arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

AI judges evaluate other AI systems by applying written clinical criteria | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Artificial IntelligenceEnglishPublished

AI judges evaluate other AI systems by applying written clinical criteria

November 17, 2026arXiv: 2611.18297v1

A new kind of artificial‑intelligence “judge” grades the output of another AI model by checking it against user‑specified rules. The judge examines a list of clinical criteria and returns an overall assessment that doctors or other systems can use. The judge is needed when a model’s answer must be interpreted against defined standards, such as whether a doctor’s note correctly records a medication that was stopped because of side effects.

When researchers report on an AI judge they should first provide the judge’s scope and technical setup, the precise criteria it uses, and how it compares to human clinicians. This includes details such as the model version, the prompt, the rubric, example answers, how the input is prepared, decoding parameters, and the method for combining scores. For multi‑step clinical encounters the report must state whether the judge saw earlier context, whether it reviewed tool calls in addition to the transcript, and at which step an error was introduced. If the judge is a panel, the report should list each judge’s model and role, how often judges disagree, and whether they make the same errors, all measured against independent clinician labels.

To evaluate the judge against clinicians a matched‑baseline design with blind assignment and adjudication is required. The study design should be set in advance, with original clinician ratings retained. For each criterion, reviewers must decide whether a clinician disagreement signals an error or a defensible difference, and how the reference label is chosen. Performance should be reported by criterion, separately for cases where clinicians agreed and where they differed, and with a full two‑by‑two table for yes‑or‑no items.

Despite these safeguards, AI judges can show “omission blindness”: they may fail to notice facts that have been removed from a clinical note. Across eight different judge designs, LLM judges were only slightly better than chance at distinguishing an intact note from one missing a fact, although they were good at detecting added or altered content. If a deployment continues to use the prompt that gave the highest scores or retrains a scribe on the judge’s verdicts, the omission may persist or become more common. Therefore, any clinical AI judge must meet pre‑specified thresholds for detecting clinically important errors before it can be trusted in practice.

Source paper

When evaluation becomes intervention in clinical AI

Artificial IntelligenceNatural Language Processingcs.CY
View on arXiv