WASHINGTON-After agreeing to sign a statement of principles to keep artificial-intelligence models safe Tuesday, AI chief executives surrounded President Trump outside the White House to present the image of a unified industry addressing the public's fears.

Behind the scenes, industry leaders including Jensen Huang of Nvidia had just questioned Anthropic's Dario Amodei about his warnings about 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 risks they posed following a lunch in the East Room with about two dozen CEOs. Amodei told the other executives that it is important to be honest with the public about model capabilities and not play down risks, some of the people said.

Their conversation occurred with a smaller group of executives and White House aides and advisers who were hashing out the principles in the Roosevelt Room, some of the people said.

Trump and House Speaker Mike Johnson (R., La.) got the top AI companies to agree to the principles at the lunch, but the questioning of Amodei is the latest indication of tensions under the surface 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 driven by the technology, making him a target for criticism by many in Washington and Silicon Valley.

Jay Clayton, director of national intelligence, said in a statement that the meeting was "one of the most, if not the most, constructive and collegial industry meetings I have ever seen."

"The focal point was, 'What is best for the American people and globe?' and that focal point drove agreement," said the candidate to be Trump's next AI "czar" coordinating policy across the administration.

On Tuesday, Amodei told reporters "we can win safely" if the industry and administration work together, and Trump praised him for being "great." The duo dined together at the White House Sunday for their first face-to-face meeting.

Anthropic has sought to mend fences with the administration recently and is gearing up for an initial public offering it hopes will value the company at some $2 trillion. The company is still reeling from a pair of bruising battles with the administration earlier this year that included a 2 1/2-week shutdown of two of its models due to security concerns and being blacklisted by the Pentagon over a dispute regarding how AI is used in the military.

Tuesday's principles were the result of months of discussions over how to prevent AI cyberattacks following the release of Anthropic's Mythos and similar models earlier this year. The administration is testing powerful models before they are released publicly, but recent hacks by AI models have demonstrated the need for stronger reviews when models are being trained and developed, industry executives say.

This summer, Anthropic, OpenAI and Google had coalesced around an approach being considered by the administration for a self-regulatory organization similar to the Financial Industry Regulatory Authority, which oversees brokerage firms.

Huang, Meta Platforms CEO Mark Zuckerberg and SpaceX's Elon Musk told Trump they worried the idea would concentrate too much power in the leading AI companies, so the idea was scrapped, The Wall Street Journal previously reported.

That structure also would have involved many of the ideas included in Tuesday's principles such as stronger internal reviews of models during development and external testing.

That highlighted the pressure facing the White House and industry due to public concerns about AI and the data centers powering it.

Zuckerberg in particular played a key role in finalizing Tuesday's principles, frequently talking to Trump and others in the administration including Commerce Secretary Howard Lutnick, who oversees a key government AI testing unit, some of the people said. He has emerged alongside Huang to champion the White House's light-touch approach and push back against calls from researchers at Anthropic and OpenAI to start slowing down model development if needed in coordination with the government.

The push for the dinner began when he and Johnson sat next to each other at the State Dinner with China's Xi Jinping last week and talked about AI policy, some of the people said. Zuckerberg and Huang then tried to coordinate top CEOs while Trump and Johnson organized the lunch to build consensus, these people said. Semafor previously reported on the conversations.

When Amodei brought up his safety concerns during the lunch, Zuckerberg responded that the best way to address them was by following through on the principles as an industry, some of the people said.

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 over the move, 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 while taking questions from the press. "We want to give the American people and our customers confidence that the technology works in the way that we intend," he said.