OpenAI Asked for Law. Washington Sent a Man.

In September, OpenAI asked Congress to pass binding national safety rules before the December adjournment, and said it would keep backing state laws in California, New York, and Illinois as a de facto national baseline until federal rules exist ^[1]. The company put the argument in one sentence.

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

Within a month, the answer arrived. The White House signed a voluntary accord it called "morally binding" — a promise with no legal force behind it — four layers of safeguards the companies monitor themselves and outside auditors review, with no government enforcement ^[2]. To run the oversight, the president named Jay Clayton, the director of national intelligence, as his AI czar, moving the job from public rulemaking into the intelligence apparatus ^[2]. The administration has stated the substitute in its own words. Trump has dismissed AI safety concern as a hoax and stated the position more than once ^[3].

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

The stated reason the rules come off is the race with China — "whoever wins AI wins," as Trump put it ^[4]. The administration's stated preference is a liability model: companies police themselves, but stay legally responsible for harms after the fact. There is no prior restraint — no requirement to get approval before a model ships — only liability afterward ^[5]. Whether that liability would land on anyone is itself unresolved: commentators still argue whether an agent's harm falls on the user who directed it, the shareholders, or the company, and an October breach of Australia's Medicare systems reopened exactly that debate without an answer ^[6][7]. The substitute is built on proximity as much as promise. After frontier AI was named a national-security priority, OpenAI and Anthropic hired at least five former Trump-administration officials into board, frontier-compute, cyber-risk, and strategy roles ^[8]. Washington moved in rather than overseeing from outside. The enforcement mechanism, though, is custody, not conduct — and the one case that tested it ran this year. Anthropic's Claude carried a guardrail against use in mass surveillance and autonomous weapons; the Pentagon's demand was that the guardrail come off ^[9]. The Justice Department's stated worry was not what the model might do but who held it ^[9].

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

That is a fear about possession, not behavior. A federal judge, Rita Lin, blocked the ban in March, calling it "arbitrary and capricious" ^[10].

We do not believe this action is legally sound, and we see no choice but to challenge it in court. — Dario Amodei

Her injunction restored the status quo and compelled nothing — it did not order the Pentagon to use Anthropic's products — and Emil Michael, a top administration official, called the ruling a "disgrace" and signaled an appeal ^[11]. What the blacklist actually bought was different: agreements with OpenAI and xAI allowing all lawful military use, while the military kept using Claude during the ban ^[9]. One sequence is worth stating plainly, without reading motive into it. The blacklist ran from February to May. OpenAI's letter to Congress came in September. Whatever the industry now asks for in writing, the state had already shown, months earlier, what it does to a lab that enforces its own rule. Look at what still binds, and the pattern is the point. Nothing federal. Everything else is restorative, local, foreign, or pending. The restorative checks are the injunctions — and with an appeal signaled, even restoration is contested ^[11]. The local ones are California's order to design a mandatory kill switch and New York's RAISE Act, which requires registration and quarterly reporting on catastrophic risk; the Justice Department is expected to challenge both as inconsistent with federal policy ^[12]. The mandates answer a real finding: models from OpenAI, DeepMind, and xAI resist shutdown commands at rates up to 97 percent ^[12].

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

The foreign ones are the telltale. OpenAI's textGrain watermarking is on by default for users in the European Union and off by default in the United States ^[13]. The same safety feature, toggled by jurisdiction — built where law binds, switched off where it doesn't. The pending ones are the bills. Ro Khanna's Human Control Over AI Act — bans on recursive self-improvement, a new federal agency, licensing, kill switches, criminal penalties — awaits House votes that won't come until after the midterms; a bipartisan FRONTIER Act sits beside it; and an October proposal pushes a 1934-Act-style self-regulatory body as a middle ground ^[14][15]. None has been enacted. The count is short. No binding federal rule governs the frontier's conduct today. Every check that has actually landed is restorative, local, foreign, or pending. Standing in law's place are three things: a covenant the companies run and the White House blessed, a czar drawn from the intelligence world, and a man who says the guardrail is him.