How Anthropic Courted Trump

Hey, Alberto here! 👋 Each week, I publish long-form AI analysis covering culture, philosophy, and business for The Algorithmic Bridge. Paid subscribers also get Monday how-to guides and Friday news commentary. I publish occasional extra articles.

QUICK REMINDER

I’m running a 50% discount offer on annual subs that ends on June 5th June 12th (extended one week). Get yours now and enjoy the discount forever.

If you read the free edition and have been thinking about subscribing, this is the best time. I won’t do another offer anytime soon.

GET YOUR 50% DISCOUNT

Here’s a quick industry analysis post for your Sunday afternoon. See you tomorrow, Monday, with a fundamental practical guide: How to Prepare for the Next 5 Years.

In July 2025, President Trump described AI as “a beautiful baby that’s born.” He said the government should “grow that baby and let that baby thrive” and warned against stopping it “with politics” or “foolish rules and even stupid rules,” echoing what Vice President Vance had said in February of the same year at an AI summit in Paris: “The AI future is not going to be won by hand-wringing about safety.” Nine months later, in May 2026, The New York Times reported that the Trump administration was considering an executive order to create, among other things, a formal government review process for new AI models before they are released to the public.

The AI baby grew up fast.

Per the NYT, Anthropic, Google, and OpenAI were all briefed on the plans, which implicitly means the rest of the industry was likely excluded. The EO entailed, wrote the NYT, a “stark reversal” of Trump’s position on AI; he “swiftly rolled back a Biden administration regulatory process.” Up to this point, Trump’s administration had been unconditionally non-interventionist.

I guess that, despite his overt disdain for journalists, Trump reads the New York Times because he swiftly let the press know he wouldn’t be signing. In late May, several outlets reported that Trump postponed the executive order because he “didn’t like certain aspects” of it, namely the part in which he would be called out as a regulator. So the White House changed the terms. Now it’s not a formal government review process but a voluntary commitment for review. Trump got what he wanted: if the press calls him a regulator, he can point at them and say, “Fake news!”

He got his alibi, but in practical terms, the EO is the same. Instead of “Submit your model here,” it will be “Wait, why didn’t you submit your model to review? Hiding something, huh?” Same outcome, sneakier mechanism. So, whether he likes it or not, I will still call him a regulator. And thus the same questions can be asked about it: Why did this happen? How did the EO come to be?

The framing is national security in general, with a specific emphasis on cybersecurity risks, but the NYT reported on two other factors that led to these deliberations: First, David Sacks, heavily anti-regulation, left his role as AI czar. Second, the concerns about AI’s effects on jobs, privacy, and mental health are no longer a “woke” anti-AI leftist thing but bipartisan, per Pew Research (and plenty of other sources). The true catalyst, however, was Anthropic’s Mythos Preview model.

I’ve written about Mythos before, but to give you the gist: Anthropic announced this new, powerful AI model that, they argued, “can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.” And they added: “The fallout—for economies, public safety, and national security—could be severe.” They held back the model from public release (initially; now that OpenAI has better public models, they plan to release Mythos anyway). Only a few selected partners, including the government, have access for now. The rest of us were left to be awed at Mythos’ evaluation scores, which Anthropic made sure to show on every release since. The White House interpreted this as a preview of what’s coming.

Anthropic agrees. In a May policy paper on the US-China conflict, they call Mythos Preview a “wake-up call” and use it to argue for tighter export controls, anti-distillation legislation, and accelerated government adoption of American AI. The dangerous model became the centerpiece of a policy agenda; Anthropic is a lobbying firm that happens to have a technological branch.

In a June blog post—of a marketing nature but cleverly disguised as a political call to action—they argue that recursive self-improvement (AI that makes itself better autonomously) is coming sooner than expected, and the implications are so large that the industry should perhaps coordinate a temporary pause on AI development, including, of course, those reckless Eastern actors that prefer to work on the open. For years, Anthropic has argued publicly for government oversight of frontier AI.

CEO Dario Amodei has repeatedly warned against open-weight release of powerful models. The most recent example of this strain of cautionary messaging was from co-founder Jack Clark, who wrote an essay in May where he assigned a 60% chance—now an overly conservative number in light of their updated beliefs—that, by the end of 2028, an AI model will be able to “autonomously build its own successor” via recursive self-improvement.

In general, Anthropic’s scaling policy is built around the idea that increasingly powerful models require increasingly strict controls. Controls that only responsible people like themselves—and unlike, say, OpenAI or Meta or open labs—can safely implement. Everything they do follows a coherent plan. The result is that the regulatory framework the Trump administration has signed into law—even after the watering down—looks, in practice, like a generalization of what Anthropic already does. They’ve raced their way into Trump’s ear.

It’s all for safety reasons, sure, but incidentally, Anthropic’s safety posture has always been aligned with its long-term strategic and competitive interests. (OpenAI GPT-5.5 shows similar cyber capabilities, and it was released normally; to this day, without known cybersecurity issues.) I don’t doubt their honesty, but I can easily imagine how happy they must be now that an unreleased Mythos Preview enabled such a nice belief-market fit. As I wrote in April: “[G]ating and publicity are independent choices, and conflating them lets Anthropic have both at once—the ethical high ground of restraint AND the commercial-reputational win.”

To users, they say: I’m sorry, but this is too dangerous for you to have. To investors, they say: We are the best. To the government, they say: You can’t let this happen with unrestricted models. Anthropic has perfected the art of making the bull case and the bear case identical: “this technology is so powerful it might be dangerous” is both a reason to invest and a reason to regulate, and Anthropic benefits either way.

I’m starting to suspect that Anthropic’s whole strategy since the first Pause AI letter failed in March 2023—more than three years ago!—was to become the #1 AI company and from that position of power indirectly threaten the world—by warning us about what the technology they’re building is capable of—with things like cybersecurity, bioweapons, and recursive self-improvement, and thus force the hand of those that can impose their will onto the reckless. That is, the US government.

And so Trump had no choice.

Per the EO, the NSA, in consultation with a handful of officials, will run a classified—that is, opaque—benchmarking process to determine which models qualify as “covered frontier models.” Developers who participate get 30 days of government access before public release and the privilege of being designated “trusted partners.” Developers who don’t participate get a raised eyebrow and some questions in the backrooms. There’s no mention of open-source AI, and the order doesn’t mandate per se, but it effectively creates a two-tier system: either you’re a friend of the government or you aren’t. If the three top labs voluntarily submit to review and earn “trusted” status, every lab that doesn’t is a second-class citizen in the eyes of enterprise buyers, procurement offices, and investors. You don’t need to ban open AI to halt it.

To top it off, the Washington Post reports that Trump “is considering taking a government stake in leading artificial intelligence companies.” Once the US government links its future to the AI industry, it will set a protective incentive structure. Such convenient timing, right when Anthropic is preparing the IPO. Creating a regulatory-financial moat just before it goes public makes it worth even more to investors precisely because almost no one else will clear the bar.

But wait. Don’t they hate one another?

Didn’t the military cut off government use of Anthropic’s technology after the two sides failed to agree on terms? Isn’t Anthropic suing the Trump administration over a $200 million contract dispute? Didn’t Hegseth officially designate Anthropic a “supply chain risk”? Yes. That’s precisely why they’re doing this. Anthropic is simultaneously courting the government and taking it to court. The government is simultaneously designating Anthropic a supply chain risk and welcoming it into its classified network. Trump claims no regulation and days later he signs the executive order.

This is the political version of that popular term used in pro wrestling: kayfabe or “be fake.” We, the unwitting viewers, cheering back and forth, are the actual target. Because what else would remain of an AI industry but an effective oligopoly?

In the policy paper that I mentioned above, Anthropic describes an ideal 2028. They call it “Scenario One,” in which “a small number of AI labs lead at the frontier with the most intelligent, capable, and performant models. All are based in the US.” We know who the small number refers to: Anthropic, Google, and OpenAI. Their pricing power will be absolute. The subsidies on AI tools will be over. When the three top labs are favored over the rest and the government, which is itself their client, gatekeeps the releases, the competitive pressure that keeps AI affordable disappears. AI has entered its non-democratic era.

The industry no longer needs individual users to survive. The “IPO lens” has replaced the “user lens.” You and I are not the engine of their business anymore and will never be again, but the government, enterprises, and investors are. Everything for the people without the people, I guess. In professional wrestling, they pretend to hurt each other. In AI, they court each other at court.