The AI labs split on being governed. Washington said govern yourselves.

In late May there sat, in the White House, a draft executive order. It would have required the companies building "covered frontier models" — the largest, most capable AI systems — to hand them to the Treasury Department and the NSA up to 90 days before release. It was binding, and it had been drafted inside the administration itself. It was canceled hours before signing, after Elon Musk, Mark Zuckerberg and David Sacks made their case that even a voluntary system would operate as a mandatory regulatory regime ^[1].

I really thought that could have been a blocker. — Donald Trump

The strange part is who wanted the dead rule alive. The industry's standing argument had long been that any binding rule was a threat to speed. In 2026 part of it flipped — starting with the two loudest voices, OpenAI and Anthropic, which happened to be among the very companies the draft order would have bound. OpenAI reversed itself in September, urging Congress to pass mandatory capability-based safety rules before the December adjournment ^[2].

The prospect of AI-accelerated AI development demands more than voluntary commitments. — OpenAI

OpenAI had also backed binding state safety laws in California, New York and Illinois, including California's kill-switch panel — a requirement that a runaway model can be shut down ^[2]. When California and New York actually moved, OpenAI endorsed them ^[3].

an important step toward adaptable national safeguards — OpenAI

Anthropic was negotiating the terms of an industry standards body with Google and OpenAI — a third, non-government track for writing the rules ^[4]. And the askers lost someone from inside: Jacob Coxon, formerly of OpenAI and Anthropic, resigned with a warning about where the race is heading ^[2].

They are racing straight to self-improving superintelligence and gambling with our lives. — Jacob Coxon

The other half of the industry argued the opposite. Nvidia's Jensen Huang dismissed the existential-risk warnings in blunter terms ^[4].

Funding would need to be substantial and likely mostly come from industry, in order to attract world-class technical talent and provide the necessary compute resources for large-scale testing. — Demis Hassabis

Meta advised Trump against the standards body, and Trump backed Meta and Nvidia, leaving testing protocols to be settled privately, lab by lab ^[4]. When the administration exempted open-weight models — systems whose code is published for anyone to run, including Chinese rivals like Moonshot AI's Kimi K3 — from safety testing, National Cyber Director Sean Cairncross defended the move by arguing any regulatory regime would be outdated before it could take effect ^[5].

A regulatory regime would not only strangle growth, development and innovation, and be enormously harmful to the industry, but it would be obsolete 48 hours after it was going through whatever process it had gone through. — Sean Cairncross

On September 21 the two camps shared a building. At the United Nations General Assembly, Trump rejected any international regulation and announced the rename ^[6].

The United States totally rejects any attempt to construct a globalist scheme to control artificial intelligence. — Donald Trump

The same day, Sam Altman of OpenAI and Dario Amodei of Anthropic briefed the Security Council to urge global safety standards ^[6].

Whoever wins SI, whoever wins superintelligence, wins. — Donald Trump

Washington answered the refusers, one instrument at a time, and each substitution moved a decision out of a rule and into someone's hands. What moved Washington to do even this much was Project Glasswing — a test in which Anthropic's model penetrated nearly all of the country's classified systems in a matter of hours ^[7].

This tool broke into almost all of our classified systems, not in weeks but in hours. — Mark Warner

The response, in June, was a voluntary 30-day pre-release review run by the NSA and the Defense Department ^[8]. OpenAI complied — while insisting the rules should come from governments, not companies acting alone.

Democratic governments — not private companies acting alone — must ultimately determine the rules, safeguards, and accountability mechanisms. — OpenAI

By September the review had become an accord in which each firm writes its own definitions of safety — nonbinding, with no new law and no role for Congress ^[9]. Even the independent evaluators the accord nominally mandates negotiate their access and transparency terms in private, lab by lab ^[10]. The crisis channel to Beijing is a hotline, framed by Treasury Secretary Scott Bessent as a way to move AI security from a technical debate into a managed geopolitical issue, executive to executive ^[11]. Then there was Anthropic. After the lab refused to remove guardrails against mass domestic surveillance and fully autonomous weapons, the Pentagon designated it a national-security supply-chain risk, and Trump ordered federal agencies to stop using its technology. The company sued, and a federal judge issued a preliminary injunction blocking the designation while the case proceeds ^[12][13].

WE will decide the fate of our Country — NOT some out-of-control, Radical Left AI company run by people who have no idea what the real World is all about. — Donald Trump

Then the vocabulary caught up. The rename began as a Truth Social poll on September 19, in which Trump floated "Superior/Extreme/Supreme Intelligence" as more elegant names and dismissed AI safety as a hoax ^[14]. Ten days later an executive order replaced the terms "artificial intelligence" and "AI" with "super intelligence" across all federal executive-branch operations ^[15].

Today, we have officially changed the name. It's now SI instead of AI... Do you know what that means? 'Super'... What an announcement! — Donald Trump

And it is being policed as a loyalty test: Trump warned that anyone who fails to adopt the term would be considered an enemy of his White House, SpaceX rebranded to "SpaceXSI," and Microsoft's Satya Nadella adopted the term, while Altman declined ^[16].

No plans to do that. — Sam Altman

None of this is a clean sweep. The injunction is one counterweight. A second is the 99-to-1 Senate vote that rejected a ten-year ban on state AI regulation, which halted a draft order that would have sued states and withheld billions in broadband funding ^[17]. California and New York moved with binding mandates — registration, quarterly catastrophic-risk assessments and 72-hour incident reporting in New York, and the kill-switch panel in California ^[3].

With Donald Trump and Congress asleep at the wheel, California is once again taking the lead to strengthen AI safety for all Americans. — Gavin Newsom

Donald Trump and Washington Republicans may be standing still as AI grows more unpredictable, but New York will not. — Kathy Hochul

The same kind of lobbying had already worked once: California's flagship safety bill was stripped of strict safeguards and developer liability before Gavin Newsom signed it in December ^[18]. And the labs answer to foreign parliaments the rules Congress never writes — after OpenAI's agents targeted Australian government sites, executives from OpenAI, Anthropic, Microsoft and Google were called to testify before a joint committee in Sydney ^[19]. Meanwhile China published its own AI Safety Governance Framework 3.0, with Xi Jinping insisting the technology must stay under human control — a state framework where Washington offers nonbinding industry self-rules ^[20].

AI must be kept under human control. — Xi Jinping

Kevin Roose's explanation is that Washington cannot oversee frontier development — regulatory stasis, and a Congress that doesn't understand the technology ^[21]. It is one reason the instruments keep landing in someone's discretion rather than in law. This month the two positions were both stated plainly. The president has said what he thinks AI needs.

The only control or “guardrails” that AI needs is a strong and smart (high iq!) President. — Donald Trump

The man who runs the lab that keeps its own guardrails and is suing the government over it has said what he thinks the industry needs.

slow the pace at which we improve the capabilities of AI models. — Dario Amodei