OpenAI, Anthropic, and Google are building their own AI standards body

OpenAI, Anthropic, and Google confirmed on Tuesday, September 16, 2026 that they have been in active discussions since July 2026 about building an AI standards body together. Their model is FINRA, the industry-funded organization that oversees U.S. brokers and investment firms. One day later, Google and Google DeepMind launched the DeepMind Institute, and one of its inaugural essays is Demis Hassabis renewing his call for a U.S.-led body that evaluates frontier AI models before release. Two announcements, 24 hours apart, pointing the same direction.

These companies spend every quarter trying to take customers from each other. Now they are sitting down to design the referee that will judge all of them. Nobody made them do it. No rulebook exists yet for testing frontier AI capability, so they have started writing an interim one. That is what proactive AI governance looks like: getting your house in order before a crisis or a court date does it for you.

What the DeepMind Institute is

The DeepMind Institute launched Wednesday, September 17, 2026. Demis Hassabis, DeepMind's CEO, chairs it. Shane Legg, DeepMind's co-founder, acts as managing editor, and James Manyika of Google sits among the directors. The stated purpose is to surface differing viewpoints on AGI from inside DeepMind and from the wider research community. Said plainly, the people building the most capable systems disagree with each other about what those systems will do.

Hassabis first called publicly for a US-led AI standards body in July 2026. The September 17 launch gives that argument an institution to live in. It also arrives beside the three-way talks rather than in isolation.

The 30-day mechanism

The proposal has a shape you can picture. AI developers would voluntarily submit frontier models up to 30 days before release. The body would start by co-designing test protocols with the AI companies themselves. Then it would shift toward independent evaluations that stay undisclosed, so a company cannot tune a model to pass a test it has already seen. If the voluntary version holds up, passing those evaluations could become a condition of deploying frontier models in the U.S. Hassabis has described a framework that could go further, up to a coordinated slowdown among frontier AI developers if safety findings demanded one. Built well, the testing setup behind it could later connect to public measurement work of the kind NIST already does.

Money is the part most readers skip, and it decides everything else. Hassabis has said the funding "would need to be substantial and likely mostly come from industry." That money buys world-class technical talent and the compute that large-scale testing requires. An industry-funded referee, judging the industry that funds it. Aidan Gomez, co-founder of Cohere, a company notably absent from the three-way discussions, has put the open question plainly:

Who writes [the rules], who gets to participate, and whose interests the rules are protecting.

Gomez deserves an answer, and the people designing this body would be wise to give one in public and early. FINRA carries weight because its rulebook and its enforcement records are visible to the firms it governs and to the customers it protects. A body funded by the same 3 companies whose models it tests will be judged on those terms.

Why this reaches your own AI decisions

Picture the executive reading this on a phone between meetings, with a vendor contract open in another tab. The practical piece lands close to home. Within a few years, the model running your contact center or your student advising tool may carry an evaluation record from a body like this one. Your procurement questions change to match. Dario Amodei has called for slowing AI development where safety demands it, a position that drew public support from Sam Altman, Elon Musk, and Hassabis. When the builders put a slowdown on the table, buyers should be asking what a slowdown would do to a roadmap they have already promised the leadership team.

I wrote about the same pattern when Nvidia, Microsoft, and 38 rivals formed an alliance to set their own AI security practices. It shows up in paperwork too, which is why AI contract terms that keep pace with the technology matter more than another policy PDF nobody opens.

The judgment that stays human

Above the Algorithm is the term I use in my forthcoming book, The Great Relearning, for the work that stays human once the system handles the rest. Someone decides what gets tested and what counts as a passing result. Someone carries the risk an organization accepts on behalf of people who never chose to be in the test set. A standards body is that work, institutionalized. Its evaluators decide what "safe enough" means for a model that will touch mortgage applications and medical intake forms.

I advise California State University on AI governance. The questions that take longest to settle are never about model capability. They are about who signs off, and what happens when the sign-off turns out to be wrong. Across 20 years shaping Cisco's $1.1B innovation program, the projects that survived were the ones where somebody's name sat next to the decision while it was still reversible.

Who this is for

The people most affected by these evaluations will never read a test protocol. A parent whose mortgage application gets scored by a model she cannot inspect. A nurse working next to a triage tool she did not select. Whether a frontier model gets 30 days of independent testing or zero days is a question about their lives, settled in meetings they will never attend.

That is the reason to watch what OpenAI, Anthropic, and Google build here, and to say so while the design is still soft. Voluntary frameworks harden fast. The ones that harden well tend to be the ones outsiders pressed on early.

Bring one question to your next leadership meeting. If an industry-funded AI standards body started publishing evaluation results for the models you already run, what would you have to change, and how fast could you change it? Answer it now, while the cost is a meeting instead of a migration. I'm easy to find.

Here is what makes Alex a credible voice on this topic: He advises California State University on AI governance and spent 20 years shaping Cisco's $1.1B innovation program, where release decisions turned on the same question this standards body is trying to answer at industry scale.

Sorting out who signs off on AI in your organization? book a conversation →