For generations, American businesses have practiced something political scientists call regulatory capture. The concept is simple enough. A company cannot repeal the law, so it does the next best thing and takes control of the people the government put in charge of writing and enforcing it. The railroads did it to the Interstate Commerce Commission. The airlines did it to the Civil Aeronautics Board, which spent 40 years deciding who was allowed to fly which routes at which fares and never once approved a new trunk carrier in all that time. Perhaps most famously, AT&T invited regulation and then helped shape it to stifle competition in the telecom industry for close to a century. More recently, the health insurance industry helped shape the Affordable Care Act. It came out of that negotiation with a federal mandate requiring Americans to buy its product and taxpayer funded subsidies to help them afford it. None of this is new. What is new is that we are watching it happen in real time, out in the open, in artificial intelligence.
Dario Amodei, who runs Anthropic, published an essay last week called “We Must Pace the Frontier.” Its central proposal is that the major AI companies agree to slow down together, and that they prove they are slowing down by handing outside evaluators “desks in our offices, access badges, and company laptops.” He names one nonprofit, METR, as the sort of organization that should get those badges. Anthropic is adopting the arrangement on its own, without waiting for Congress to require it. Sam Altman agreed within a day. “I agree with Dario that we need to pace the frontier,” the OpenAI chief executive said, and he committed his own company to the same embedded-evaluator access. Elon Musk, whose data centers Anthropic rents, said Dario is right.
Presented that way, these are companies volunteering to be watched. The question is who ends up wearing the badge.
David Sacks, who served as President Trump’s artificial intelligence and crypto czar until March and now co-chairs the President’s Council of Advisors on Science and Technology, 𝕏answered that question on Friday night. He told Amodei and Sam Altman to go ahead and pace themselves, since the two of them are the frontier, and then listed what he wanted them to quit doing. “Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability. Stop pretending METR is independent when it is intertwined with Anthropic’s investors and staff. Stop pretending you need those same evaluators to police competitors who aren’t even at the frontier.” If they insist on their own regulatory framework as the price of slowing down, Sacks wrote, “we’ll know this was just another bid for regulatory capture.”
That is the man who ran this President’s artificial intelligence policy accusing the two largest AI companies in the country of trying to write the rules for everyone else. The strongest evidence for his case is not in his post. It is in Amodei’s essay.
Amodei writes that “the most effective method of pacing is via regulation that targets all US frontier AI companies,” and that Congress should pass bills “that focus on transparency and on third-party auditing.” He also writes that “for antitrust reasons, it’s helpful for the US government to mediate or at least enable these discussions.” The chief executive of one of two dominant firms is asking the federal government to sit in on a conversation among competitors about how fast each of them will move, because he understands perfectly well that the conversation is otherwise a Sherman Act problem. He is not hiding this. He put it in writing.
Notice what the essay does not propose. There is nothing in it about product liability. Transparency mandates and third-party audits are fixed costs, and fixed costs are the oldest barrier to entry there is. A company with thousands of employees absorbs them as a rounding error. A startup in a garage cannot absorb them at all. Liability for what a product actually does scales with revenue, deployment, and market share, and no compliance department makes it cheaper. Of the two ways to make AI safer, the one Amodei recommends raises the drawbridge, and the one he leaves out would cost Anthropic money. 𝕏Nathan Leamer, a tech policy hand in Washington, said naming METR as the evaluator is “a tell.” It is. 𝕏Neil Chilson, formerly the chief technologist at the Federal Trade Commission, put the same point more plainly: go ahead and pace, because literally no one is stopping you.
That brings us to the week’s other story.
On September 8, a pretraining researcher named Jacob Coxon 𝕏announced he had resigned from Anthropic because the industry is “racing straight to self-improving superintelligence and gambling with our lives.” The post has 165 million views. The Wall Street Journal reported it produced a groundswell of urgency for AI regulation in Congress. At least 22 elected officials responded with calls for legislation, and within days Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act, which creates a cabinet-level federal agency to police AI, advised by a board of artificial intelligence experts, carrying a corporate death penalty and 20 years in prison for violators.
Parker Thayer 𝕏walked the timeline. The Journal published its exclusive 18 minutes before Coxon posted anything, which means the rollout was planned in advance. Brian Chau went looking at the resume, and 𝕏Thayer confirmed what he found. The insider who set off a stampede in Congress had worked at Anthropic for no more than four months. His own post says his three years of pretraining research were split between two companies, which means nearly all of it happened at OpenAI, not at the company he was resigning from in protest. Before either job, in 2021, he was a fellow at 𝕏Newspeak, a British outfit the Guardian once described as trying to hack British politics, which Chau reports was bankrolled by the Long-Term Future Fund and the Effective Altruism Infrastructure Fund. Newspeak is directly connected to prison abolition activism and the radical environmental group Extinction Rebellion.
The first three accounts to amplify Coxon, all within 15 minutes, were the general counsel of Encode AI, the head of policy at the AI Policy Network, and the head of the AI Futures Project. All three organizations are funded through the Survival and Flourishing Fund, which directs the giving of Jaan Tallinn, an Anthropic investor. Coxon himself took a $20,159 scholarship in 2022 from Good Ventures, the foundation behind Open Philanthropy, since renamed Coefficient Giving. Dustin Moskovitz, who runs Coefficient Giving, led Anthropic’s Series A funding. Holden Karnofsky co-founded Open Philanthropy and joined Anthropic himself in January of last year. He is married to Anthropic’s president, Daniela Amodei, who is Dario’s younger sister and co-founded the company with him.
The amplification is the part that ought to bother people. Coxon’s account had almost no prior activity, and the post drew 165 million views. Elon Musk, who owns the platform and can see the underlying data, 𝕏said he does not think this has ever happened for a post from a new account with almost no prior history. Thayer sampled 3,500 accounts that reposted it and found 76 percent of them were foreign, concentrated in India, Indonesia, and Mexico. Notably, Steve Jurvetson, a venture capitalist, ran the same analysis on Thayer’s own skeptical post and found 66 percent of its engagement came from Americans. Thayer asked X publicly to run an engagement-botting check, and X has not published one. But, the most consequential AI policy moment of the year, the one the Journal credits with producing a groundswell in Congress, may have been carried to Congress by accounts that cannot vote here.
The people who own Anthropic also fund the movement demanding that Anthropic be regulated. Moskovitz has more than $1 billion out the door in AI safety advocacy and another $1 billion committed this year. As 𝕏Kevin Bass laid it out, that is not a coincidence of adjacent interests. It is one network financing both sides of a debate it intends to win.
Run the sequence forward and the shape of it is plain. Congress panics. It passes the Sanders bill, or a milder version with the same architecture: a federal body, an approval process keyed to model size, and a board of experts to advise it. Anthropic clears the threshold, and its competitors, along with every startup that has not raised $10 billion, do not. The experts seated on that board come from the same nonprofits the same donors fund. The third-party evaluator with the desk and the badge is METR, which Amodei named himself, and which Sacks says is intertwined with Anthropic’s investors and staff. The company ends up holding the pen that writes the rules its rivals have to follow, and it gets there without a single illegal act by anyone.
The competitor who gets locked out has a name. The serious threat to both companies is not another startup burning venture money on a closed model of its own. It is Meta, which has spent the past two years arguing that the answer to everything Anthropic and OpenAI say they fear is to open the models up rather than lock them down. Mark Zuckerberg’s case, published this summer, is that widely deployed open systems “have proven more secure because more people can identify vulnerabilities, harden the systems, and easily upgrade,” and he points to the Hugging Face incident, where open models were used to patch the very holes that attack exposed. Restricting who may hold these capabilities, he argues, “leads us down the path of centralization and lack of checks and balances.” The long-term answer “isn’t to withhold capabilities but to establish a balance of power where superintelligence is broadly distributed.”
Reasonable people can argue about whether Zuckerberg is right. What is harder to argue is that he is losing. Epoch AI, which tracks model capability, found that since January the best open-weight models have trailed the best closed models by an average of four months. Four months is not a moat. It is a head start, and head starts close on their own. A federal approval process keyed to model size does not close on its own, and it falls hardest on the freely downloadable model that Anthropic and OpenAI can neither sue, acquire, nor outspend.
None of this is happening on a blank calendar. Anthropic confidentially filed its IPO paperwork with the Securities and Exchange Commission on June 1 and is targeting a public offering next month, with Morgan Stanley, Goldman Sachs, and JPMorgan Chase running the deal. A funding round in May valued the company at $965 billion. CNBC reported in August that the filing will list backlash against AI as a risk factor, which is a company warning its future shareholders that public anger is the thing most likely to hurt the stock. OpenAI landed on the other side of that decision. Sam Altman said this week that his company will not go public this year, calling it an ill-advised moment and pinning it on safety. “I would say not 2026,” he said. “We got a lot of stuff to do, like meeting this moment of what is going to be required for safety and alignment.”
Neither company has ever turned an annual profit. Anthropic lost somewhere between $10 billion and $15 billion from 2021 through 2025 and is expected to lose roughly $2 billion more this year. OpenAI is on track to lose about $14 billion in 2026, and some estimates run considerably higher. Both are preparing to ask public markets to fund businesses that burn extraordinary sums today against the promise of a durable advantage tomorrow. A federal approval process keyed to model size is that advantage, and it is the cheapest version available, because the taxpayers build and staff it. Compare Anthropic and OpenAi to Meta. That company made over $60 billion in profit last year.
The endorsements are arriving on schedule. At a private fundraiser on Thursday, according to the New York Times, former President Barack Obama told House Minority Leader Hakeem Jeffries that artificial intelligence “is moving very fast in private hands, and if we don’t get on top of it, I think it can be dangerous.” He urged Democrats, should they win the House in November, to put together a framework and a public conversation about governing it. Obama is the man who watched the health insurance industry help shape his signature domestic achievement and walk away with a federal mandate requiring Americans to buy its product. He is now lending his name to the proposition that this industry needs rules, at the exact moment this industry is asking for them.
Amodei is asking the country to defer to his judgment about what the rest of us may be permitted to build. His wife, Cami Clark, holds no formal title at Anthropic, which has not stopped reporters from calling her the first lady of AI and the most consequential person in the company’s orbit without a job there. In March of 2012, four years after Jeffrey Epstein pleaded guilty in Florida and registered as a sex offender, Clark emailed him to ask whether he would invest in her businesses. One of them was a self-described free luxury porn company. The Epstein files, reported out by the Wall Street Journal and Forbes, give no indication that he put money in.
Amodei wants evaluators with desks, badges, and company laptops inside the firms they sit in judgment of, and he wants Congress to make that the law for everybody. Before any of it gets written into the United States Code, somebody in a hearing room ought to ask who hired the man wearing the badge. Anthropic, to its credit, published the essay admitting it is the one who hired him.
Lastly, if the rhythm and cadence of this piece sounds like AI played a role, it did. I asked Anthropic’s own Claude to piece together all the threads and Claude, itself, concluded it sounds like a well-orchestrated political campaign to engage in regulatory capture and then wrote everything above that you just read with only a few edits from me.
Bo French is…
