Responsible scaling policy

Q: Is a responsible scaling policy legally binding?

Mostly no. RSPs are voluntary company policies. Anthropic notes that some laws, such as California's SB 53 and the EU AI Act's codes of practice, now require frontier developers to publish risk frameworks, which pushes in the same direction.

In short

A responsible scaling policy (RSP) is a public, voluntary promise by an AI company about which safety measures it will put in place as its models get more capable, and when it would stop or slow down if those measures are not ready.

Also called: RSP, Frontier safety framework
Topic: Safety, ethics and risk
Level: Some background helps
First described: The term was set out by METR in September 2023, and Anthropic published the first named Responsible Scaling Policy the same month²¹.

The idea is an "if-then" rule. If a model reaches a dangerous level of capability, such as being able to help someone make a biological weapon, then the company must have stronger safeguards in place before it trains further or releases the model³.

The research group METR described RSPs in September 2023 as a statement of what capabilities a developer can handle safely with its current protections, and of the conditions under which it would be too dangerous to keep scaling up until protections improve². METR also said voluntary commitments would not be enough on their own and were not meant to replace regulation².

Anthropic published the first policy under that name, version 1.0, effective 19 September 2023¹. It sorts safeguards into "AI Safety Levels" (ASLs), with higher levels requiring tougher security and misuse protections³. In May 2025 Anthropic switched on its ASL-3 protections when it launched Claude Opus 4, as a precaution, saying it had not yet determined that the model definitely crossed the threshold⁴.

Other labs use different names. Anthropic says OpenAI and Google DeepMind adopted broadly similar frameworks within months of its first RSP³. At the AI Seoul Summit in 2024, 16 organisations, including Amazon, Google, Meta, Microsoft and OpenAI, agreed to publish safety frameworks setting out thresholds at which risks would be intolerable⁵.

These policies change often. Anthropic rewrote its RSP as version 3.0 in February 2026, separating what it will do alone from what it recommends for the whole industry, and adding regular public Risk Reports³. Version 3.4 took effect on 8 July 2026¹.

Example

When Anthropic released Claude Opus 4 in May 2025, its RSP triggered extra protections aimed at chemical, biological, radiological and nuclear weapons questions, plus tighter security around the model's weights⁴.

Why it matters

RSPs are the main way frontier AI companies publicly promise to handle catastrophic risks. Because they are mostly voluntary and the companies rewrite them, it is worth checking what the current version actually commits to.

Questions people ask

Is a responsible scaling policy legally binding?

Mostly no. RSPs are voluntary company policies³. Anthropic notes that some laws, such as California's SB 53 and the EU AI Act's codes of practice, now require frontier developers to publish risk frameworks, which pushes in the same direction³.

What are AI Safety Levels (ASLs)?

They are Anthropic's tiers of safeguards. ASL-2 is the baseline; ASL-3 adds tighter protection of model weights and targeted measures against weapons misuse⁴. Higher levels are meant for more capable models³.

Do other AI companies have responsible scaling policies?

Many have similar documents under other names. Signatories of the 2024 Frontier AI Safety Commitments, including Google, Meta, Microsoft and OpenAI, agreed to publish safety frameworks with risk thresholds⁵.

Related terms

Sources

Back to the glossary