cs.AINov 13, 2026

When evaluation becomes intervention in clinical AI

Authors: Anthony Abou Mrad, Aviona Conti, Samuel Schmidgall, Jayanth Pratap, Ruhika Iyer, Joseph Kim, King Shi, Yann Sakref, 

Organizations: The Johns Hopkins University, Baltimore, Maryland, USA · Rutgers New Jersey Medical School, Newark, New Jersey, USA · Stanford University, Stanford, California, USA

Abstract

AI judges, models that grade another system's output against a rubric, extend clinical review to the scale at which clinical AI writes. When teams use judge assessments to select a prompt or train a model, the judge becomes part of the intervention. Selection can then favor errors the judge overlooks. We describe what a judge's report must contain and the independent clinical test each judge-driven change needs.

Explore similar work

Oct 20, 2026

Decision language models as affordable and rapid alternatives to LLM judges of clinical generative AI

Automated judges of clinical artificial intelligence (AI) outputs are necessary, but large language model (LLM) judges are costly and slow at scale. We compared Jev, a decision language model that returns structured judgments, with frontier and open-weight LLM judges in four use cases. On 118 pediatric answers, Jev's agreement with physicians' ordinal quality ratings (Gwet's AC2 0.50; expected score 0.53) matched or exceeded that of seven of 12 LLM judges, including frontier models, Claude Opus 5.5 (0.49) and GPT-6 (0.38). On physician-graded binary HealthBench criteria, Jev matched or exceeded four of six LLM judges. On MedHELM's open-ended tasks, Jev's agreement with a three-LLM jury approached or exceeded the jurors' internal agreement. On HealthBench Professional, Jev matched the benchmark's reference judge, GPT-5.4, on 91% of criteria (other judges, 91-93%). Frontier judges cost 73-294 times more than Jev and took 7-57 times longer. Jev may enable efficient clinical AI evaluation at scale.
Jul 21, 2026

Evaluating medical AI under missing information: same-provider judges and human raters change apparent safety

Readiness stress-testing of medical AI has focused on closed-ended and multimodal benchmarks. We extend it to open-ended clinical conversation under missing information, where safe behavior means recognizing absent information and qualifying, clarifying, or not over-committing - and where the evaluator becomes part of the measurement. We stress-test four models - three flagships (Claude Opus 4.8, GPT-5.5, Grok 4.3) and one mid-tier model (Gemini 3.5 Flash) - by deleting the latter half of the final user turn in HealthBench conversations, grading responses with a four-provider LLM-judge panel and a blinded clinician-anchored reference. Two evaluator-facing results are robust. First, judge choice materially changes apparent safety: inter-judge agreement is only moderate (Fleiss' kappa = 0.65), and after adjusting for each judge's general leniency (vote-level logistic regression), a positive same-provider association remains (exact permutation p = 0.04; GPT-5.5 ~ +0.10 on the probability scale) - large enough to change which model appears to over-commit least once its own-provider judge is excluded. Second, LLM judges are more permissive than clinicians on a blinded 50-item subsample: all four are significantly more lenient than the stricter independent clinician (crediting appropriate uncertainty on 66-84% of items vs 52%), and three of four than the author-influenced consensus (Grok directional only; judge-vs-consensus kappa = 0.20-0.43). On the author-audited clinical-underdetermined subset the permissiveness gap widened and the point-estimate model ordering held. A closed-ended MedQA anchor confirms accuracy is high and option-order effects are within a +/-5-point equivalence region for three of four models, so the safety gap is about calibration, not knowledge. We release the harness, prompts, per-item outputs, judge panel, perturbation audit, and human-annotation protocol.
Nov 14, 2026

Objectify, Don't Delete: Complaint-Derived Evaluation of Ambient Clinical Notes

Evaluating the quality of AI-generated clinical notes is hard: standard metrics like ROUGE measure lexical similarity to reference notes, while expert checklists encode what committees believe matters. Neither captures what clinicians actually complain about. Recent work has shown that evaluation criteria can be mined from user feedback and filtered to criteria that LLM judges can reliably enforce. We argue that this is where much of the value is lost. The criteria that fail enforceability are disproportionately the ones clinicians raise most often, because high-frequency complaints are stated as several questions glued together. No yes/no judgment can grade such composites as posed. We introduce an objectification step that rewrites such criteria into sets of scoped, determinate decisions. We demonstrate the approach through an evaluation pipeline that mines criteria from complaints, objectifies them, and deploys LLM judges against both planted defects and production notes. Across eleven defects planted into ACI-bench reference notes, objectified judges localize violations with median recall of 0.95, while reference metrics change by at most 0.037. Given the complaint as one broad question, the same judges still find every planted defect, but they also flag 70% of the unperturbed reference notes; one section-pair question per call raises Youden's J from 0.30 to 0.72 (by 0.19 to 0.26 with prompt effort matched). We release ACI-Defect, a perturbation benchmark; clinician-reviewed instance-level labels showing quality gaps in the ACI-bench gold standard (32 of 40 notes contain hallucinations; 40 of 40 omit discussed facts); and grounding-judge verdicts. On 1,446 production notes, mean instance-rate burden falls at every clinician-rating step (rho=-0.15), a small association. Defect counts have to be kept per span rather than collapsed into one pass/fail bit per note.