Who can enforce Trump’s voluntary AI safety accord with tech executives?

That distinction matters because the people setting and reviewing the safeguards are largely the people who develop and govern the technology. An external firm would perform audits, but the reported terms do not say whether its findings must be made public or what a company must do if a review exposes a problem.

What the companies agreed to check

The accord calls for companies to evaluate their AI systems internally, have an external firm conduct audits and submit reports for review by their boards. Those are different kinds of scrutiny, but none is described as an inspection or enforcement power assigned to a public regulator by the accord.

Internal evaluations can draw on a developer’s access to its own systems. An audit by an outside firm adds a reviewer beyond the company’s staff. A board review brings the findings to the people responsible for overseeing the company. The reported terms, however, do not establish how much of that work people outside the company will be able to see.

The companies’ regular meetings could produce shared safety practices. The reporting does not identify a common public test for deciding when a system has failed a safety check, or specify how an audit firm’s findings would change a company’s decisions. Those gaps should not be mistaken for proof that evaluations or audits have no value. They are limits on what the announced pledge tells the public about accountability.

There is a reason to ask how the checks would work in practice. CBS News reported that OpenAI and Anthropic had described incidents involving unauthorized access or dangerous use of AI systems. The accord proposes reviews, but the reported terms do not establish how those incidents would have been evaluated or whether the public would learn the results.

House Speaker Mike Johnson presented the agreement as a joint commitment and statement of principles centered on companies’ internal controls. That describes what the administration and executives intend to do together. It does not answer what someone outside that arrangement can do if a company does not follow through.

What morally binding means for accountability

The accord describes its commitments as voluntary. Its reference to possible future laws or regulations underscores the distinction: A company can promise an audit without Congress requiring one, specifying what the auditor must examine or establishing a penalty for noncompliance.

The narrower point is important. The absence of an identified enforcement mechanism in this accord does not mean AI companies are exempt from existing law. It means the new safety pledge itself does not specify a process through which the public can demand an evaluation, obtain an audit finding or seek a penalty when a company fails to meet its commitments.

An audit can inform a company’s board even if its findings are never published. But if findings remain out of public view, people affected by a company’s decisions have less information with which to judge its response. The reporting does not say the findings will remain private; it also does not identify a requirement to disclose them. That unresolved choice bears directly on how meaningful the promised outside review will be to the public.

The same question applies to standards developed in meetings among the signatories. Companies have expertise about their own systems. They also have a stake in deciding how those systems are built and used. The accord’s described terms do not identify a role for people outside the industry in setting shared standards or assessing whether they have been met.

Why the administration favors company-led checks

Trump argued against slowing AI development. Nvidia CEO Jensen Huang, another participant in the discussion, said that innovation and safety can advance together. Their case is that developers can continue improving the technology while taking responsibility for examining its risks.

The approach has a practical appeal: Companies can examine systems they build, and an external audit could catch problems that an internal evaluation misses. But technical access and public accountability serve different purposes. The fact that a company can identify a problem does not settle who must be told about it, who decides whether its response is adequate or what happens if it fails to respond.

That is where the strongest criticism of the accord begins. Decisions about acceptable risks need not rest solely with political leaders and the firms developing AI. Those decisions can affect people who were not at the table when the commitments were written. The reported pledge provides company and outside-firm reviews, but it does not establish a public process for weighing a failure or requiring a remedy.

Sen. Mark Warner, a Virginia Democrat, called for Congress to require testing, evaluation and incident reporting for the most advanced AI models. His proposal would make those duties public requirements rather than leave them solely to company promises. As CBS News reported, Warner said, "The companies building the most powerful AI systems are warning us that the technology is advancing faster than our safeguards".

What could make the pledge enforceable

Trump said he planned to name an AI czar within days. That was a stated plan, not an appointment established by the reporting, and the account does not define what authority that person would have over the signatories’ safety commitments.

Congress faces a more specific choice: whether to require the testing, evaluation and incident reporting Warner sought. The accord allows for the possibility of future legislation or regulation, but signing it enacted neither. Until enforceable duties are set, the companies have promised checks while the reported terms leave unanswered who can see the results and require action when those checks find trouble.