Last night The Information reported that Anthropic, OpenAI and Google have discussed forming a shared standards organization for AI safety. The idea: the three largest frontier labs define the standards their own models get evaluated against.
For companies weighing whether to point a frontier model at something sensitive, common standards would help. You would know where you stand. The flip side is already in an OECD analysis: high fixed costs, scarce compute and concentrated infrastructure make entering this market hard enough. Compliance costs defined by three large vendors would be one more barrier for everyone else.
What is already on the table
On 14 July, Google DeepMind chief Demis Hassabis proposed a US body modeled on FINRA, the financial industry regulator. Labs submit frontier technology voluntarily, evaluators get 30 days and look at cyber, biological and manipulation risks. Later, passing would become a condition for deploying frontier tech in the US. The Council on Foreign Relations reported that the White House had already been weighing the idea.
Dario Amodei takes a different route in “We Must Pace the Frontier”: independent evaluators such as METR, working inside the labs themselves. And OpenAI’s Chris Lehane has long pushed for a global framework with both the US and China on board — his comparison was international nuclear cooperation.
A forum already exists, incidentally. Anthropic, Google, Microsoft and OpenAI have run the Frontier Model Forum for years, pooling technical evaluations and funding safety research.
Why this one stuck with me
Two things stand out.
First: this is exactly what David Sacks shot down yesterday. The PCAST chair rejects an antitrust exemption that would let the labs coordinate on safety — to him that is a cartel. A body the three biggest vendors run themselves is the version that needs no exemption, because from the outside it looks voluntary. Whether that holds up legally is another question.
Second: the Future of Life Institute published a safety index this summer in which Anthropic scores best at C+ (2.66), ahead of OpenAI at C (2.28) and Google DeepMind at C (2.01). Whoever sets the standard also decides where the bar sits. With grades like those, it is hard to assume the bar goes higher than the weakest of the three can clear.
Which leaves the question no report answers: who pays for safety? The price is delayed launches and revenue that never arrives. As long as the labs are racing each other for the same market, that is the hard part — not what the body ends up being called.
Sources: