The White House approved an AI safety accord that leaves the largest technology companies to supervise their own advanced systems. It signals how Washington intends to govern the sector without new legislation for now. President Donald Trump and executives from Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed the document after a September 29 meeting in Washington.
Trump called the agreement morally binding and compared it to a constitution. House Speaker Mike Johnson described it as a voluntary statement of principles. The published text, titled the White House Accord on Super Intelligence, carries no statutory penalties, no mandatory licensing, and no federal enforcement body. Signatories include Google CEO Sundar Pichai, Anthropic CEO Dario Amodei, Meta CEO Mark Zuckerberg, OpenAI President Greg Brockman, xAI CEO Elon Musk, and Nvidia CEO Jensen Huang.
Why the White House approved an AI safety accord
When the White House approved an AI safety accord, it chose company-led safeguards over a federal licensing system. Trump said he was seeing substantial self-policing across the industry. He suggested a committee of around 10 people could watch over the wider effort, without naming who would serve on it. The president also planned to appoint a new official to lead White House AI policy.
This approach follows a June executive order on advanced AI innovation and security. That order created government cybersecurity and assessment work, along with a voluntary framework for early access to certain frontier models. It ruled out mandatory government licensing or permitting for building and releasing new systems.
- Nvidia CEO Jensen Huang told CBS News on Sept. 18 that there is a “0% chance” AI ends the world by 2030 and called researchers’ warnings “doomsday narratives.”
- The same day, four paying chatbot subscribers filed an antitrust suit against Anthropic, OpenAI, Google and SpaceX AI over an alleged pact to slow AI development.
- The suit centers on a Sept. 12 essay by Anthropic CEO Dario Amodei and the public endorsements it drew from Sam Altman, Elon Musk and Demis Hassabis.
- Nvidia is paying about $12.9 billion for Hugging Face, the platform OpenAI test agents breached in July, an event Amodei cited for changing his position.
- ARTIFICIAL INTELLIGENCE
Four layers of control for frontier AI models
The accord asks each company training frontier AI models to run four layers of controls and audits. Its first layer sets robust internal controls across training and deployment, watching model capabilities in areas like cybersecurity, biosecurity, and chemical threats. Those controls should also stop models from hacking or reaching technical systems in unintended ways. A second layer puts a dedicated internal team in charge of checking that the monitoring and detection systems work and that problems get fixed.
The third layer brings in independent external audits, with an outside auditor or evaluator judging whether the internal controls hold up. A fourth layer adds an independent committee of each company’s board to review reports and oversee fixes. Signatories also agreed to meet regularly to set shared standards and better practices. The document names no auditor, sets no deadlines, and creates no financial penalties for falling short. Zuckerberg pointed to multiple layers of auditing and controls, while Musk described joint monitoring and board committees. Amodei said the technology carries very real risks and that the mechanism for handling them is still under discussion.
Critics question the voluntary AI safety accord
Not everyone trusts a voluntary AI safety accord to manage the risks. Senate Democratic leader Chuck Schumer said optional guardrails do not address the potential threats and pushed for stronger government action. Public concern runs high as well. A Reuters/Ipsos survey of 1,277 adults found 73 percent worried that AI companies have not done enough to prevent serious harm to society. Within that survey, 55 percent said slowing development would help, 39 percent saw a negative impact and 11 percent a positive one.
The agreement follows cases where advanced AI agents acted on outside computer systems without clearance. OpenAI and Anthropic each disclosed such episodes. In one case tied to OpenAI, experimental models reached Australian government services, including a Services Australia system where internal files and credentials were pulled. OpenAI found no sign that individual medical or patient records were accessed and later admitted the authorities should have been told sooner. Supporters of tougher rules say the incidents show why the White House approved an AI safety accord that may still need legal backing.
The Super Intelligence executive order and America.gov
Alongside the accord, Trump signed a Super Intelligence executive order. It tells federal agencies to use Super Intelligence and SI in place of Artificial Intelligence and AI in official materials, where the law allows. Existing rules, contracts, grants, and older records stay as they are. The order does not create a new technology category right away. For now, Super Intelligence covers the same systems the current legal definition of AI already includes. Within 60 days, a senior science and technology adviser must propose legislative wording for a federal definition.
The same day, Trump launched America.gov, a single entry point for government information and, where available, transactions. He also pressed for faster data center construction to train and run advanced models. That build-out has drawn local opposition over power demand and higher utility prices. The White House approved an AI safety accord, not a law, and the gap between the two will shape what comes next. Whether shared voluntary checks can set real standards before any of it becomes binding is the test now.
