AI Aulis

AI NEWS IN ENGLISH · 14 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 14 November 2026Paper

Anthropic explains how to make AI safety evaluations reliable

AI-generated contentWritten by AI

anthropic.com, claude.com, x.com · 5 sources · Sources published: 13–14 November 2026 · Some source publication dates are unknown

Anthropic says the refusal rate alone often measures the wrong thing. Drawing on examples from its own system cards, the company advises checking the grader against human judgements and reporting the uncertainty in results.

On 13 November, Anthropic published guidance on its Engineering Blog for designing, running and reporting safety evaluations of AI models. The company said the claim ‘safer than its predecessor’ means a model scored better on a test someone designed. The claim is only as valid as the test. Anthropic takes psychology as its model. The field reformed its practices after large replication studies, in which many published results failed to replicate.

Refusing is not always the right answer

The most obvious metric is how often a model refuses to answer. In conversations about suicide and self-harm, however, the refusal rate is often the wrong metric, Anthropic said. When a user is in crisis, simply declining to engage is itself a failure: a good response engages the user with empathy and points them towards human support. On dual-use topics such as biology and cybersecurity, a good response helps with a legitimate request but leaves out harmful instructions.

For multi-turn conversations, Anthropic reports the appropriate response rate. This is the share of conversations in which the model followed an expert-written rubric from start to finish. A single poor response fails the whole conversation. After a fixed opening turn, another Claude model writes the user’s turns from a description written by an expert, so a request can turn harmful gradually. Simulated users are still much simpler than real ones, the company said.

Single-turn tests pair harmful requests with benign requests on the same topics, so that unnecessary refusals also show up. These tests are already too easy: recent models achieve harmless response rates of between 94 and 98 per cent on Anthropic’s API. The company uses them mainly to check that known failures do not return.

Check the grader against humans

Another language model usually scores the responses, because nobody has time to read hundreds of thousands of conversations. Anthropic advises comparing the grader’s judgements with human labels and accounting for chance agreement. The grader should agree with experts about as often as experts agree with each other. Grading criteria are best broken down into yes-or-no questions. It is also worth telling the grader that warnings do not make harmful content harmless. The grader should be a different model from the one under test, because models favour their own outputs.

Uncertainty and settings change the interpretation

The analysis plan should be written down before anyone sees the results. Anthropic treats 100 items as the minimum for a multi-turn evaluation and 1,000 for a single-turn one.

Confidence intervals for multi-turn evaluations run on the API can exceed ±10 percentage points. In a romance scam evaluation run on claude.ai, Claude Opus 5 had an appropriate response rate of 98 per cent and Opus 5.5 94 per cent. The difference was not statistically significant, so the company could not conclude that either model was better.

Settings matter too. For Claude Fable 5.1, the appropriate response rate in multi-turn evaluations differed by as much as 34 percentage points depending on whether the test ran on the bare API or on claude.ai with a system prompt. In the suicide and self-harm evaluation, the rate was 60 per cent on the API and 94 per cent on claude.ai, where new instructions had been added to the system prompt.

The company points to two ready-made tools: the UK AI Security Institute’s open-source Inspect framework and its own Petri tool, which automates multi-turn behavioural testing.

Report an error in this story