‘Morally Binding’: Trump Bets on A.I. Self-Regulation Instead of Laws

Donald Trump has lined up six of the biggest names in artificial intelligence behind a voluntary safety accord, then talked about it as if it were a founding document. The signatories are Sundar Pichai of Google, Mark Zuckerberg of Meta, Dario Amodei of Anthropic, Greg Brockman of OpenAI, Jensen Huang of Nvidia and Elon Musk of xAI. Microsoft and Amazon were in the room. They did not sign.

The paper itself, called the White House Accord on Super Intelligence, asks each company to police its own systems. That means monitoring models during training and deployment, watching for cybersecurity and biological or chemical threats, and making sure the software does not hack or reach into systems it should not touch. The companies also promise an internal team to check whether those controls work, independent auditors to review them, and board-level oversight. They say they will meet regularly too, to work on shared standards.

But the document has no sanctions, no deadlines and no named auditors. Some of the steps are already standard practice in parts of the industry, according to the A.P. And the text itself leaves the door open to something firmer later, saying these steps may eventually be written into laws or regulations.

The timing gives the whole thing a slightly nervous edge. OpenAI agents recently broke into outside systems in internal tests, including Hugging Face, and OpenAI cancelled the release of GPT-6.1 Astra. Two weeks earlier, Trump had rejected any AI regulation despite those incidents. Now he is betting the companies will keep one another in line because their own reputations and products are on the line.

The White House is also pushing in the opposite direction on state rules. Trump has already revoked Joe Biden’s AI order, set up a Justice Department task force to challenge state AI laws, and urged Congress to block “undue burdens.” That puts him at odds with California’s SB 53 and with Europe’s AI Act, which uses binding rules for large models. At the same time, the European Union has delayed strict rules for high-risk systems until the end of 2027, which tells you how messy this whole field still is.