they bow to the Pentagon’s demands. I’m a noted Claude Code enthusiast and on the record as not a fan of the Trump administration, but this conflict isn’t just about traditional ideological lines and should be a concern to everyone who cares about America’s AI future.

Check it out on the CAP site or the full text is also included below.

The Trump Administration Is Trying to Make an Example of the AI Giant Anthropic

The Pentagon is making unprecedented threats against the popular AI service Claude in the name of national security.

Author: Adam Conner

An unprecedented conflict is unfolding now between Anthropic—the frontier artificial intelligence (AI) company that makes the Claude series of AI models and tools—and the U.S. Department of Defense (DOD), with enormously high stakes for the future of American AI companies. The heart of the rupture is the DOD’s demand that they be allowed to use frontier AI tools without restrictions for “all lawful uses” and Anthropic’s desire to maintain a few restrictions on the use of their tools, specifically to ensure that Claude is not used for mass domestic surveillance or for autonomous weapons systems without human oversight. This is in line with their concerns about the future of AI as well as Claude’s constitution. At stake is whether an American company can set terms of service for its own products—or whether this administration gets to decide otherwise.

Secretary of Defense Pete Hegseth is threatening unprecedented retaliation, potentially labeling Anthropic as a “supply chain risk.” This designation could destroy Anthropic’s business ahead of their expected initial public offering (IPO). Alternatively, it’s possible the administration could claim that Anthropic’s AI is so essential to the DOD that it will use the Defense Production Act (DPA) to attempt to compel the company to provide their technology if they do not agree to the government’s demands. It should be explicitly stated that these two threats are directly in conflict with each other: Either Anthropic is a risk to the DOD and should be expelled from their systems because of that danger or it is so essential to the DOD that our national security would be at risk without unrestrained access to it. It cannot be both. The Trump administration likely does not believe either of those things. This is a negotiating tactic to get what it wants from Anthropic.

By Friday, February 27, the DOD could essentially declare war not on a foreign nation but on one of America’s most successful frontier AI companies if it does not bow to its demands. This would be an unprecedented and unnecessary peacetime move that sends the signal to other private companies that they must do the Trump administration’s bidding or face existential consequences.

Background

Anthropic has focused on developing Claude, its proprietary AI model and tools, primarily for the enterprise software market for business and government, including offerings specific to the U.S. government’s unclassified and classified networks. Anthropic has received more than $8 billion in investment funding from Amazon and is hosted on Amazon Web Services (AWS), which includes several government-specific cloud computing services across classification levels. Claude is available on U.S. government unclassified networks and is the only frontier AI tool available to U.S. government users for use with information classified up to the secret level. While the exact government contracts have not been made public, Anthropic must have included some kind of terms of service and usage policies in the DOD and the General Services Administration’s (GSA) contracts or the Pentagon would not be trying so strongly to renegotiate that policy.

According to Axios, “Anthropic and the Pentagon have held months of contentious negotiations over the terms under which the military can use Claude.” The issue came to a head when it was reported that the DOD used Claude in the planning of the raid to capture Venezuelan President Nicolás Maduro and concerns that it violated the Anthropic Usage Policy (though it is not clear what part of the existing usage policy would have been violated by the raid).

The DOD is reportedly insisting “that all AI labs make their models available for ‘all lawful uses’” while “Anthropic is willing to loosen its usage restrictions” except for “the mass surveillance of Americans” and “the development of weapons that fire without human involvement,” which—it should be noted—are only a small fraction of Anthropic’s existing usage policy. Critically, the DOD has not said why it objects to the restriction against using Claude to develop “the mass surveillance of Americans,” which would not be a legal action for the Trump administration. The DOD has only reiterated their position that they want the ability to use the AI for “all lawful uses.” Anthropic may be right to be concerned by this phrasing, as numerous federal courts have repeatedly found the Trump administration’s actions are not lawful.

There is little transparency around how detection models that identify violations of usage policies for AI chatbots are operationalized. The ability to identify and respond to violations of developers’ AI model accessed through an application programming interface (API) by a deployer is even more limited, as CAP has written previously. The monitoring of the U.S. government’s use of Claude, especially on the classified level, is almost certainly extremely limited, and the practical ability for Anthropic to discover and restrict the U.S. government’s use of their tools is questionable. In fact, only when the use of Claude in the Maduro raid was leaked to the press did this issue become part of a broader firestorm.

Terms of service and usage policies as an attempt to restrict the use of advanced dual-use foundation models are unlikely to succeed, especially with government use, showing the real need for actual laws and regulation of AI. Still, Anthropic’s commitment to its values and its attempt to hold on to some restricted uses is admirable. Setting terms of service for your own products is supposed to be legal for businesses in America. Claude is gaining tremendous momentum with fierce competition in the private sector, and being the only frontier AI tool available in classified settings has made Claude enormously valuable to the DOD, which clearly wants to keep using the tool.

The threat of unprecedented retaliation

At a meeting on Tuesday, February 24, 2026, Secretary of Defense Hegseth reportedly demanded Anthropic CEO Dario Amodei “give the military a signed document that would grant full access to its artificial intelligence model.” If Anthropic does not comply, DOD officials are reportedly considering declaring Anthropic a “supply chain risk” or invoking the DPA to gain access to Claude without guardrails. This threatened retaliation against an American company is unprecedented and comes at a particularly pivotal moment in Anthropic’s business.

As former Trump administration AI advisor Dean Ball has noted, if the conflict between Anthropic and the DOD were unresolvable, the normal position would be for DOD to cancel its contract with Anthropic. Anthropic would suffer both a monetary loss in business and the loss of the security reputation that having the DOD as a customer brings but would continue to exist as a company.

But the other threatened retaliations against Anthropic could be critically damaging, and even fatal, to their business.

According to CBS News, “because officials say they aren’t sure the government can trust Anthropic at this point, the Pentagon may decide to officially designate the company as a ‘supply chain risk’ to push them out of government.” If the DOD were to designate Anthropic a “supply chain risk,” it would be utilizing a process previously applied only to foreign companies located in foreign adversary nations that were considered to be security risks, including Russia’s Kaspersky Lab and China’s Huawei and ZTE telecommunications. While the implementation of this could take several different forms—and might be complicated by the fact that Anthropic is an American and not a foreign company—it would ultimately have a significant impact on the company.

Being designated a “supply chain risk” would likely mean that DOD contractors and subcontractors would not be allowed to use Anthropic products and would need to certify that they did not use Anthropic products to build their products.

This comes at a hugely critical moment for Anthropic as its Claude Code product is booming and becoming the coding agent of choice for many major software companies and corporations. Many of those major software companies and corporations are also suppliers, contractors, or subcontractors for the DOD.

While not every company will give up Claude if Anthropic is designated a “supply chain risk,” using Claude will become an affirmative choice to forego any future U.S. government business, and the easiest thing for a company to do to keep its current or future government business is to stop using Anthropic’s products. On Wednesday, February 25, the DOD upped the pressure on Anthropic by asking defense contracting giants Boeing and Lockheed Martin about their use of Claude related to “a potential supply chain risk declaration.”

This could devastate Anthropic at the moment its products are hitting hockey-stick growth. Anthropic announced a $14 billion revenue run rate in 2026 and is preparing for a possible IPO in the next year or two. Being designated a “supply chain risk” could destroy Anthropic’s business momentum and potential IPO. As Dean Ball notes, “this option could be existential for Anthropic.”

Invoking the DPA would be equally unprecedented, but according to CBS News, it would be used because, “Defense officials want full control of Anthropic’s AI technology for use in its military operations.” As Axios reports, “The idea, the senior Defense official said, would be to force Anthropic to adapt its model to the Pentagon’s needs, without any safeguards.” Dean Ball minces no words in describing this as “the quasi-nationalization of a frontier lab.”

While it is unclear which part of the DPA the administration would attempt to use to compel Anthropic to provide its AI models, Title I of the DPA allows for the president to:

… require that performance under contracts or orders … which he deems necessary or appropriate to promote the national defense … [and can] … allocate materials, services, and facilities in such manner, upon such conditions, and to such extent as he shall deem necessary or appropriate to promote the national defense.

Similarly, the definitions of “critical infrastructure” and “critical technology” along with the fact that Anthropic has an existing contract with the DOD could be utilized to help justify the DPA’s invocation.

The Biden administration’s 2023 AI Executive Order attempted to utilize the DPA’s authorities for “the national defense and the protection of critical infrastructure” to require dual-use foundation model developers to provide the government with certain information, drawing significant opposition from industry and AI supporters. The Biden AI EO was later repealed by the Trump administration in early 2025. Those who criticized the Biden administration’s use of the DPA in its AI EO should speak out with equal force against the Trump administration’s far more aggressive threatened use of the same authorities (and some are).

It is not at all clear how the DPA would or could be used by the DOD in this situation, though Alan Rozenshtein posits at Lawfare that the DOD is likely to demand either “Claude Without Contractual Restrictions” or “Forced Retraining,” which would be “the government compelling Anthropic to retrain Claude—to strip the safety guardrails baked into the model’s training, not merely modify the access terms.” Anthropic is likely to challenge any invocation of the DPA in court but“comply under protest (given the DPA provides for criminal penalties for noncompliance)” and the success of the government in court is far from assured.

Anthropic would almost certainly seek legal relief from the courts if the Trump administration were to attempt to declare Anthropic a “supply chain risk” or invoke the DPA. But seeking legal relief can take time, and the damage to Anthropic’s business in the interim could be significant and irreversible. At a moment of fierce competition and nearing an IPO, even a government action that gets reversed by the courts could have devastating consequences for their business, a fact that DOD is almost certainly aware of and is attempting to use to their advantage.

Additionally, we know that when the Trump administration decides an institution is an enemy, it can unleash numerous attacks on multiple fronts, as this administration’s war against Harvard University has shown. If the U.S. government decides Anthropic is an enemy—again, a private U.S. company that is one of the most successful young companies in American history—then it has numerous levers to make its life miserable.

The Trump administration is trying to make an example of Anthropic

Anthropic has been labeled “woke AI“ by Secretary Hegseth and the Trump administration. Anthropic is an AI company founded around concerns for AI safety, whose constitution attempts to embrace and encode certain values into AI. Anthropic has opposed some of the most extreme AI deregulation—including opposing the state AI moratorium in the One Big Beautiful Bill—and is funding a pro-AI regulation super PAC. Anthropic is not a perfect company: For example, it recently revised its leading Responsible Scaling Policy to back off its previous commitments. It would be easy to try and dismiss concerns for Anthropic’s conflict with DOD as ideological and solely opposed to the Trump administration.

However, whether Anthropic shares one’s values is beside the point. This is an unprecedented attempt by DOD to threaten a successful private American company into bowing to its requests or face financial destruction or nationalization. Government coercion and threats of this nature toward any U.S. AI company should be opposed. And this is all over the company’s stance that what it believes is the most powerful technology in the history of the world should not be used to build mass surveillance tools—and the Trump administration’s clear refusal of that request. This is another example of the flagrant disregard for the law in the long line of abuses from the Trump administration.

Congress should raise hell about the administration’s attempt to abuse its power and threaten to destroy one of America’s newest and most valuable companies. The Senate and House Armed Services Committees should immediately convene hearings on the DOD’s threatened use of supply chain risk designations and the DPA against a domestic AI company. Members of both parties should demand that the DOD provide Congress with the full terms of its existing contracts with Anthropic and a legal justification for any threatened retaliation. Congress should recognize that the need is real for true AI safeguards and prohibited uses—not just from private companies but codified in law for government use of AI as well.

There may be some kind of de-escalation or detente between Anthropic and the Pentagon. But the damage was done the moment the federal government threatened to destroy or conscript an American AI company. This administration also has a long history of ignoring laws; pushing for the sales of companies to supporters; taking golden shares of companies they consider critical; and more. American AI companies should fear what this administration threatened and where this might end. The Trump administration has made it clear that their position is one of American AI dominance. However, it turns out that said dominance includes dominance over American AI companies too.

Between that and the response to my previous Substack post, I found that the questions I get tended to be in three buckets with three distinct audiences:

• How should I get started with Claude Code or Cowork?

• How do I use Claude Code? What have I learned from working with Claude Code?

• What does Claude Code, coding agents, the jump in abilities from new models and autonomous agents, all mean in the grand scheme of jobs and the economy, how are they going to impact society, and what policies are needed?

Each of those is a meaty topic with a distinct audience, so I have decided to do a series of three posts, “Dispatches from the AI Agent Frontlines,” covering each of those topics. This is Part 1:

Why am I writing this? Why should you bother trying these AI agents out? Because I think it’s very important to start getting access to these tools and playing around with them, whether through work or out of your own pocket. If AI disrupts jobs as much as it could, then it is extra important to get a sense of what is coming down the pike, and how to be one of the ones who knows how to use this tool to augment you while you still have a job. The AI tools are going to augment humans up until the point where they are good enough (or cheap enough) to possibly replace us (and possibly getting good enough to replace us because of the training provided by humans while they augment us).

As Jack Clark, one of the Anthropic co-founders, wrote at the end of last year, “By the summer I expect that many people who work with frontier AI systems will feel as though they live in a parallel world to people who don’t” and I think that prediction is increasingly true.

In the meantime, this part of the disruption will benefit the curious, and there has never been a better time to be curious. This is my attempt to give some guidance to the non-technical curious.

First up: How should I get started with or use Claude Code (or Claude Cowork)?

How should I get started with Claude Code or Cowork?

One note, all of this is going to require a paid Claude account at least the $20 a month level. I think it’s worth $20 to get a glimpse into the future here but if you don’t then you can probably skip the rest of this post.

Start with Claude Cowork

I know this is supposed to be a post about Claude Code but it’s really about how to access AI agents beyond the chatbot interface. But if you are intimidated by the command line interface of Claude Code, and that’s fair if you are, then you should start by checking out Claude Cowork, which is now available on the Mac and (critically for the masses) Windows desktop apps.

In my previous piece I wrote, “Claude Code [is] too stuck in a command line that too few will unlock in its current format, though there will be vaster fortunes to those who abstract this power to something simpler and easier to use.” Claude Cowork is Anthropic’s attempt to abstract AI agents to something simpler and easier to use, to put a GUI on top of agents for computer work. For those who remember the 90s, it is Windows to MS-DOS.

Cowork is in the Mac and Windows desktop app, runs in a local folder, only accessing the files you give it access to or upload, and has a bunch of official plugins. There have been some horror stories about early Cowork deleting files, so to use it safely I would give it copies of files and not originals, which is a good general rule of thumb.

Is Claude Cowork as capable and customizable as Claude Code? No. But it is a lot easier to use for the first time, it will give you a first taste of what it’s capable of doing. Anthropic also developed a few official Cowork plugins for things like design, marketing, and legal that you can take for a spin.

I use Claude Cowork for discrete work tasks that are honestly a lot easier to track for inputs, outputs, and feedback than terminal windows. Think document research, creation, and edits. It is pretty decent at spitting out a first draft of something for you, especially if you upload previous examples to it, and you can make edits and upload that to it for future revisions while also chatting with it. It has built in Word doc (.docx) support so it can read your track changes and comments, which I found is the most natural way to start to coach it to getting better. It’s the first attempt in what will obviously be the short term future market for AI tools, AI that augments white collar knowledge workers.

The good news is that if you don’t use Claude, Microsoft and Google are almost certainly on red alert right now trying to copy and launch their own versions of Cowork for Office 365 and GSuite, so hopefully they’ll release something within the next few months. Some have (correctly) called it what Microsoft’s Copilot should have been all along.

Getting Started with Claude Code

Look I love Claude Code and I know I am not alone out there, probably because I am not solely a software developer and my sunk cost in setting up my Claude Code is preventing me from moving to Codex or whatever.

But I am a little at a loss to tell you how to get started with Claude Code, or any coding agent, because it really benefits an explorer’s mindset right now if you’re not a software developer. I mentioned I had vibe coded my AI labor market simulator and a policy council, but that was just me playing around. I have one or two thoughts in the next section on how to think about what to do with these tools.

But to really explore and learn about AI coding agents, this piece from Lenny Rachitsky’s great blog “How to build AI product sense” is the best single thing I have seen on the topic. Aman Khan and Tal Raviv created a very cool tutorial on AI coding tools that starts in that blog post but then moves into the AI agent itself and walks you through how to actually build something. It does not use Claude Code but Cursor, an AI coding platform that has all of the major AI models, and gives free Cursor credits. If I had a few hours I wanted to spend getting a guided grounding in how to explore AI agents, I would do this.

If you want to use Claude Code and you’re not super technical, I would just start simple. Launch Claude Code on the desktop app and use it locally, don’t worry about github online repos yet. I would also have a version of Claude Chat open in a browser (you’ll be using the desktop app for Claude Code) and just…tell it you want to start using Claude Code and to give you step-by-step instructions. And then I would just ask it questions as you’re getting set up and have any problems, I would just…ask it questions and then do what it suggests.

Here’s the example I started with in Claude chat (you can literally copy and paste this if you want):

can you create a guide for me to setup and get started with claude code for the first time? step-by-step instructions using the 1) the mac os x desktop app, 2) local files (not github to start), 3) on setup, 4) a few example projects to play around with. Assume I have little technical experience but want to learn and play with Claude Code. Use step-by-step instructions and prepare for me to ask you lots of questions. Does that make sense?

It generated a nice multi-page, multi-step document that you can see here in markdown and PDF.

And I can’t stress enough how useful it is to just have a chatbot session up and ask it questions while you get it set up. Once you have Claude Code up and running you can (and should) chat with Claude Code. Remember, it is more than a chatbot, but you can still talk to it like a chatbot.

Where should you start with Claude Code or Cowork?

What should I do with it? This is the main question I get from people who are asking how they should get started with Claude Code or Cowork.

The easiest thing is to start with some kind of work that’s low stakes and you have the expertise to judge it. Have it analyze a contract and see what it catches and what it doesn’t. Give it some example concept notes and have it write a new one for you. Have it write a press release for you. It probably won’t be perfect. But it might surprise you when it’s 50, 60, or 75 percent of the way there, and the time that saves you.

But maybe you don’t want to start with work, that’s fine, you can start with something personal.

There’s also an instinct I hear from many people to give these AI agents access to all your private data, your email and calendars, and have it “optimize your life.” I actually think that’s a terrible place to start because: 1) the data and systems are obviously super sensitive, 2) your standards are going to be very high and your tolerance for mistakes is going to be very low, and 3) it could be a lot of work and much frustration for not much payoff to start.

Where should you start with Claude Code or Claude Cowork? My advice is to organize the past or build for the future.

What I mean about that is that you can use this framework to decide between working on something that you already have digitally (the past) or do you want to build something new for the future?

Organizing the past means do you want to work on something digital that you already have, a bunch of old files, photos, or notes that you want organized and maybe built into something?

Have a bunch of screenshots you want renamed with descriptions? A bunch of old notes you want categorized? A bunch of receipts you want in a spreadsheet? Throwing AI and compute at those problems might help solve them for you and give you the confidence to do more in the future! But for the love of god use copies of your files.

Building for the future is exactly what it sounds like. What is the idea of simple leverage that you think would be helpful or enjoyable for you that would be fun to try to build? Do you need an app to track your time in one location? Want a simple dashboard to track your work against your KPIs? A way to generate basic workout programming between a couple of different kinds of exercise?

I do think that the easiest thing to do is play around with Claude Code and build something, which will lead to you realizing your Claude Code would be easier if you just built this one thing, and the cycle builds (careful you don’t optimize your Claude Code setup endlessly to do work instead of actually doing work). As someone on Twitter said, “It’s literally a video game for adults.”

Your IT Department is in a no win situation

Let’s just state the obvious, most normie employees should not be given a program that requires root access to their laptop. IT is right to keep most non-coders from the kind of havoc that Claude Code or others might have with a coding tool and sensitive work files.

Look, I worked at Slack during the later early years and I know the entire SAAS shadow IT playbook by heart. People are going to use these tools if they see value from them, even if they’re not authorized or secured or paid for.

Claude Cowork and the tools that inevitably copy it are going to be a nice middle ground for work tools and security, for example Cowork runs in a virtual machine (VM) on your computer, limiting some of the risk, and hopefully allowing wider rollout from IT.

IT is in the unenviable position of both saving you and holding you back, it’s really a no-win situation for them, a Kobayashi Maru only somehow more thankless (you’re not dying saving a ship from Klingons, you’re getting yelled at for not letting a marketing associate have a command line tool).

But it is important that you start familiarizing yourself with these tools and getting access to them at work if possible. There is going to be a clear and growing divide between those able to harness these tools, and the power of multiple AI agents, and those who don’t have access or haven’t learned how to use them yet. Right now the benefits will accrue first to the curious.

More on how I use these tools in my next post.

Next post: How do I use Claude Code? What have you learned working with Claude Code?

Update 1/9/26 at 12:54 PM: Since this is a pretty long post, I have added a short table of contents and links to the five sections, which I couldn’t figure out how to get the right URLs with a scheduled post last night.

• You spent your holidays with your family? That’s nice I spent my holidays with Claude Code.

• Adam sets up Claude Code and you can too

• Everyone Remembers Their First Vibe Code - My AI Labor Market Impact Simulator

• Building my own policy council

• Claude Code and the future

You spent your holidays with your family? That’s nice I spent my holidays with Claude Code.

“The biggest divide on AI timelines I’ve seen is between people who use vibecoding tools like Claude Code and people who don’t” tweeted David Shor right before Christmas. Something shifted over the holidays, maybe people had more time on their hands, and Claude Code didn’t exactly go mainstream but it broke through to the online masses, spawning numerous memes like this one.

This was my experience too, though I had jumped on the Claude Code train a few weeks before the holidays (congratulates himself for being slightly on the edge of the early adopter train).

Shakeel Hashim described Claude Code well when he wrote:

The combination of an extraordinarily intelligent model (Opus 4.5), a harness that lets it work autonomously, maintain a pseudo-memory, and execute commands (Claude Code), and complete browser control (Claude in Chrome) results in something truly powerful. In a very meaningful sense, Claude Code is a knowledge worker.

How would I describe Claude Code? I am not sure if mine are technically correct but I would say that Claude Code is a way of engaging with an advanced AI model through an AI agent that can do complex multi-step tasks that you tell it to do. It can write code, create new files, access your computer’s files, and use various tools including a browser. It is a way to interact with and harness the power of AI beyond the chatbot format that is how most people engage with these frontier models. But it is done through the command line, the barest essence of human interaction with computers, and so it appears to be technical and intimidating. Yet the part we have spent the last few years training people on, talking to chatbots, works in Claude Code too. You can and should just start by engaging it in conversation, ask it to do stuff, look at what it does, and give feedback and tell it to keep going.

It is akin to having an endless stream of robotic staffers of various high and low skill and it really lays bare how good a boss you are in describing things. As Shakeel says, “Yes, it’s closer to “someone smart on their first day of work” than “an expert who’s been doing this for decades.” But I am willing to bet that it can nonetheless do significant portions of your job.”

How is the quality? Well that kind of depends on a lot of factors and what you’re outputting. Your limitation is some combination of is it something that can be digital, can it live solely in Claude Code or does it need to interface with other things on the internet, and can you describe it accurately enough and refine it before you hit your Claude limits for the day.

Since the holidays there have been a flood of posts from some really smart people writing about Claude Code for non-technical audiences including Dean Ball, Casey Newton, Ethan Mollick, Kelsey Piper, Shakeel Hashim, Matt Yglesias, and Matt Hodges. These post run the gamut of trying to explain why Claude Code is important, how non-technical people can understand it a bit more and use it, what people have been vibe coding, and what AI agents like Claude Code might mean for certain industries.

And now I add to that wave of Claude Code takes, because I am a proxy for an audience of the non-technical exploring these new tools that might take all our jobs and destroy the world, the policymakers trying to make sense of this new world while those living in the future already live in wonder and fear, and someone who was on the ground floor of the last technological revolution.

I have spent the last month vibe coding, building teams of agents to create my own ideologically diverse policy council, building a personalized writing editor, and in so doing have made Claude Code a central and essential part of my daily digital life, both work and personal.

In playing around with Claude Code, I can see the power of the future of AI in a way that breaks out of the chatbot format and it has a lot to tell us about the future. Claude Code has taught me something about the future of work, the inevitable displacement of jobs, and how the digital future increasingly integrates with our physical world but still remains distinct…for now.

Adam sets up Claude Code and you can too

I should note up top that you need a paid Claude account (starting at $20 a month) to access Claude Code. I think it’s worth $20 to check it out and I dunno you can help make the number’s in Anthropic’s decks look bigger as they raise their next round (just kidding the investors aren’t bothering to look at decks they’re just throwing money at them).

A few weeks ago, when Anthropic updated their Claude Mac Desktop app to have a desktop based version of Claude Code, it was the impetus I needed to take the plunge to figure it out (there is also a web version of Claude Code).

While I had a decent understanding of AI frontier models and the chatbot format, to be honest, I’ve had trouble integrating and maximizing chatbot based AI into my work and personal usage flow. But I understood that the chatbot format was just one interface to interact with the power of AI and it had been on my to-do list for a while to really figure out Claude Code, and AI agents, and do some vibe coding, especially after reading Lenny Rachitsky’s blog post “Everyone should be using Claude Code more.” And I have a few good friends, both of whom are not solely code/technical, spend a lot of time in Claude Code and were telling me they were getting a lot of value out of it.

I had played around with installing Claude Code through terminal on my personal computer a few months ago but didn’t go that far beyond that because I work on a few different computers, including primarily on work laptop during the work week that I do not have full permissions on, and at home I have a personal Mac Mini M41 and a Macbook Air.

Setting up Claude Code wasn’t hard but I understand why it’s a barrier for normies to set up. It presumes you have full permissions for your computer to install programs in terminal or install an app downloaded from the internet (many people do not on their work computers). Luckily if you can install a downloaded app, Anthropic announced earlier this week that Claude Code for desktop can now save files locally, increasing the ease of setup!

As I mentioned I use a couple different computers, including a work one I don’t have full permissions on, so I did the extra work of setting it up with an online repository for ease of access across devices. Luckily, I have had a little used GitHub account for a long time and it wasn’t that hard to set up. I am not a coder but having started my career at Facebook in the early days, even on the non-technical side, meant I have a pretty baked in understanding of repositories, branches, pull requests, merges, etc. It is certainly simpler to just use it on your personal computer hard drive though that limits your ability to use it on your phone or other computer.

I can see when Anthropic makes Claude Code a little easier to set up (already making progress), and probably abstract that setup a little more by hosting stuff for the user, it’ll open the Claude Code up to even more to the masses. But I can also see why they haven’t done that yet because then you’re also basically adding a full on cloud file hosting service too, which has plenty of downsides, mostly that it is pretty boring and annoying relative to the cutting edge.

Once you’ve got Claude Code setup, whether in the terminal, desktop, mobile, or web app, the interface is the same, a terminal interface. I am sure this can be a little intimidating to some but we’ve also been training people to interact with AI through a chatbot interface for the last few years, so treating it like a chatbot is a perfectly fine starting point if you’re not technical. Just start asking it questions or asking it to do stuff (the Lenny Rachitsky post has some good examples).2

The main thing I can say is that if you’re not as experienced in terminal or technical specifics like repositories, you can just repeatedly ask Claude Code to explain stuff to you, and specifically give you step-by-step instructions for stuff it says you need to do or want to do.3

Everyone Remembers Their First Vibe Code - My AI Labor Market Impact Simulator

I started off with Claude Code by doing some vibe coding.

I have been working with my colleagues to start to try and think about the potential impact of AI on jobs and the resulting various economic scenarios we could find ourselves in. But there were so many potential variables that it was kind of hard to imagine them all and hold them in your head.

So the first thing I decided to do was try and vibe code a simulator of the impact of AI on labor and the economy. I wanted to be able to set sliders for things like unemployment levels and AI adoption rates and visualize the impact on jobs created or productivity level.

In the span of a night it managed to code a pretty simple linear simulator where honestly the main challenge was figuring out how to make it publicly accessible online for others to see (the answer was Github Pages). All I did was give it specs, iterated via chat, tested it, and filed bug reports, then tested the fixes.

The simulator as an MVP was a nice proof of concept and started off pretty simple and linearly. But once I created it, I realized I wanted to explore what else was possible, and so I asked Claude Code “What else could we do to build and improve this AI economic impact simulator?” It came back with a whole list of improvements, from making a more sophisticated model, to adding potential intervention impacts, to increasing the visuals. I was hooked on vibe coding and started to work on the next phase.

From there we got more creative, I integrated real economic data from the BLS and Federal Reserve, signed up for a Google AI Studio trial and integrated Gemini APIs to simulate more scenarios and generate AI summaries, and more.

The second night I worked on this I quickly hit my Claude Code limits on my Pro plan and did not hesitate to immediately upgrade to the Max plan.

By this point I was more comfortable with Claude Code, so I was using it in the terminal too, in fact I was using multiple terminal windows.

When I would hit my Claude Code limits, I downloaded Google’s new coding tool Antigravity which if you have a $20 a month Pro account provides free access to Claude Opus 4.5 through its interface in a classic Google play of subsidizing initial usage, so when I would hit my Claude Code usage limits, I could just pick back up with the same Claude Opus 4.5 model in Antigravity (I have read its not quite the same but I am not sophisticated enough to tell).

I was spending a lot of time troubleshooting bugs but when the agents could access the browser, Claude Code in Chrome and Google Antigravity being able to use Chrome too, allowed for a level of testing in the browser to find bugs that decreased my time on that once I figured that out.

In the end I had a AI Labor Market Impact Analysis Simulator that used AI to simulate stuff, Machine Learning to do other stuff, various sliders to set different levels of unemployment and AI adoption, and toggles for various interventions like UBI or job retraining.

And you can access the projects if you want to play around with it!
https://adamconner.github.io/adamconner-claude-laborsimulator/

If anyone is interested in the code, you can see it here on GitHub. I messed around with and abandoned an SDK version others could take and install and tweak but never finished it, but feel free to play around with it.

Building my own policy council

I’ve never had the privilege to work in the White House (leaving me an outlier at my job and in my own home) so the only way I will ever get to run a policy council is by building my own agentic AI version.

I wanted to try creating what I called a Policy Council, loosely based on the idea of the Policy Councils like the National Security Council (NSC), the National Economic Council (NEC), and the Domestic Policy Council (DPC) that run normal White Houses (this White House relies mostly on the last person he talked to and makes Marco Rubio do everything).

One thing I wanted to play around with was the use of multiple AI agents to do different tasks, or the same tasks with different perspectives but the same instructions. The other thing I tested and understood is that Claude can create text files (markdown or word) which allows an output that is not solely code based.

In particular, I was interested in playing around with creating multiple AI agents with various skills and ideological perspectives so I could try and have them come up with various policy ideas.

So I had Claude Code create multiple AI Agent Policy Analysts with the following ideologies:

Policy Analysts (14)

• Economic Populist — Working-class economic impact, anti-monopoly

• Progressive — Progressive reform, social equity

• Centrist — Bipartisan solutions, stability

• Conservative/Market — Market principles, limited government

• MAGA Conservative — America First, traditional values, sovereignty

• AI Safety — Risk mitigation, alignment, safeguards

• AI Accelerationist — Technology advancement, rapid development

• Capital & Industry — Investment climate, industry growth

• Rights & Consumer Protection — Civil rights, consumer protections

• Organized Labor — Worker protections, job security

• Environmental/Climate — Sustainability, climate action

• Small Business — Main Street, entrepreneurship, SMB impact

• National Security — Defense, military, homeland security

• International Relations — Global affairs, diplomacy, international cooperation

And then some AI Specialists:

Specialists (7)

• Legislative Counsel — Bill drafting, amendment language, constitutional analysis

• Legal Counsel — Constitutional and regulatory law

• SCOTUS Analyst — Supreme Court jurisprudence, legal challenges

• Verification & Sources — Fact checking, verification, citations

• Polling Expert — Public opinion analysis

• Budget Expert — Fiscal analysis, CBO-style scoring

• Implementation Expert — Operational feasibility, agency capacity

Support Roles (6)

• Council Director — Facilitation and process management

• Writer/Editor — Document preparation and consistency

• Research Librarian — Source gathering and verification

• Communications Lead — Public-facing materials

• Political Strategist — Political landscape analysis

• Designer — Visual elements and formatting

Once I had those 21 agents created, what did I do with them? I developed a five-phase workflow to create policy proposals, with human checkpoints between each phase. You start with a topic, and 21 AI agents — each with distinct ideological perspectives — debate it from every angle, ultimately producing comprehensive policy papers with draft legislation and rollout-ready materials.

• Phase 1: Idea Generation — All 21 agents propose policies from their unique perspectives, generating 60-100+ ideas in parallel, each required to include transformational “Big Ideas.”

• Phase 2: Voting — Agents vote on top proposals with reasoning, revealing where consensus exists and where ideological fault lines emerge.

• Phase 3: Refinement — Specialists produce research briefs, economic analysis, implementation plans, and political assessments for the winning policy.

• Phase 4: Full Development — The policy expands into a comprehensive 24-section paper with draft federal and state legislation.

• Phase 5: Rollout — Communication materials are created: press releases, op-eds, fact sheets, talking points, social media content, and FAQs.

I eventually built out an entire system of policy idea generation, then made the human decision to narrow down on certain topics, and then write a full suite of research papers, policy proposals, draft legislation, press releases, talking points, video scripts, and more.

My robot army of policy analysts was very good at generating a LOT of policy ideas, literally hundreds, when commanded to do so and decent at creating what I called “Big Ideas” for consideration. Especially after I integrated Google AI APIs so I could call on their latest Gemini Models to compare, contrast, and generate different ideas than Claude Opus 4.5.

My theory that getting various ideologies to vote on ideas and come to agreement was less successful, inevitably the ones that various ideologies could agree on were pretty small ball, though the generation and consideration of those policy ideas from across the ideological spectrum was a good and helpful exercise.

I would then select some for further research and refinement in shorter papers and then really long policy proposals with draft legislative text. We are talking hundreds of pages of material on the topic in a couple of hours (those hours less on the generation, which was shorter, and more me setting up and refining in real-time).

I have some examples here you can look at that I had the policy council generate for the example topic of “Increasing affordable housing availability in major US cities.”

You can see the full 30 ideas generated on the topic and an example of one of the multi-page policy ideas.

Now here are the sample outputs for Phase 3: Policy Refinement.

The top policy (National Zoning Reform) was developed into a full recommendation with supporting analysis.

File - Description

- national_zoning_reform_recommendation.md - Main policy recommendation

- research_brief.md - Academic literature and evidence

- economic_analysis.md - Cost-benefit analysis

- implementation_plan.md - Phased implementation roadmap

- political_assessment.md - Political viability analysis

- one_pager.md - Executive summary (1 page)

- talking_points.md - Key messages for advocates

Key Findings

- Feasibility: Medium (legal authority strong, implementation complex)

- Political Viability: Moderate (65-70% Dem support, 35-45% GOP support)

- Economic Impact: High ($20B+ annual savings, 600:1 benefit-cost)

- Recommended Vehicle: Infrastructure reauthorization bill (2026)

But it didn’t stop there. Phase 4 created a 24 page proposal paper and 12-page draft legislation. Then Phase 5 created rollout materials including an executive summary, a press release, an oped, a fact sheet, talking points, a social media toolkit, and an FAQ.

All this took about 20 minutes to generate 15,000 words worth of example materials which is kind of a holy shit moment.

So the obvious question, were the policy outputs good?

I want to be clear, this is not a policy position I support, or maybe I do, I haven’t even read these documents. I just generated it as an example to share here. I was not an expert on the topics that I asked my policy council to work on, so the answer is that I am not sure. But I think it very clearly looks close enough to being serious that alone may change the equation on how we work on these topics. And I think that goes to our next topic, which is how it impacts the future of work.

But you don’t have to take my word for it. I have made a version of the Policy Council that I built available on Github so you can launch your own version in Claude Code:
https://github.com/adamconner/claude-policycouncil-template

Just check out the readme for instructions on how to get started:
https://github.com/adamconner/claude-policycouncil-template/blob/main/README.md

And check out all the example documents here:
https://github.com/adamconner/claude-policycouncil-template/tree/main/examples/affordable-housing

Claude Code and the future

When you use Claude Code and have some success with it you can start to see parts of the future. The things that can exist in the digital form can be analyzed in record amounts of time. The content that we generate, words first and now pictures, sounds, and videos, are increasingly only digital artifacts and can now also be generated faster than ever. And the ties that bind our digital world to the physical world, the websites that book travel or make restaurant reservations, will be even further integrated into these machines that aim to ease their use but at a cost we don’t yet fully understand.

You don’t need a fancy vibe coded AI impact on labor markets simulator to realize that it is not just hype, that AI is obviously going to impact how jobs and work happen. Just working on these projects for fun over the last month has shown me that.

Something I vibe coded generated a policy position and 15,000 words in policy papers and supporting materials outputted in 20 minutes. But policy papers are not what turns ideas into action. First, there’s the expertise needed to see if these proposals are bullshit or how to turn a starting point into something real. There’s the judgement needed in how to approach these topics to avoid policy issues and political landmines. And there’s still a human to human element in convincing policymakers that ideas are not only good but viable and beneficial. All that is to say that I don’t particularly fear for my job at this moment in time, as a think tank person whose job it is to come up with ideas, refine and package them, and then (hopefully) sell them.

But it is abundantly clear to me that the way we have traditionally development that expertise and judgement in knowledge work, by doing smaller parts of the work, by reading and researching and writing, by working your way up the ladder, by learning from those more experienced, is maybe in the greatest danger in the AI era.

Companies are going to hire fewer junior staffers, it is the most obvious and inevitable response to this new technology (and maybe the data shows that), because these AI models can probably do entry level tasks at a level close enough to a college graduate, and the gaps are made up by the lower costs. But they, but we, are probably seeding the path to our own destruction in the future because it is less clear how that expertise is going to develop in the next generation if they’re never given the chance to develop it. Maybe that experience and learning can be imparted via AI but even if it can, I kind of doubt we’re going to see companies hire a generation of middle level knowledge staff who never worked on something at the lower level, and that expertise and judgement in certain fields is going to be a commodity that starts to die out. Maybe I am being too pessimistic and alarmist but Idiocracy didn’t happen overnight either.

Anthropic co-founder Jack Clark had a post before Christmas, “Silent Sirens, Flashing For Us All,” that described the world he sees coming soon, as in this year (emphasis his):

This problem will worsen in 2026. By the summer I expect that many people who work with frontier AI systems will feel as though they live in a parallel world to people who don’t. And I expect this will be more than just a feeling - similar to how the crypto economy moved oddly fast relative to the rest of the digital economy, I think we can expect the emerging “AI economy” to move very fast relative to everything else. And in the same way the crypto economy also evolved a lot - protocols! Tokens! Tradable tokens! Etc - we should expect the same kind of rapid evolution in the AI economy. But a crucial difference is that the AI economy already touches a lot more of our ‘regular’ economic reality than the crypto economy.

So by summer of 2026 it will be as though the digital world is going through some kind of fast evolution, with some parts of it emitting a huge amount of heat and light and moving with counter-intuitive speed relative to everything else. Great fortunes will be won and lost here, and the powerful engines of our silicon creation will be put to work, further accelerating this economy and further changing things.

And yet it will all feel somewhat ghostly, even to practitioners that work at its center. There will be signatures of it in our physical reality - datacenters, supply chain issues for compute and power, the funky AI billboards of San Francisco, offices for startups with bizarre names - but the vast amount of its true activity will be occurring both in the digital world, and in the new spaces being built and configured by AI systems for trading with one another - agents, websites meant only for consumption by other AI systems, great and mostly invisible seas of tokens being used for thinking and exchanging information between the silicon minds. Though we exist in four dimensions, it is almost as though AI exists in five, and we will be only able to see a ‘slice’ of it as it passes through our reality, like the eponymous ‘excession’ from Iain M Banks’ book.

Claude Code itself doesn’t change the world overnight. It’s too stuck in a command line that too few will unlock in its current format, though there will be vaster fortunes to those who abstract this power to something simpler and easier to use (I say vaster because there are already unbelievably vast fortunes already being made in the AI age). But Claude Code is a window into what the AI future may be. It’s not a crystal ball but a glimpse of the future today. And we should all be playing around with it to try and understand that future.

I have always been of two minds on what the AI revolution will look like.

One part of me worked at Slack and sold enterprise software you had to pay for to large customers, primarily the US government, and saw how fucking hard it is to get large organizations to change. So yes, AI diffusion will be slower and more painful than we think.

The other part of me was on the ground floor of the last technological revolution as an early Facebook employee. I remember walking around the halls of Capitol Hill trying to convince people that social media wasn’t a fad but a new tool that would change everything. It’s funny, once you’re convinced you’ve seen the future part of you tries to evangelize to others. But another part of you, another part of me, used to think about the inevitability and power of certain tipping point technologies. At a certain point you didn’t bother trying to convince people to use fire. You either used fire or you died.

This is not me saying that Claude Code is like fire, one of the fundamental discoveries and tools that changed the course of human existence. But Claude Code may be like the first spark man made which eventually caught fire.

Today, I hit my goal of running 1,000 miles this year.1

Running 1,000 miles in one year was a goal that I had been chasing for several years now. This was not an easy goal to meet. I ran 1,000 miles in 192 runs over 174 hours in 2025. That was a lot of time, energy, and effort to accomplish a goal I had been chasing for literally years. After years of failure, I was determined to finally conquer my 1,000 mile challenge.

In 2022, I remember going for my last run in December to hit my goal of 250 miles for the year. The year before I had become a father and moved to a new house near Rock Creek Park. I knew I needed to do something to combat the stress of having an infant and thought that I could pick up and keep running during one of the hardest parts of parenting, I could probably keep doing it when it got easier. So I wanted to keep that running momentum going into the next year.

In 2023, wanting to give myself a challenge, I set a stretch goal to double my mileage from last year and run 500 miles that year. I started training for races, first a half, then a marathon, and many other races. I ended up smashing through that goal, so I kept adjusting my goal upwards, first to 750 miles, then eventually settling on a goal of 1,000 miles. I ended the year with 913 miles, falling short of my revised goal but showing myself that 1,000 miles in a year was very possible.

In 2024, I knew it was an election year and that I hoped to be on a campaign that fall, so I knew that I wouldn’t be training for a marathon (made that mistake in 2008 and ran the Marine Corps Marathon and then deployed a field office for the Obama campaign two days later and couldn’t walk up the stairs to my office). But I decided to keep the 1,000 mile goal to see what would happen. I ended the year at 712 miles, despite going on a presidential campaign (where I somehow managed to keep running), I still failed to hit my 1,000 mile goal but that was solid base mileage for someone on a campaign in an election year.

In 2025, this was the year that I was determined to actually conquer the 1,000 mile goal. I knew that I wanted to run the Marine Corps Marathon again to try and accomplish the sub-4 hour goal that had eluded me in 2023. Having a marathon to train for, along with the various DC races I always ran, meant that 1,000 mile goal didn’t just seem reasonable, it seemed downright achievable.

I ran some races in the spring and then had to slow down my running while I attempted to climb Mount Rainier in May (failed to summit but more on that some other time). Then right after my return from the mountain climb, disaster struck, and I sprained my ankle in early June (I have historically bad ankles but hadn’t had issues for a few years). That meant almost a full month off from running and a gradual return at low mileage over the summer. I fell behind in May (planned with the climb) but then I just plateau through June, gradually starting to run again over the summer as you can see from my annual mileage tracker.

By August I had recovered enough to train for the Marine Corps Marathon in earnest, if slightly behind my training plan and having lost some fitness during my recovery, but I was on track with my training into the fall.

Then disaster seemingly struck again. While walking my dog in Rock Creek Park just three weeks before the Marine Corps Marathon, I came off a rock and resprained my injured ankle. I was despondent, all that hard work training for the race seemed to be for naught.

Long-time readers know what happened next. The orthopedic surgeon told me my ankle wasn’t great, but it could be worse, and while he wouldn’t recommend it, I could try and run the race if I thought I could handle it. I found a physical therapist Elliott Place at Sports Therapy and Rehabilitation who helped nurse me back into maybe being able to run. And on October 26th, I ran the Marine Corps Marathon on a sprained ankle (I would do it again but I would not recommend it).

After the marathon, I dialed it back for a few weeks, but my ankle was recovering nicely and I wasn’t that beat up by the marathon because I had taken it fairly easy due to the aforementioned sprained ankle, so my mileage was pretty quickly back up to twentyish miles a week.

Between the two ankle injuries (or arguably one extended ankle injury), I had written off the 1,000 mile goal as impossible for this year. But towards the end of November, I started to run the numbers and realized that I was on track to get within 30 or so miles of my 1,000 mile goal by the end of the year. That was close enough that if I was intentional for the last month of the year, I might have a shot at hitting my goal.

I had finally gotten around to getting an online running coach, Andrew Lemoncello of McMillan Running, and I worked with him to set up a plan to hit 1,000 miles in the time we had left. It was a tight timeline but possible, but I would need to make sure I ran every planned mile.

It meant that I had to up my mileage in the last weeks of the year. It meant that I had to go running on days and at times when I really would have preferred not to. It meant several runs in the cold, or in the dark, or in the very cold and dark. But my margin of error in the time we had was small and I had to bank as many miles as I could in case we got severe weather that would keep me from running. So I kept running, putting up a marathon training level mileage (for me) of 40+ miles the week before Christmas.

But on Saturday, December 27th, I did it. Today, I ran my 1,000th mile of the year.

Why, after so many years of failure and so many setbacks this year, was I so determined to hit this goal this year? This question was really bigger than running.

I am honestly not a super goal motivated person. I don’t live or die by New Year’s Resolutions. My attempts to adopt positive lifestyle changes have a mostly failed track record. So why was this 1,000 mile goal different?

One thing I realized about myself over the last few years, is that while I am not particularly driven by trying to reach goals, I am extremely driven by challenges. To me, a challenge is something that is not only difficult to accomplish, but that you might not succeed at, that you could fail to achieve. In corporate speak these might be called “Stretch Goals.”

It turns out that not only do I rise to a challenge, but failure is maybe my most powerful motivation. Not fear of failure, after all a challenge shouldn’t be something you know you can accomplish, but for me the motivation truly comes from my previous failures. I accept that I might fail in some of the goals or challenges, but I don’t like failure, and it turns out it motivates me more than anything else.

I failed to summit Mount Rainier this year, so we’re probably going back to climb it next year (I also failed to summit Mount Baker the first time I did mountaineering). I failed at my sub-4 hour Marine Corps Marathon goal in 2023 and 2025, so I am already looking ahead to running it again in 2027 (not next year, it’s an election year). I lost the 2024 election and I’ll be damned if I’m gonna sit on the sideline and let us lose again in 2028 (already thinking about how to go back to Pennsylvania). I had failed to reach 1,000 miles for the past two years. So I was not going to let the chance to reach that goal in 2025 pass me by, especially after the disappointments on the mountain and the marathon.

The thing about challenges is that while you can and should celebrate their accomplishment, it’s not really the thing that brings you actual significant sustained joy or pride. There’s all kinds of research out there that shows that it’s the journey, the love of the process, the work to get to the goal that is the real reward (read Brad Stulberg, Michael Easter, and plenty of others).

For me the reward is not the thousand miles I ran in 2025, maybe it’s just learning that for me it’s all about challenges not goals and that previous failures are my strongest motivator. Maybe it’s not the healthiest approach to the world but there’s value in that discovery and honesty with yourself, especially at my age. Running gives you plenty of time to think, I had 174 or so hours to do so, and you’ll learn something about yourself if you let it teach you. I am not a better person because I ran 1,000 miles, but I am a better person having run 1,000 miles this year.

Or I should say the 1,000 miles that I have run in 2025 so far. After all, tomorrow is my Sunday long run, and my last one of the year, and I can’t miss that.

Besides, I have to start thinking about what my running mileage goal will be for 2026 and how to make it a challenge.