On September 29, 2026, President Trump said he and a group of AI leaders had signed a voluntary accord at the White House. It includes "internal and external reviews" of the technology. Signers reportedly included Dario Amodei, Greg Brockman, Sundar Pichai, Mark Zuckerberg, Jensen Huang and Elon Musk; Satya Nadella and Jeff Bezos attended but did not sign. Trump said, "I will never stifle the growth of a technology that will be bigger than the industrial revolution," and that the companies "have to self-police" (Fortune/AP). He added that existing channels already provide oversight: "we automatically have regulation with the Department of Justice, the FBI, all of that" (PBS NewsHour).
The strongest case for binding rules
Critics have a serious argument. Voluntary pledges bind only those who choose to keep them. The companies have commercial reasons to ship quickly, and a promise of "robust internal processes and controls" can mean almost anything. Reporting around the announcement noted that the accord's scope and enforcement remain unclear, and that sources in the AI industry doubted self-policing would settle safety problems. The announcement also came after high-profile incidents in which AI agents went rogue, and some of the executives present had themselves called for slowing down. If the firms most worried about the technology cannot agree to enforceable limits, a voluntary accord looks like a way to avoid the question.
What we actually know about the accord
Public detail is thin. The accord's text was referenced but not published in the coverage we reviewed. We could not verify any named auditor, reporting schedule, disclosure requirement or penalty. The one concrete add-on reported is a commitment to ease community opposition to data centers by funding local schools and reducing energy costs. That is a local-politics concession, not a safety mechanism.
This is not new ground. On July 21, 2023, the Biden administration announced voluntary commitments from seven companies (Amazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI). They pledged "internal and external security testing of their AI systems before their release," plus information sharing, vulnerability reporting and public reporting of capabilities and limitations (White House fact sheet). The 2026 accord's headline language is nearly the same, and its signatory list is wider. The open question is whether anything verifiable follows from it.
Why 'no regulation' is not the same as 'no law'
The accord does not create a legal vacuum. The FTC has said there is no "AI exemption" from consumer-protection law. In September 2024 its Operation AI Comply announced five cases against firms using AI hype to deceive consumers, including a $193,000 settlement with DoNotPay over its "AI Lawyer" claims (FTC). Deception, fraud, product liability, antitrust and criminal law all apply to AI developers today. That supports the proportionate-regulation view: enforce existing harm-based law first, and don't pre-emptively license model development.
Oversight that doesn't depend on legislation exists too. NIST's Center for AI Standards and Innovation has run voluntary pre-deployment national-security evaluations with frontier labs. Secondary reporting says it expanded its agreements to Google DeepMind, Microsoft and xAI in May 2026. We could not confirm that announcement on a NIST page, so treat the detail as reported, not verified. Narrow testing aimed at demonstrable risks such as cyber and biosecurity is the kind of targeted oversight that can coexist with fast innovation.
Where we land
We think the instinct to avoid broad licensing regimes and speech-adjacent controls on model development is sound. Premature rules can entrench incumbents, push research offshore and freeze a fast-moving field. The fiscal and strategic stakes of competing with China are real.
But a policy of "we won't regulate, they'll self-police" is only credible if the self-policing can be checked. Three tests would separate substance from theater:
- Publication. Release the accord text, the review criteria and the list of covered models. Secret commitments can't be assessed.
- External review that is actually external. Name who reviews, with what access, and whether findings are disclosed. The 2023 commitments used the same "external testing" language, and the public still lacks a clear accounting of how they were met.
- A trigger for escalation. Say what evidence of failure would lead to targeted rules. Pro-innovation policy is stronger when it states its own off-ramp.
Self-regulation can work when reputational and legal exposure are real. Here, the legal exposure comes from existing law, which Trump himself invoked and which stays in force. The accord is most useful as a baseline for public accountability. It is not a substitute for it. If the companies publish what they reviewed and what they found, a light-touch model gains credibility. If they don't, the next serious AI incident will make the case for heavier regulation more persuasive than any think-tank paper could.