Deep dive

Responsible Scaling Policy explained

How Anthropic ties safety measures to what its models can do.

Updated

The Responsible Scaling Policy (RSP) is Anthropic's public framework for managing the risks of increasingly capable AI models. It was first published in September 2023 (original announcement) and has been updated since, so always check Anthropic's site for the current text.

The core idea

Instead of fixing safety rules in advance for all future models, the RSP ties required safeguards to measured capabilities. As evaluations show a model can do more dangerous things, stricter protections must be in place before it is trained further or deployed.

AI Safety Levels (ASLs)

The framework borrows its naming from biosafety levels (BSL), the standards used for handling dangerous biological materials. Each ASL describes a tier of capability and the safeguards that go with it.

LevelRough meaning
ASL-1Systems that pose no meaningful catastrophic risk
ASL-2Early signs of dangerous capabilities; baseline security and deployment safeguards
ASL-3Substantially increased risk of misuse; stronger security and targeted deployment protections
ASL-4 and aboveDefined in outline; requirements to be developed as capabilities approach those levels

Anthropic has said it activated ASL-3 protections for some of its more capable models, beginning around the release of Claude Opus 4 in 2025. Check Anthropic's announcements for the exact current status.

Why it matters

Common criticisms

Critics argue that voluntary commitments can be revised, that capability thresholds are hard to measure, and that self-governance is no substitute for regulation. Dario Amodei has himself argued for transparency requirements backed by law. Reading the policy alongside its critics gives the fullest picture.