AI chiefs challenge safety warnings hours after signing Trump's pact

By Zara Kincaid · 2026-10-01

The Roosevelt Room split

Six signatures on a single page, and the accord was barely posted. On 29 September, President Donald Trump put the "White House Accord on Super Intelligence" on Truth Social, bearing the names of Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman, Elon Musk and Jensen Huang.^[1] Within hours of that signing, Nvidia's Huang and several co-signatories confronted Anthropic's Amodei inside the Roosevelt Room over his public warnings about AI risk, according to reporting on the exchange.

The accord runs to four layers of control: strong controls over training and deployment, an internal monitoring team with real authority, a partnered independent external auditor, and a designated independent board committee.^[1] The document leaves the door open to eventual legislation, noting it "might make sense to codify these steps into laws or regulations" at a future point.^[1] The collegial framing collapsed into a confrontation faster than most index cycles turn over.

The fault line: who decides when to sound the alarm

Huang and Zuckerberg have both argued the industry can govern itself without triggering public alarm or inviting new regulation. Amodei's position runs in the other direction: labs must be transparent about what their models can and cannot do, and that transparency carries a cost in public comfort.

Amodei published an essay calling on the AI industry to slow down, with a three-part plan for doing so. He said Anthropic would unilaterally commit to the first step: granting third-party evaluators permanent, employee-level access to its systems to verify safety measures, report on incidents and assess model alignment during training.^[3] That commitment, made six days before the White House lunch, is the concrete expression of the divide. Self-policing, in Amodei's framing, requires independent verification; in Huang's framing, the public warnings that justify such verification are the problem.

The accord's own text does not resolve this. It requires internal controls and external auditors but sets no disclosure standard for what executives may say publicly about the risks their models carry.^[1] That gap is precisely where the confrontation lives.

FTC probe: scope, targets and what comes next

The Federal Trade Commission opened an industry-wide investigation on 30 September 2026, targeting Anthropic, OpenAI and the AI research group METR over potential dangers their technology poses to consumers.^[2] The accord was signed, the confrontation happened and the regulator moved, all within 24 hours.

A senior FTC official said the commission plans to issue formal demands for information and to compel testimony from executives at the targeted companies.^[2] Compelled testimony is a different instrument from a voluntary accord: it requires answers, not signatures.

Voluntary self-policing rests on companies deciding what to disclose and when. Formal FTC demands require companies to produce documents and send executives to answer questions under legal obligation. The accord's signatories now face both tracks simultaneously: the self-regulatory framework they agreed to on 29 September, and the investigative machinery the FTC activated on 30 September.

Zara Kincaid

Zara Kincaid writes about artificial intelligence and search. Her focus is what happens to businesses when the front page of the internet stops being a list of links and starts being an answer.