Wiki · Reference entry
AI Safety Level
Anthropic's tiers, modeled on lab biosafety levels, that pair how capable a model is with the safeguards it must have.
Laboratories that handle dangerous germs are graded by biosafety level, or BSL. The higher the level, the more dangerous the organisms allowed and the stricter the rules, from gloves and a lab coat at the bottom to sealed suits and airlocks at the top. AI Safety Levels, written ASL-1, ASL-2 and so on, borrow that structure. Each level pairs a band of dangerous capability with the security and deployment protections a model in that band requires. The levels are defined in Anthropic's responsible scaling policy, and the definitions change when the policy is revised.
Why it matters
It ties the safeguards to measured capability, so the scheme is only as good as the tests that do the measuring.