How it works
The Responsible Scaling Policy (RSP) is Anthropic's policy for scaling AI models safely. It defines capability levels (ASL, AI Safety Levels) and requires that, before training or releasing a more capable model, the company implements safety measures proportional to the identified risk.
The idea is that higher capabilities require stronger safeguards. The policy includes safety evaluations, red teaming tests and pause criteria: if a model reaches a dangerous capability level without adequate safeguards, training or release is halted.
Practical example
Before training a model with advanced coding or biology capabilities, Anthropic assesses whether the model could be used to cause harm (for example, assisting in the creation of biological weapons). If the risk is high, the company implements filters, monitoring and access restrictions before proceeding. Results and decisions are documented in system cards.
What it's for
The RSP serves as a reference for AI governance in companies that develop frontier models. It inspires other industry policies and helps customers and regulators understand how the company manages risks from more capable models. It is an example of a voluntary commitment to safety, complementing (not replacing) regulations such as the EU AI Act.