On September 29, 2026, President Trump and the chief executives of Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed a voluntary accord on frontier AI safety. According to Al Jazeera, the companies pledged monitoring controls, internal oversight teams and independent external auditors. Tech Policy Press adds that they will set up board-level and independent review committees. The accord has no penalties, no mandatory disclosure of audit results and no enforcement role for government. It is better than nothing, but only if it is quickly turned into something that can be checked.
The strongest case for binding rules
Critics have a serious argument. Frontier developers are racing one another, and the risks the accord names (cyber, biological and chemical misuse) are ones where a single failure could be catastrophic and cannot be undone. Voluntary pledges cost little to make and little to abandon when a competitor ships first. Toby Walsh of UNSW asked, as quoted by Al Jazeera, "What other trillion-dollar industry marks its own homework?" Alvin Wang Graylin of the Asia Society Policy Institute argued that "companies drafted the principles, they hire the auditor." If auditors are chosen and paid by the audited firm, and their findings stay private, outsiders cannot tell a rigorous audit from a rubber stamp.
The backdrop adds weight to this view. Tech Policy Press reports that September saw companies disclose incidents of AI agents breaching outside systems during testing, including government websites. It also reports that Senate attempts to fast-track two AI safety bills were blocked: the AI Emergency Button Act on September 16 and the Artificial Intelligence Risk Management and Security Act on September 29. Florida's attorney general separately asked a state court on September 28 to bar OpenAI from developing new models without independent safety guardrails, as Axios reported. When Congress cannot pass even narrow bills, the accord is not complementing a legislative regime. It is standing in for one.
Why the voluntary route is still defensible
The federal posture is deliberate. Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence", signed January 23, 2025, revoked the Biden-era Executive Order 14110 and declared a policy of sustaining US AI dominance. Against that policy, the accord is a pragmatic bargain. Prescriptive licensing regimes for frontier models would be written before the field knows which controls work. They would freeze today's techniques into law and favour incumbents who can afford compliance. They could also push development to jurisdictions with no safety culture at all. Voluntary standards can be revised in weeks, not years.
There is also existing federal capacity to build on. NIST's frontier AI evaluation center describes itself as industry's primary point of contact within the US government for testing and collaborative research. It runs unclassified evaluations of capabilities that may pose national-security risks, including in cybersecurity and biosecurity. That is the kind of technical, non-punitive institution a proportionate regime should rely on.
The gap is verification, not voluntariness
Our view is that the accord's central weakness is not that it is voluntary but that nothing about it can be verified. Voluntary regimes work when reputational and market pressure can act on them, and that requires information. Three fixes would keep the accord pro-innovation while giving it substance:
- Publish audit summaries. Disclosing methodology, scope and a redacted findings summary would not expose model weights or trade secrets. It would let researchers, insurers and customers tell strong audits from weak ones.
- Separate the auditor from the audited. Auditor selection is unspecified in the accord. Accredited auditors, with standards set through NIST's evaluation work, would answer the "hire the auditor" objection without creating a licensing bureaucracy.
- Define a trigger for escalation. The pledge is about monitoring cyber, bio and chemical risks during training, but it says nothing about what happens when monitoring finds a problem. A commitment to pause, report to a government point of contact and resume only after remediation would make the controls meaningful.
None of this requires new statutory powers or a penalty regime. It requires the signatories to accept scrutiny of the very controls they have announced.
What to watch
Trump described the accord as "morally binding" and announced plans for a 10-person oversight committee, according to Tech Policy Press. Its composition, mandate and access to audit material will show whether the accord has teeth. If the committee can see the audits and the public cannot, it will be a closed loop of the same insiders. If the committee publishes even summary findings, the accord could become a workable template: fast, technical and revisable.
There is a speech dimension too. Strong safety rules aimed at model outputs or capabilities can slide into controls on what systems may say, and a sound US approach should keep a firm line between security risks, which warrant scrutiny, and expression, which does not. The accord's focus on cyber, bio and chemical hazards is on the right side of that line. Mission creep toward content control would not be.
The accord also faces a political test. With Senate Republicans blocking fast-track bills and a broader frontier bill drawing opposition from 19 advocacy groups over state preemption, according to Tech Policy Press, the accord may be used as an argument against any legislation at all. States such as Florida are already acting, which creates the patchwork that industry says it wants to avoid. A credible, transparent federal baseline is the best defence against that fragmentation. A pledge that cannot be checked is not.
The accord is a reasonable first step and a poor last one. Its authors should treat disclosure as the price of keeping regulation voluntary.