Tech Leaders Sign ‘Morally Binding’ AI Safety Deal Under Trump
Six of the biggest names in AI sat down with Trump on September 29 and signed a one page promise to police themselves. No law behind it. No fines if they skip it. Just a document called the White House Accord on Super Intelligence, and a handshake.
Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman, Elon Musk, and Jensen Huang all put their names on it. That’s Google, Anthropic, Meta, OpenAI, xAI, and Nvidia in one room, agreeing to the same four rules.
What the companies actually promised
The accord runs just over 300 words. Short enough to read twice before your coffee gets cold.
It asks every signatory training frontier models to build four layers of oversight. First, internal controls that watch a model’s capabilities and alignment while it’s being trained and deployed, with specific attention to cybersecurity, biosecurity, and chemical risks. Second, an internal team whose job is making sure those controls actually work and fixing anything that doesn’t.
Third, an outside auditor comes in to check the company’s own claims. Fourth, a dedicated board committee gets the reports and is supposed to make sure problems get fixed, not just logged.
Put plainly: watch yourself, double check that you watched yourself, pay someone to verify you watched yourself, then tell your own board about it.
Same day, a different headline
Trump didn’t stop at the accord. He also signed an executive order the next day renaming “Artificial Intelligence” to “Super Intelligence” across every federal agency. We covered that renaming order in detail here, including the 60 day clock running on whether it becomes actual law.
Jensen Huang leaned straight into the new branding, calling data centers “SI factories” at a White House event the same week. Musk reshared the signing photo with two words attached. Super Intelligence.
Here’s the part nobody’s saying out loud
Nothing in this accord has a deadline. No source lists a required testing schedule, a named auditor, or a consequence for a company that just… doesn’t bother.
“Frontier models” never gets a real definition either. That matters more than people think, because without one, a company gets to decide for itself which of its own systems even count.
And the external auditor each company brings in? They pick it. Nobody outside the company is assigned to check their homework.
House Speaker Mike Johnson called it a statement of principles, voluntary on the industry’s side. That’s a fair description. It’s also the honest one, and it’s worth sitting with before anyone calls this regulation.
Why this still matters anyway
Look, a voluntary pledge with no teeth isn’t nothing. Six companies that compete for the same researchers, the same chips, and the same government contracts just agreed in writing to a shared minimum bar.
That’s a real signal, even if it’s not enforcement. It gives lawmakers something to point at and ask, in six months, whether anyone actually followed through.
It also sets up a baseline the next administration, or the next Congress, can build real rules on top of. Voluntary accords have a habit of becoming the first draft of binding ones, eventually.
For now, the honest read is this: a public commitment, signed at a White House luncheon, with every incentive for the companies involved to be seen doing the right thing, and very little forcing them to actually do it. If you want the broader context on where federal AI oversight currently stands, our piece on AI safety testing and security risk lays out what’s actually being checked before models ship, accord or no accord.
