Anthropic Releases Version 3.0 of Responsible Scaling Policy, Reflecting Two Years of AI Safety Learning
Key Takeaways
- ▸Anthropic has released version 3.0 of its Responsible Scaling Policy after more than two years of implementation, with updates focused on transparency, accountability, and lessons learned
- ▸The RSP framework uses conditional "if-then" commitments, where AI Safety Levels (ASLs) trigger stricter safeguards as models cross defined capability thresholds
- ▸AI capabilities have evolved dramatically since the original 2023 policy, from simple chat to autonomous web browsing, code execution, and multi-step actions
Summary
Anthropic has released the third version of its Responsible Scaling Policy (RSP), a voluntary framework designed to mitigate catastrophic risks from increasingly capable AI systems. The updated policy reflects more than two years of implementation experience and introduces new measures to increase transparency and accountability in the company's decision-making processes around AI safety.
The original RSP, introduced in September 2023, was built on the principle of "if-then" commitments, establishing AI Safety Levels (ASLs) that trigger increasingly stringent safeguards as models cross specific capability thresholds. Since then, AI capabilities have evolved dramatically—from simple chat interfaces to systems that can browse the web, write and execute code, use computers autonomously, and take multi-step actions. Each new capability has introduced corresponding new risks, validating the RSP's foundational premise.
In assessing their "theory of change," Anthropic acknowledges mixed results. The RSP successfully served as an internal forcing function, compelling the organization to treat critical safeguards as launch requirements and spurring faster progress on safety measures. The company also credits the policy with helping catalyze a "race to the top" among AI companies, with similar voluntary frameworks emerging across the industry. However, Anthropic recognizes that some aspects of their original vision—particularly around creating multilateral action at capability thresholds and coordinating with governments on safeguards beyond what one company can achieve alone—remain works in progress.
The version 3.0 update aims to reinforce successful elements of the framework while addressing identified shortcomings. The policy continues to anticipate that the most challenging safeguards at higher capability levels will require coordination with governments worldwide, reflecting Anthropic's recognition that no single company can adequately address the risks posed by frontier AI systems operating in isolation.
- The policy has successfully functioned as an internal forcing function at Anthropic and helped inspire similar voluntary frameworks across the AI industry
- Anthropic acknowledges that the most stringent future safeguards will likely require coordination with governments, as they exceed what any single company can achieve unilaterally
