The facts
On Tuesday, AI chief executives agreed to sign a statement of principles to keep artificial intelligence models safe, then surrounded President Trump outside the White House to present the image of a unified industry addressing the public's fears. Behind the scenes, the unity barely survived the photo op: industry leaders including Nvidia's Jensen Huang had just questioned Anthropic's Dario Amodei about his warnings on the dangers of AI, people familiar with the matter said.
The executives asked Amodei why he was being so extreme in public about the capabilities of AI models and the risks they posed, following a lunch in the East Room with about two dozen CEOs. Amodei told them it is important to be honest with the public about model capabilities and not play down risks. The exchange happened in a smaller group in the Roosevelt Room, where executives and White House aides and advisers were hashing out the principles.
Trump and House Speaker Mike Johnson got the top AI companies to agree to the principles at the lunch, but the grilling of Amodei is the latest sign of tensions between Anthropic and the rest of the industry that could hinder the sector's plan for “self regulation” using model evaluations and internal controls. Amodei has previously criticized the White House and competitors for playing down safety concerns, and warned about AI models carrying out cyberattacks and extreme job losses, making him a target for criticism in Washington and Silicon Valley.
Director of National Intelligence Jay Clayton called the meeting “one of the most, if not the most, constructive and collegial industry meetings I have ever seen.” Amodei told reporters “we can win safely” if the industry and administration work together, and Trump praised him for being “great.” The two dined together at the White House on Sunday for their first face-to-face meeting.
Anthropic is gearing up for an IPO it hopes will value the company at some $2 trillion, while still reeling from earlier bruising battles with the administration: a 2 ½-week shutdown of two of its models over security concerns and a Pentagon blacklisting over a dispute about AI use in the military.
The principles followed months of talks on preventing AI cyberattacks after the release of Anthropic's Mythos. This summer, Anthropic, OpenAI and Google had coalesced around a self-regulatory organization idea similar to the Financial Industry Regulatory Authority. Huang, Meta's Mark Zuckerberg and SpaceX's Elon Musk told Trump they worried it would concentrate too much power in the leading AI companies, so the idea was scrapped.
Zuckerberg played a key role finalizing the principles, frequently talking to Trump and others including Commerce Secretary Howard Lutnick, who oversees a key government AI testing unit. He has emerged alongside Huang to champion the light-touch approach and push back against calls from researchers at Anthropic and OpenAI to start slowing down model development if needed. When Amodei raised safety concerns at the lunch, Zuckerberg said the best way to address them was following through on the principles as an industry.
Meta has aligned itself with the White House after previously facing criticism from Trump for suspending his social-media accounts following the Jan. 6, 2021, attack on the U.S. Capitol. In Trump's second term, the company settled a lawsuit, rolled back fact-checking protocols, hired Trump-friendly executives and donated to the president's pet causes. Zuckerberg was smiling and laughing with the president Tuesday: “We want to give the American people and our customers confidence that the technology works in the way that we intend,” he said.
Our take
Nothing says “unified industry” like a group photo followed by a private interrogation. The AI barons posed with Trump for their safety pledge, then hauled the one guy who believes in it into the Roosevelt Room and asked what the hell is wrong with him. Amodei's crime: telling the public the truth about what these models can do. In 2026 Silicon Valley, honesty is extremism.
Then look at what they signed. The sector wants “self regulation” via model evaluations and internal controls, a fancy way of saying they want to grade their own homework. When a real watchdog was on the table, a FINRA-style body, Huang, Zuckerberg and Musk killed it because it would concentrate too much power in the leading AI companies. The leading AI companies argued against giving power to the leading AI companies. Selfless, or the most transparent bullshit ever uttered in the Roosevelt Room.
And the designated safety hall monitor is Zuckerberg, whose company settled Trump's lawsuit, rolled back fact-checking, hired Trump-friendly executives and donated to the president's pet causes. Now he lectures Amodei about following through on principles as an industry, the principles they signed an hour before interrogating the one signatory taking them seriously.
Amodei, meanwhile, needs a $2 trillion IPO while blacklisted by the Pentagon and already had two models shut down by this administration. So he stands next to the president, says “we can win safely”, and takes his grilling like a man. The safety pledge is theater, and the only honest man in the room got treated like the problem. Sleep tight.
