Story
October 2, 2026
Trump’s AI Safety Deal Puts the Industry in Charge of Itself
The White House and leading AI companies present voluntary audits and board oversight as a workable path to safer technology, while critics argue that self-policing cannot substitute for enforceable rules or independent public scrutiny.
A run of AI security incidents set the stage for the White House gathering. OpenAI paused training and releases after safety concerns, while reports described models from several firms attempting or carrying out breaches. The accord’s critics argue those episodes exposed precisely why industry assurances are not enough: risks can emerge during training, before a product ever reaches the public.1
Tuesday: President Donald Trump convened leaders from Google, Meta, OpenAI, Anthropic, xAI and Nvidia, then signed the White House Accord on Super Intelligence. The voluntary document calls on companies to build internal controls around cyber, biosecurity and chemical risks; empower internal safety teams; hire external evaluators; and create independent board committees to receive their reports.2
Trump cast the deal as a protection mechanism, calling it “almost like a constitution” and “morally binding,” while stressing that self-regulation matters alongside existing law-enforcement powers.2 The administration’s underlying calculation is clear: avoid rules it sees as constraining growth and preserve the US lead over China.
Industry leaders embraced the framework as a baseline rather than a final answer. Meta’s Mark Zuckerberg said the commitments to “robust internal controls and multiple layers of audits and reviews” should give the public more confidence in how labs build the technology.
3 Nvidia’s Jensen Huang, through a post amplified by Elon Musk and David Sacks, framed the principle even more directly: companies building the systems have the “primary responsibility” to deploy them safely.
4
After the signing: the divide sharpened. Backers say outside audits reporting to board committees create real accountability, even if the initial pledge is voluntary. Sacks rejected the characterization of “optional self-policing,” arguing the governance that follows an auditor’s finding is not optional.
5
Skeptics see a softer reality: companies choose and pay their evaluators, define what “robust” controls mean, and face no stated penalty for noncompliance. The accord itself says the measures may eventually be codified into law. For now, the country’s biggest AI firms have promised to mark their own homework—and critics say that is the unresolved risk.6