Which AI labs take safety seriously?
A handful of companies are racing to build systems more capable than any human — and they are not approaching the risk the same way. This scorecard compares the frontier labs on what they've actually committed to: safety frameworks, pause commitments, government testing and transparency. Sort it, filter it, and judge for yourself.
Last reviewed · we update this as labs publish new commitments.
The clearest divide isn't East versus West. The strongest safety commitments — Anthropic's and Google DeepMind's — and some of the weakest — xAI's and Meta's — are all American. Meanwhile China's Zhipu AI signed the very same Seoul safety commitments as OpenAI and Google, and every Chinese lab operates under binding national AI rules.
Recklessness, in other words, isn't a nationality. That matters, because the argument for racing ahead usually rests on the idea that someone else — invariably China — can't be trusted to slow down. Read the fuller case on our China myth page.
Each lab is scored 0–2 (not met / partial / met) on five dimensions, for a total out of 10: a published safety framework with capability thresholds; a pause commitment to halt if those thresholds are crossed unsafely; government testing access for national AI safety institutes; transparency via system cards and risk evaluations; and overall safety governance (mission, structure, team stability).
Scores reflect publicly available commitments as of the review date and are the Foundation's assessment, not an audit. Think a rating is out of date or wrong? Tell us — we revise it.
Voluntary safety is only as good as its enforcement.
Even the best-scoring lab here is grading its own homework. Frameworks can be revised, paused, or quietly dropped when the competition heats up. That's exactly why we need binding, external guardrails — not just corporate promises. Get our briefings on the fight for real oversight.