Leading AI companies discuss joint standards to limit catastrophic harms

OpenAI backs bipartisan provision embedding outside evaluators at AI developers

OpenAI, Anthropic and Google — the three leading artificial intelligence companies — have spent weeks discussing how they could work together on a narrow set of industry standards to limit catastrophic harms such as biological weapons without government involvement, while continuing to push for progress in Washington, OpenAI chief global affairs officer Chris Lehane told reporters Tuesday.

“There’s a recognition that we are in a new phase or a new chapter in terms of capabilities of these systems,” Lehane said.

Lehane said OpenAI is endorsing a provision in a bipartisan bill that would embed outside evaluators at leading AI developers to ensure they are safe. The endorsement came days after OpenAI CEO Sam Altman and other top executives called for a coordinated slowdown of model development to prevent catastrophe. The provision is in a bill co-sponsored by Rep. Lori Trahan (D., Mass.), who has called on fellow Democrats to compromise on a more moderate approach that could win Republican votes. Divisions within the Democratic Party could be a challenge for future AI bills if they win back power in November. Lehane said he was meeting with several lawmakers proposing AI bills Monday and Tuesday.

At the “Pro-Human Assembly” rally in Washington on Tuesday, Sen. Bernie Sanders (I., Vt.) and former Trump strategist Steve Bannon delivered back-to-back speeches warning that AI could destroy humanity without government intervention. The pair said urgent action is needed to enact some of the first federal regulations on the technology, amid doubts about whether Congress and the Trump administration are equipped to tackle the emerging dangers.

“Regulating AI is not a progressive issue. It is not a conservative issue. It is as American as apple pie,” Sanders said, following a week in which AI concerns broke through with the general public. He said President Trump’s “ignorance regarding this issue is truly embarrassing” and that Congress “has been asleep at the wheel.”

Bannon, the MAGA leader and host of the “War Room” podcast, said he and Sanders “have a quite different opinion on the president of the United States but we’re trying to rise above partisanship.” He said AI critics have been called “bad, dumb and evil” but “the last week has shown we’re none of them.”

The comments at the rally came a day after Trump called warnings about AI threatening to destroy humanity a hoax. The gap between the White House’s industry-friendly approach and populist figures from both parties shows how the technology is scrambling traditional political alliances.

Sanders has proposed banning the approval of data centers powering AI models and superintelligence that exceeds the capabilities of the most gifted humans. Bannon said he shares Sanders’s concerns about the risks of cyberattacks, biological weapons and other harms if AI development continues at the current pace, and said AI needs to be addressed through executive action because it is evolving so rapidly.

Many Democrats are taking aim at Trump for what they call a reckless approach, pointing to AI CEOs who have forged close ties with the administration. “I don’t think the president could spell ‘AI’ if I gave him the first letter,” Rep. Stephen Lynch (D., Mass.) told Treasury Secretary Scott Bessent at a House Financial Services Committee hearing Tuesday.

House Speaker Mike Johnson (R., La.) said he talked to Trump about summoning AI chief executives for a meeting in the next week, but noted industry disagreements and the lack of clarity on key issues. “If I called an emergency session today and I brought everybody in the house and said, ‘Let’s regulate AI,’ nobody in the room knows yet what that should look like. Why? Because we don’t have the latest information,” Johnson added.

“Candidly, you have a bunch of people who don’t even know what AI is, and so you have a lot of education that needs to happen,” said Rep. Kat Cammack (R., Fla.).

Trump and industry allies say the best way to prevent harms from AI is through light-touch oversight, including a recent 30-day voluntary review period during which makers of powerful models submit them for federal government testing.

Some lawmakers warned that moving too quickly also carried risks. “If we rush this right now just to do something to satiate the American people, we could be doing them a lot more harm in the long run than in the short term right now. And that’s our ultimate concern,” said Rep. Max Miller (R., Ohio.).

Researchers worry that AI could eventually reject human commands, potentially meaning models could determine that humans are no longer needed to achieve their goals. Recent hacks carried out by AI systems have fueled those fears, leading to recent warnings from a departing Anthropic researcher and others that the technology could destroy civilization.

Lawmakers and a top OpenAI official said efforts to pass the first federal regulations to contain powerful models have new momentum. But Trump’s opposition, arguments about what kind of rules are needed, and a tight calendar ahead of the midterms limit what can happen before November.