Responsible Scaling
Anthropic's framework for managing the risks of increasingly powerful AI systems, which informed the company's refusal to remove all guardrails for military use.
Responsible Scaling is Anthropic’s framework for managing the deployment of increasingly capable AI models. The policy establishes escalating safety requirements as model capabilities increase, including evaluations, safeguards, and oversight mechanisms.
Relevance to the Dispute
Anthropic’s Responsible Scaling Policy provided the intellectual and institutional framework for the company’s position in the Pentagon dispute. The policy holds that certain AI applications pose unacceptable risks regardless of who is requesting them — a principle that directly conflicted with the Pentagon’s “all lawful purposes” demand.
Core Principles
The Responsible Scaling framework is built on the idea that:
- More capable models require more rigorous safety measures
- Some applications cross red lines regardless of the requester’s authority
- Safety evaluations should be conducted before deployment, not after
- The company, not the customer, bears ultimate responsibility for how its technology is used
Tension with Military Use
The military’s argument — that Pentagon oversight should replace company safeguards — directly challenged the Responsible Scaling framework’s assumption that the model developer retains responsibility for deployment outcomes.