What is Responsible Scaling Policy?

What it is

A Responsible Scaling Policy (RSP) is a written framework an AI lab publishes to say, in advance, what it will do as its models get more capable. Anthropic introduced the name and the approach. Other labs publish similar documents under different names, often called frontier safety frameworks.

The core idea is an "if, then" commitment. If a model shows a dangerous capability, then the lab must have specific safeguards in place before it continues training or releases the model.

How it works

An RSP has a few moving parts:

• Capability thresholds. Defined lines that mark when a model could cause serious harm. Typical areas include help with chemical, biological, radiological or nuclear weapons, and the ability to do AI research and development autonomously.
• Evaluations. Tests the lab runs on a model to check whether it is near or past a threshold.
• Required safeguards. Security and deployment protections tied to each threshold, such as stronger protection of the model weights or tighter limits on how the model can be used.
• Safety levels. In Anthropic's version, tiers called AI Safety Levels, modeled loosely on the biosafety levels used for laboratory work with dangerous pathogens.
• Reporting and updates. The policy is versioned, and the lab revises it as it learns. Labs also publish results, often in a system card.

You can think of it as a building code for model development: rules that say what has to be true before you can add the next floor.

Why it matters to you

If you build on a frontier model, an RSP is the closest thing to a public rulebook for how your provider decides whether a new model ships and with what restrictions. It can explain why a release is delayed, why a capability is limited, or why a model arrives with extra usage conditions.

It is also worth reading with a skeptical eye. An RSP is a voluntary, self-written commitment. It can be rewritten by the same company that wrote it, and outside parties cannot always check the test results behind it. Its value depends on how specific the thresholds are and how honestly the lab reports against them.

Related ideas

An RSP sits next to AI alignment work and red teaming, which supply the evaluations that decide whether a threshold has been reached. For a lab's description of how a model should behave day to day, see model spec. For the concern about AI speeding up AI research, see recursive self-improvement.