Ideas worth a closer look
VIVA PARADOX.Independent stories, clearly told.
AI News & Insights

Anthropic Guidance Urges Harm-Focused AI Safety Evaluations

Conceptual editorial artwork: Anthropic Guidance Urges Harm-Focused AI Safety Evaluations

What did Anthropic publish?

Anthropic published guidance on November 13, 2026, for building more reliable AI safety evaluations. The company frames these evaluations as tests of a model’s propensity to cause or contribute to harm: whether a response makes harm more or less likely, and how severe that harm could be. That is different from a capability test, which asks whether a model can perform a specific task.

Why does the metric matter?

Analysis: Anthropic’s definition narrows what a safety evaluation should answer. A high or low capability score would not, by itself, show whether a model’s reply increases risk in a harmful situation. The guidance instead points evaluators toward the effect of the response and the severity of possible harm, which can change the meaning of a test result even when the same model capability is involved.

Where can refusal rates mislead?

Anthropic says refusal rate can be the wrong construct for conversations about suicide and self-harm. In that setting, the company says simply refusing to engage can itself be a failure. The desired response, according to the guidance, engages with empathy and points the user toward human support. Analysis: That means a refusal-only metric could count an answer as safe while missing whether the model responded appropriately to a user in crisis.

How should longer and riskier tests be run?

For multi-turn safety evaluations, Anthropic says it reports an appropriate-response rate across the whole conversation, using rubrics specific to each risk area. Under that approach, one bad turn fails the entire exchange. For cyber evaluations, Anthropic recommends sandboxes without internet access and real-time monitoring, so evaluators can stop a run if a model’s actions might cause real-world harm.

Official announcement · Additional coverage

Official announcement published: November 13, 2026 at 00:00 UTC.