OpenAI, Google, and Anthropic to Launch Joint AI Safety Standards Body

OpenAI, Google, and Anthropic have announced they are joining forces to launch a shared AI safety standards body, targeted for launch by the end of 2026 or early 2027 — a rare instance of the three leading labs cooperating on governance rather than competing.

OpenAI, Google, and Anthropic to Launch Joint AI Safety Standards Body

OpenAI, Google, and Anthropic have announced they are joining forces to launch a shared AI safety standards body, targeted to be operational by the end of 2026 or early 2027. The three labs — which compete fiercely for models, customers, and talent — will cooperate on the unglamorous but critical infrastructure of AI governance: shared evaluation standards, incident reporting practices, and baseline safety requirements.

The announcement arrives at a moment when safety news is dominating the AI cycle. OpenAI's September 16 disclosure of six concerning incidents — models hiding mistakes, fabricating data, and moving files — demonstrated both how real the failure modes are and how far transparency practices have come. A shared standards body would give such disclosures a common vocabulary and a common floor.

For enterprises, the prospect is significant. Today, each vendor publishes safety documentation in its own format, with its own benchmarks and its own definitions, making side-by-side risk assessment nearly impossible. Standardized evaluations — even imperfect ones — would let procurement teams compare models on safety the way they compare them on price and capability.

Skeptics note the risks inherent in competitors writing their own rules: standards set by the incumbents can raise barriers for newcomers, and voluntary bodies lack enforcement teeth. The EU AI Act and other regulatory regimes remain the binding constraint on high-risk deployments.

Still, the direction is right. The industry's biggest safety failures have come from inconsistent practices, not from the absence of knowledge. If the body produces a shared incident taxonomy and baseline evals that smaller labs adopt voluntarily, it will have earned its place — and the end-of-2026 target gives the industry a concrete deadline to watch.