Anthropic's Responsible Scaling Policy uses an if-then structure: if models pass a certain capability, Anthropic imposes a set of safety and security requirements.

Judge as of 2024-11-11. True: Anthropic's Responsible Scaling Policy as published by then sets capability thresholds that, once reached or approached, trigger specified safety and security requirements (AI Safety Levels / required safeguards). Mostly true: this structure is present, with a minor difference in detail, for example thresholds triggered by evaluations indicating the capability might be reached rather than definitively passed. Mixed: the policy only partly has this structure. Mostly false: the policy refers to capabilities and safeguards, but they are not linked conditionally. False: the RSP has no such structure.

Set before the research began.

1. E1 Our updated Responsible Scaling Policy Anthropic · 15 October 2024 Own record

   “Capability Thresholds: Specific AI abilities that, if reached, would require stronger safeguards than our current baseline.”

2. E2 Our updated Responsible Scaling Policy Anthropic · 15 October 2024 Own record

   “Required Safeguards: The specific ASL Standards needed to mitigate risks once a Capability Threshold has been reached.”

3. E3 Our updated Responsible Scaling Policy Anthropic · 15 October 2024 Own record

   “Chemical, Biological, Radiological, and Nuclear (CBRN) weapons: If a model can meaningfully assist someone with a basic technical background in creating or deploying CBRN weapons, we require enhanced security and deployment safeguards (ASL-3 standards).”

4. E4 Our updated Responsible Scaling Policy Anthropic · 15 October 2024 Own record

   “Autonomous AI Research and Development: If a model can independently conduct complex AI research tasks typically requiring human expertise—potentially significantly accelerating AI development in an unpredictable way—we require elevated security standards (potentially ASL-4 or higher standards) and additional safety assurances”

Anthropic's official announcement of its updated Responsible Scaling Policy, dated 2024-10-15 (before the claim date), defines Capability Thresholds as specific abilities that, if reached, require stronger safeguards (E1). It defines Required Safeguards as the ASL Standards needed once a threshold has been reached (E2). The announcement also gives concrete conditional rules for CBRN weapons, which require ASL-3 safeguards (E3), and for autonomous AI research and development, which require elevated security standards (E4). Together these show the if-then structure the claim describes. This rests only on Anthropic's own published policy announcement, which here counts as the official record of the policy itself.