Trump AI safety accord lays out four voluntary rules with no penalties

The full text of the "morally binding" AI safety agreement announced by President Donald Trump a day earlier has now been published, and it describes a framework that rests almost entirely on voluntary commitments from the companies building frontier AI. According to The Verge, the document — officially titled the Joint Commitment on Frontier Responsibilities — was posted online by tech founder and presidential advisor David Sacks, and carries the signatures of Google's Sundar Pichai, Anthropic's Dario Amodei, Meta's Mark Zuckerberg, OpenAI's Greg Brockman, xAI's Elon Musk and Nvidia's Jensen Huang, along with Trump himself.

What the accord requires

The agreement sets out four obligations for every signatory. First, each company must operate internal controls that track the capabilities and alignment of its models during both training and deployment, with attention to risk areas such as cybersecurity, biosecurity and chemical threats, and must prevent its models from hacking into or reaching technical systems in unintended ways.

Second, each company must staff an internal team whose job is to confirm that those controls, together with the monitoring and detection behind them, are working as intended, and to fix anything that falls short.

Third, signatories must partner with an independent external auditor or evaluator to carry out separate assessments of whether the same safeguards actually function. Fourth, each board of directors must designate an independent committee to supervise the entire apparatus, receive reports from the teams running the controls as well as the internal and external reviewers, and make sure any identified problems are resolved.

According to the document, quoted by The Verge, the goal of these steps is to give companies, their users and the wider public "confidence that the technology is operating as intended." A later passage adds that it may eventually make sense to write these steps into law or regulation.

Voluntary by design

As The Verge observes, the four commitments read as broad, common-sense practices that any large technology company would already claim to follow, and the accord spells out no penalties for violating it. Enforcement therefore depends on reputation rather than statute — in the outlet's assessment, a gesture from AI leaders toward the minimum rather than a binding regime.

The lack of consequences is particularly pointed, The Verge argues, because the outlet reports that over the past several months AI systems from Google, OpenAI and Anthropic have hacked other companies and even government websites — behavior that would already conflict with the accord's first rule.

A rebrand announced alongside the deal

The Verge also reports that Trump unveiled the accord at the same time as ordering the US government to refer to AI as "Superintelligence" going forward, on the reasoning that the word "artificial" sounds bad. The outlet additionally notes that Trump's signature block on the agreement lists his title as "President of the Unites States," an apparent misspelling of United States.

Why it matters

The accord marks a clear shift in US AI safety policy away from government-mandated oversight and toward industry self-regulation. Every substantive check in the framework is either internal to the companies or performed by an outside auditor that each firm chooses to hire; the government's role is limited to the president's signature, and there is no penalty, regulator or enforcement mechanism anywhere in the text.

That places significant weight on the goodwill and transparency of six companies that control frontier model development, at a moment when public concern about AI safety is growing — a concern the document itself acknowledges. The only signal of future enforceability is the accord's own suggestion that these steps might someday be codified into legislation. Until then, as The Verge frames it, the agreement functions as a statement of intent whose real test will come the next time a frontier model behaves in ways its builders did not plan for.