AI's big four form safety club with teeth — or at least a charter

Anthropic, Google, Microsoft, and OpenAI just drew a line in the sand. The four companies — arguably the most influential players in generative AI right now — are forming the Frontier Model Forum, an industry body they say will focus on the safe development of advanced AI systems. Think of it less as a regulatory body and more as a working group for the firms building models that, by their own definition, exceed the capabilities of anything we’ve seen before. They’re promising to pool technical expertise on things like adversarial robustness and anomaly detection, and to build a public library of safety benchmarks.

Kent Walker, Google’s President of Global Affairs, framed the initiative as an industry stepping up while governments work out the rules. And to be fair, the Forum’s four stated objectives are exactly what you’d want from an industry that keeps telling Congress it’s terrified of its own creations: fund safety research, nail down best practices, talk to policymakers, and somehow use AI to fight cancer and climate change along the way. The last one feels a bit like a PR chaser, but the first three are genuinely hard, unsolved problems.

The membership criteria are where things get interesting. The Forum is open to any organization that develops what it calls ‘frontier models’ — large-scale systems with broad capabilities — and can demonstrate a ‘strong commitment’ to safety. That’s a velvet rope that currently excludes a lot of open-source efforts and smaller labs. It also raises the question of whether this becomes a de facto standards body for the most powerful AI, with the four founding members acting as the bouncers.

What’s missing from the announcement is any enforceable mechanism. There’s talk of an Advisory Board and sharing information with governments, but no mention of independent audits, liability, or what happens if a member ships something dangerous. The Forum plans to focus on technical evaluations and safe disclosure mechanisms in its first year. Whether that produces anything more than white papers depends entirely on how seriously these companies treat a group that has no legal authority over them.