Responsible Scaling Policy (RSP)
/ɑːr-ɛs-piː/
What is the Responsible Scaling Policy?
The Responsible Scaling Policy (RSP) is Anthropic's framework for governing AI development as models become more capable. It defines specific capability thresholds — called AI Safety Levels (ASL) — and requires that corresponding safety measures be in place before a model at each level can be deployed.
Think of it like automotive safety standards: as vehicles become faster and more powerful, they require progressively stronger safety features (seat belts, airbags, crash testing). Similarly, as AI models advance through ASL-2, ASL-3, and beyond, they must pass increasingly rigorous evaluations for dangerous capabilities — from biosecurity risks to cybersecurity threats to autonomous behavior.
Key Characteristics
- Threshold-based: Specific capability levels trigger specific safety requirements
- Empirical: Based on measurable evaluations, not subjective assessments
- Progressive: Requirements increase as model capabilities advance
- Legible: Published externally so customers, regulators, and competitors can verify compliance
- Iterative: The document evolves — Anthropic acknowledges early versions won't be perfect
Why the RSP Matters
For organizations deploying AI, the RSP represents a new model for AI governance that's neither "move fast and break things" nor "stop all development." It creates a middle path: advance as fast as you can, but with concrete checkpoints.
The RSP has had outsized influence on the AI industry. After Anthropic published its framework, other major labs adopted similar approaches. It also provides a template for enterprise AI governance — organizations can use its structure to define their own capability thresholds and safety requirements for internal AI deployments.
Historical Context
Anthropic published its first RSP in 2023, making it one of the first AI labs to formalize capability-based safety requirements. The concept was developed primarily by Paul Christiano (who designed the initial framework) and refined through extensive internal drafting — the co-founders note it went through more iterations than any other Anthropic document. The RSP drew inspiration from biosafety levels (BSL-1 through BSL-4) used in laboratory safety, adapting the concept of tiered containment to AI capabilities.
Related Reading
- Scaling Laws - The research showing predictable capability growth
- Enterprise AI - How RSP principles apply to business deployments
- Dario Amodei - Anthropic CEO and RSP champion
Mentioned In
Dario Amodei at 00:00:00
"The RSP is described as 'the holy document for Anthropic' — a framework with capability thresholds that trigger mandatory safety requirements before deployment."
Jared Kaplan at 00:00:00
"The RSP creates healthy incentives — every team in the company aligns with it, and it's legible both internally and externally."