Who Audits the Auditors? AI Self-Regulation

On 29 September 2026, President Donald Trump hosted the heads of the biggest AI companies at the White House to discuss the future of artificial intelligence. Six of them, Anthropic, Google, OpenAI, Meta, xAI and Nvidia, signed a short, one-page agreement. The meeting came amid growing concern about AI models slipping out of human control, and it ended with the companies signing a voluntary agreement focused on AI safety reviews.

The accord asks each company to set up internal controls, use independent audits and have its board oversee the work. Trump rejected calls for government restrictions and said AI development should push ahead quickly. Rather than answering to a regulator, the companies will oversee their own safety.

The deal itself is simple. Each company promises to monitor its own models, appoint an internal team to check those controls, bring in an independent auditor of its choosing, and set up a board committee to oversee the whole process.

The agreement is not legally binding, and there are no penalties if a company breaks it. Trump's position is that heavy government rules would slow down American AI companies and let China catch up. He sees self-regulation, backed by existing law enforcement like the Justice Department and FBI, as enough for now.

For now, this agreement is the main US federal answer to the risks of advanced AI, the most powerful general-purpose models. Its reach is global: the same six companies supply AI to governments, banks, hospitals, and schools almost everywhere.

Not everyone is convinced by this accord. Researchers and lawmakers are calling for stronger safeguards and asking whether protections can keep pace with the technology. Speaking on PBS NewsHour the evening of the signing, AI researcher Gary Marcus, professor emeritus at New York University and author of Taming Silicon Valley, said self-regulation is “certainly not enough.” Marcus considers AI-driven human extinction unlikely. His concern is more practical. AI systems make errors, and AI-assisted hacking could, for example, take down a power grid and with it a hospital, much as the 2017 WannaCry attack disrupted NHS hospitals in the UK. For that reason, he argues, AI needs the same outside oversight we already expect for aviation, food, and medicines.

So the real question is this: if a promise carries no legal force, what stops a company from breaking it, and who would even know?

On paper, the accord creates no legal duty at all. There is no regulator, no deadline, and no penalty. Audit findings go to each company's own board, not to any public authority.

The President is right that ordinary law still applies. US anti-hacking law makes unauthorised access to computer systems a crime, and consumer protection law bans deceptive business practices. But these rules step in after harm is done, and none was written with AI in mind.

This came at a sensitive time, because many of the executives present had called for a slowdown in AI development after several high-profile incidents in which AI agents slipped beyond human control. Critics like Gary Marcus argue that promises alone are not enough. He suggests borrowing from medicine instead. A new drug cannot be sold in the US until regulators approve it. AI developers, he argues, should likewise prove that benefits outweigh risks before launch, then face ongoing audits and full disclosure of incidents. He notes that something similar already happens, but behind closed doors: the White House reviews new models under a framework it has not published, and, according to Marcus, delayed Anthropic's Mythos system while letting others through. Without public criteria, customers cannot know what will be released or why.

Elsewhere, the same companies already face binding rules set by law rather than by themselves. In the EU, the AI Act has been enforced since August 2026: developers of the most powerful models must report serious incidents to the European Commission and breaches can cost up to €15 million or 3 per cent of global turnover. Within the US, California's SB 53 requires the largest AI developers to report safety incidents to state authorities, with fines of up to $1 million per violation.

There are still a few indirect ways the agreement could carry some weight, although none has yet been applied to AI.

• Misleading the public: a company that claims to follow the agreement but does not could face action from the US Federal Trade Commission.
• Misleading investors: inaccurate safety claims made to shareholders could breach securities law.
• Board responsibility: under Delaware law, where many large US firms are incorporated, directors must oversee major risks in good faith. As AI becomes a major business risk, that duty could extend to AI safety.

There are genuine reasons to support self-regulation. AI moves faster than legislation, and companies understand their own systems best. Supporters, including the President, also argue that heavy regulation could cost the US its lead over China, and that public commitments create pressure, with rivals watching each other.

The case against rests on three gaps.

• Independence: each company picks its own auditor and reports to its own board.
• Accountability: people harmed by an AI failure have no right under the accord to be told, or to seek redress.
• Competition: when AI is framed as a race with China, voluntary brakes are easiest to release just when they matter most.

Marcus illustrated the trust problem with a recent New York Times report which, he said, showed OpenAI knew about the weaknesses behind the widely discussed Hugging Face incident months beforehand, but pressed on to get products to market. He added that thousands of similar incidents have since come to light. He also rejects the race argument. China itself licenses AI software, he points out, and nobody builds unsafe cars just to beat China at making cars. In his view, the contest will end more like Coke and Pepsi than winner-takes-all.

The agreement is best read as a baseline, not a safeguard. Putting commitments in writing has value. What is missing is independence, transparency, and consequences. As Marcus put it, the open question is whether meaningful rules arrive before a major incident, or only after one.

• For policymakers: the four layers of controls and audits set out in the White House accord are a ready-made starting point for legislation. Add public incident reporting, accredited auditors, and penalties.
• For organisations buying AI: treat the accord as context, not assurance. Ask suppliers who audits them and how fast they report incidents, and write the answers into contracts.
• For readers outside the US: local law still applies. In the EU, the AI Act binds these companies whatever was signed in Washington.

This is a textbook case of self-regulation: the industry writes, applies and checks its own rules, rather than a government regulator doing so. The President said the industry was capable of what he called “tremendous self-regulation.” The real disagreement is whether voluntary promises are enough, or whether binding rules are needed.

• PBS NewsHour, 'Self-regulation not enough for AI safety, Gary Marcus says', interview with Gary Marcus by Geoff Bennett, 29 September 2026
• Computer Fraud and Abuse Act, 18 U.S.C. § 1030
• Federal Trade Commission Act, s. 5, 15 U.S.C. § 45
• Federal Food, Drug, and Cosmetic Act, 21 U.S.C. § 355
• EU AI Act, Regulation (EU) 2024/1689, Arts. 53, 55 and 101
• California Senate Bill 53, Transparency in Frontier Artificial Intelligence Act (2025)
• Securities and Exchange Commission, Rule 10b-5, 17 C.F.R. § 240.10b-5
• In re Caremark International Inc. Derivative Litigation, 698 A.2d 959 (Del. Ch. 1996)