Anthropic shares 10 recommendations for building better AI safety evaluations
AI safety company Anthropic published 10 recommendations for building better AI safety evaluations on November 13, 2026, borrowing lessons from psychology's replication crisis to improve how safety claims are tested.
When an AI company claims its new model is safer than the previous one, that assertion rests on the quality of its tests. Anthropic now wants the industry to scrutinize how these safety evaluations get constructed, recognizing that a score is only meaningful if the test actually measures what it claims to measure.
The company published its engineering blog post titled "Building Strong Safety Evaluations" on November 13, 2026, outlining recommendations for designing, running, and reporting the evaluations that determine whether a model is ready for release. These evaluations shape critical decisions about model deployment, safeguards implementation, and future safety research focus.
A safety claim is only as sound as the eval behind it.
Learning from psychology
Anthropic draws parallels to psychology's replication crisis, when the field confronted that many findings couldn't be replicated. The company proposes that AI safety adopt methodological reforms before facing its own reckoning. One key recommendation is preregistration, where researchers document their analysis methods before seeing data to prevent cherry-picking favorable results.
The company also urges validating automated grading systems against human expert judgments, noting that many evaluations now use AI models to score other AI models. These graders need verification from people with actual subject matter expertise, according to Anthropic.
Beyond simple refusal rates
A notable recommendation addresses how labs measure model responses to sensitive requests. Anthropic calls for moving beyond simple single-turn refusal rates toward multi-turn simulations that better reflect real conversations where context builds and requests evolve.
The company stresses that appropriate responses often depend on nuance rather than blanket refusal, particularly in areas like self-harm and cybersecurity. Anthropic also recommends assembling multidisciplinary teams for evaluation development and implementing real-time monitoring during tests to catch problems as they occur.
Industry impact
The recommendations arrive as Anthropic continues to publish more about its safety methods, including October 2026 reports on unintended model actions and containment practices. While the company frames the advice as useful across the industry, these practices require significant time, expertise and resources, with no mechanism compelling competitors to adopt them.
Anthropic also calls for transparent communication about uncertainty and limitations in evaluation processes, recognizing that no test is perfect but honesty about constraints can improve overall safety practices in the rapidly evolving AI landscape.
Related
