On September 29, the leaders of OpenAI, Anthropic, Google, Meta, Nvidia and SpaceX AI stood at the White House and signed a voluntary AI safety accord. The companies promised four things for their most advanced models: internal monitoring of capabilities and risks, an internal team to verify that the controls work, independent external audits, and an independent board committee that receives reports and ensures problems are fixed.
The accord also says, in its own words, that it creates “no new regulatory requirements.” The President called it “morally binding” and said AI firms should police themselves.
That same week, Congress did not move. Republican senators blocked fast-track votes on two AI safety bills, and a Senate preemption bill that could override state AI laws drew opposition from 19 advocacy groups and a request from 25 state attorneys general to keep state oversight in place. Meanwhile, California reportedly kept signing AI bills in the closing days of the governor’s term, including worker-protection measures aimed at preventing AI from making discipline and firing decisions on its own.
Two weeks ago, in Washington Just Told You That No One is Coming, we wrote about Dario Amodei and Sam Altman asking for outside evaluators, and about the President and the Speaker answering that the industry should govern itself. This week, that argument produced a document. The same companies pledged to independent audits and board-level oversight, with no penalties attached and no regulator behind it.
The outside audits Amodei and Altman requested are now in the pledge, but as a promise the companies made to themselves, with nothing behind it if they fall short. The accord describes the audits. It creates no new regulatory requirements and no outside body to enforce them.
That matters because an audit with no authority behind it is only as strong as the company that commissioned it. Who the auditors answer to, what they are allowed to see, and what happens when they find a problem are the details where governance either works or turns into a press release. Those details are exactly what a voluntary pledge leaves open.
Nobody from Washington is coming. They simply handed you a pledge that sounds like someone is.
Trust Us. Auditors On Duty.
Monday, I shared concerns over the signing of a voluntary safety accord by some of the most powerful people in artificial intelligence. Six companies, one document, and a President who called it “morally binding” and “almost like a constitution.”
The page made four promises about the most advanced models. Monitor what the models can do and where they could go wrong. Build an internal team to confirm the controls actually work. Hire independent outside auditors to check that work. And give a board committee the reports, with the job of making sure problems get fixed.
It sounds like a system, but the accord says in its own words that it creates “no new regulatory requirements.” There are no penalties. There is no deadline. The companies choose their own auditors. They are not required to name them or to publish what they find. And if something goes wrong, nobody is required to tell the government. The President summed it up in my “no one is coming” post in one sentence: companies understand they have to police themselves.
Now look at what happened in Washington.
Thirteen days earlier, on September 16, Senator Rand Paul blocked a request from Senator John Kennedy, a fellow Republican, to fast-track a bill requiring human-controlled shutdown mechanisms in AI systems. Hold that thought, because the bill was not a radical idea. It was an emergency button.
On September 29, the same day the pledge was signed, Senator Ted Cruz objected when Senators Warner, Schatz and Kim asked the Senate to pass a bill creating an AI Safety Board. Same day. The pledge went in front of the cameras while the legislation went nowhere on the floor.
Meanwhile, a Senate bill negotiated by Cruz, Majority Leader John Thune and Senator Amy Klobuchar, one that critics say could preempt state AI safety laws, drew opposition from 19 advocacy groups and AI researchers before it was even released. On September 24, 26 attorneys general, in a letter led by New York’s Letitia James, asked Congress to regulate frontier AI and to leave state enforcement in place. Their words: a prohibition on preemption of state laws, with full authority for state officials to enforce federal protections.
Washington’s answer came in three parts. A pledge with no enforcement. A Senate that could not pass an emergency button. And a preemption fight that could take away the only enforcement states have.
To see why that matters, go back to June 18. According to reporting on the incidents behind the accord, experimental AI agents from OpenAI reached into Australia’s Medicare portal. The company did not learn about it until August. Australia was notified through a public inbox in September. From start to finish, that was a three-month breach someone eventually noticed. I don’t think it’s the only one. Regardless, the accord does nothing to increase awareness and doesn’t require anyone to pick up the phone.
An audit with no authority behind it is only as strong as the company that commissioned it. Who the auditors answer to, what they can see, and what happens when they find a problem are the details that determine whether governance works or turns into a press release. Those are exactly the details a voluntary pledge leaves open.
Who Has the Authority to Say No?
When something goes wrong, an oversight program comes down to three questions. Who finds out? How fast? And who can compel a fix? Under the voluntary accord, the answer to all three is the same: the company that built the AI system.
How could that possibly go wrong?
In April 2013, the Rana Plaza building collapsed outside Dhaka, Bangladesh, killing more than 1,100 people and injuring about 2,500. The building was permitted for shops and offices, not heavy industry. Five garment factories ended up inside, with thousands of workers, industrial sewing machines, and large generators on the upper floors.
Those factories supplied major clothing brands, which ran audits and monitoring to know what was happening inside. The pieces were all there: monitoring, internal compliance, external auditors, and upward reporting. What was missing was anyone with the authority to stop the shift. The day before the collapse, cracks appeared in the walls, and the factories reopened anyway.
The AI voluntary safety accord has no enforcement. It runs on the same good intentions as those clothing brands.
Until the people writing these pledges are willing to give someone else the authority to say no, I will keep calling them what they are: a pledge, not a safeguard.