Musk Says AI Giants Are Grading Their Own Homework. He's Not Wrong.
The White House Accord on Super Intelligence asks AI companies to police themselves and hire their own auditors. Musk calls it grading each other's homework, and the track record on voluntary AI commitments backs him up. But there's a market argument for giving it a chance.
Elon Musk is right about the diagnosis, and probably wrong about the cure. The White House Accord on Super Intelligence, signed Tuesday at an event hosted by President Trump, leaves AI oversight to the companies building the systems. Musk's read: they'll be grading each other's homework.
That's a fair description of the document. It's voluntary. There's no regulator with subpoena power, no fines, no penalty for a company that quietly ignores the fine print. Signatories monitor their own models and bring in outside auditors, which sounds reassuring until you remember who picks the auditors and who pays them.
The Track Record Isn't Great
We've been here before. In July 2023, the Biden White House got seven AI companies to sign a set of voluntary commitments on safety testing and watermarking. Two years on, the results are mixed at best, and several of the same names are signing again. Go back further and you find the March 2023 open letter, signed by Musk and thousands of others, calling for a six-month pause on training models more capable than GPT-4.
The pause never happened. Not for a single day.
So when proponents argue that voluntary frameworks mature into real standards, I'd point at that track record. Voluntary works when incentives line up. It tends not to when there's a race on. And there's very much a race on. Meta, Google, OpenAI, Anthropic, and xAI are all shipping frontier models on overlapping timelines, and nobody wants to be the company that slows down first.
The question worth asking: what does an audit actually prove if the audited company controls the scope, the timing, and whether the findings ever get published?
The Case for Giving It a Chance
To be fair, there's a decent argument on the other side, and it's not just industry spin. Voluntary standards have a real track record in enterprise software. SOC 2 and ISO 27001 began as private frameworks, then became procurement requirements because buyers demanded them. Auditors built practices around them, and those practices hardened into norms. Something similar could happen here. If enterprise customers start asking for safety attestations before they sign seven-figure contracts, that's pressure no law has to supply.
Congress, meanwhile, hasn't passed meaningful AI legislation, and the EU AI Act is still phasing in its obligations. Skeptics who want a regulator with teeth should notice that the regulator doesn't exist yet. A voluntary accord isn't nothing. It's plumbing. Admittedly, plumbing matters if you ever want to run water through it.
My Verdict
I'm not entirely convinced. Self-policing can handle the ordinary stuff, bias testing, data provenance, model cards, incident reporting. It struggles with the thing the accord is named after. Superintelligence isn't a compliance problem you solve with a checklist and a quarterly review. It's a problem where the entities doing the reviewing have every incentive to pass themselves, and a competitor like Musk pointing that out from the outside is useful even if his motives are obvious.
History suggests otherwise on the big stuff, and the big stuff is the whole point of the document.
So here's what I'd watch over the next six months. Does the accord name the auditors, or leave that vague? Does it require any public findings at all? And most telling, does anyone ever fail? If every signatory walks away with a clean report, you've learned nothing about the models and everything about the graders.
Time will tell, though. It usually does, just slower than anyone racing to ship wants.
