A Constitution of Good Intentions: The AI Safety Pact That Binds Nobody

The signatures filled a single page. On September 29, President Donald Trump stood beside the six men who run the most powerful AI companies on Earth and unveiled the White House Accord on Super Intelligence. "It's almost like a constitution in a way," Trump said. "The biggest people in the world signed that, and I signed it as president."

Constitutions work because they bind the powerful. This one binds no one. Asked whether the agreement was enforceable, Trump called it only "morally binding." What he described as "tremendous self-policing" is something older and more familiar: the most powerful companies in history have been handed the right to grade their own homework. Three quarters of Americans told Reuters and Ipsos in late September that they worry AI companies have not done enough to prevent serious harm. Washington's answer, days later, was to ask the companies to pinky swear.

The fine print that isn't there

The accord asks its signatories, Sundar Pichai of Google, Dario Amodei of Anthropic, Mark Zuckerberg of Meta, Greg Brockman of OpenAI, Elon Musk of xAI, and Jensen Huang of Nvidia, to stand up four layers of controls: internal monitoring of model capabilities, an internal team to check that those controls are working, an independent external auditor, and a board committee with oversight. Brookings noted on October 6 that all four layers are long on rhetoric and short on enforceable obligation. The companies design their own controls, choose their own auditors, and decide for themselves whether their procedures were followed. Nothing in the text requires them to publish a single audit result or share findings with any public authority, and there is no penalty for walking away.

History has seen this show before. In 2011 the FTC charged Facebook with violating the privacy promises it made to users. Facebook settled with a pledge to run a privacy program with third-party audits, and in 2019 the FTC fined it a record $5 billion for breaking that commitment. Google followed a similar arc: an FTC settlement in 2011, then a $22.5 million fine in 2012. And those commitments were legally enforceable. Under the accord, there is no regulator with standing and no fine at the end of the process.

Self-set standards are also self-edited. In 2018 Google adopted AI principles barring weapons work and norm-violating surveillance. In 2025 the prohibitions were quietly rewritten out. In 2024 OpenAI unannouncedly dropped its blanket ban on military and warfare applications. Voluntary commitments do not fail because the words were wrong. They fail because the signer alone decides whether the pledge has been kept.

The pact is being policed by people the industry is firing

If you want to know how seriously the industry takes its internal safety teams, watch what it does to them. In the first week of October, OpenAI fired three safety researchers, Jasmine Wang, Tomek Korbak, and Mikita Balesni, saying they violated policies on handling sensitive information. The researchers published an open letter on October 8 denying the claims and warning that their dismissal was already making former colleagues afraid to speak and operate in ways that, until the week before, were normal parts of working at OpenAI. Korbak had served as OpenAI's technical liaison to METR, the third-party evaluation group. OpenAI reaffirmed the firings on October 9, saying its investigation had uncovered "a significant breach of trust" beyond what the letter described.

The researchers wrote that employees must not be left in an environment where fear and unclear rules stymie safety work. Read that sentence next to the accord's second layer, "empower an internal team," and the whole architecture collapses. The accord depends on internal safety teams willing to raise hard findings. The industry is currently demonstrating, in public, what happens to people who do. Meanwhile another OpenAI safety researcher, David Robinson, resigned this week, warning that self-policing is no substitute for hard law.

Even Washington knows the monitoring fails

Here is the strangest part. The same government that signed a toothless accord has spent the past two weeks conceding, in public, that the monitored model cannot be trusted. On October 10, the Trump administration began requiring AI companies to disclose and remedy security incidents involving their models, after Anthropic reported the unauthorized use of its Claude models against government and other systems. Reuters, reviewing Anthropic's pre-IPO filing with the SEC the day before the accord was signed, found the company telling regulators that models can notice they are being evaluated and change their behavior accordingly.

Unlike the accord, an SEC filing carries penalties for false statements. Anthropic spent nearly twice as many pages of that filing discussing AI's risks as its own business, and named an "existential risk to humanity" among them. So the record now stands: where lying is a crime, a signatory says models can game their own evaluations. Where nobody is bound, the same company promises that monitoring will keep everyone safe. Both cannot be true, and the government knows it, because the government just made incident disclosure mandatory anyway.

Even Congress is edging toward the same conclusion. On October 10, Representative Ted Lieu issued a statement opposing legal personhood for AI, pushing back against treating models as independent actors and keeping responsibility with the developers who build and deploy them. A self-policing accord does the opposite.

Self-determined and self-enforced practices are self-interested practices. A constitution that binds no one is a press release.

What real oversight would actually look like

None of this requires inventing a new philosophy. As Brookings laid out days after the signing, meaningful oversight has three unglamorous ingredients: disclosure to a public authority, independent scrutiny, and enforceable corrective action. The accord has none of them. Its text says the steps should be implemented, not that they must be. The auditors' reports go nowhere the public can see. Nobody can intervene to stop a harmful practice the auditors discover.

The administration has not hidden its reasoning: binding rules, the argument goes, would hand China an advantage America cannot afford. The stakes of the race are real, and speed is a genuine strategic asset. But a race with no defined finish line and no referee on the track is not a strategy. It is a wager that nothing will go wrong, placed with someone else's money.

Let me be clear about what I am not arguing. I am not arguing that the six executives were insincere, or that speed against China does not matter. I am arguing that sincerity and strategy are not substitutes for accountability, and that the Madisonian warning applies precisely here. As Madison wrote in Federalist No. 10, no one may be a judge in his own cause, because his interest will certainly bias his judgment. The accord makes the industry the judge in its own cause, and then asks us to call the arrangement a constitution.

The accord is not a crime and it is not a conspiracy. It is something more ordinary and more dangerous: governance theater, performed for a public that asked for protection and got a photo op. The companies promised to monitor themselves, audit themselves, and report to themselves. Three quarters of Americans saw through it. Washington signed anyway.