Research·arXiv cs.CL·13h ago

LLM judges fail to catch omissions in AI-drafted clinical notes

A new benchmark reveals a critical blind spot in LLM-based clinical note auditing: AI judges reliably catch added or altered information but fail to detect omissions, the most common error type in ambient AI scribes. Across eight judge architectures, discrimination on missing facts ranged from 0.50 to 0.63, barely above random chance, while detection of false additions reached 0.79 to 0.94. This asymmetry exposes a fundamental limitation in using language models to validate other language models in high-stakes healthcare settings, where silent gaps in documentation carry direct patient safety implications.

Modelwire context

Explainer

The paper isolates a directional asymmetry in LLM judgment: models trained to catch errors excel at spotting false additions but perform near-random on missing information. This isn't just a performance gap; it reveals that detection difficulty correlates with what's present in the text, not what the judge 'knows' should be there.

This connects directly to the August audit of deployed AI scribes that found one in three notes contained errors, with dangerous gaps in allergy and medication records. That study documented the prevalence of omissions in production systems; this new benchmark explains why automated auditing (the natural response to that finding) will systematically miss them. The DIASENTINEL multi-agent work from the same week points toward one partial solution: hybrid architectures that enforce rule-based verification rather than relying on LLM judges alone. Together, these three papers sketch a problem (omissions are common), a diagnosis (LLM judges can't catch them), and an emerging workaround (deterministic guardrails).

If clinical validation teams deploy this benchmark against their own judge models in the next 6 months and report similar omission-blindness rates (0.50-0.63 range), that confirms the finding generalizes beyond the eight architectures tested. If they don't, or if they report substantially higher omission discrimination, the result may be specific to the benchmark's design rather than a fundamental LLM limitation.

Coverage we drew on

One note in three: a verified census of three deployed AI scribes, and the instrument that counted it · arXiv cs.CL

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsLLM judges · ambient AI scribes · clinical notes

Read full story at arXiv cs.CL →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes and What Recovers It”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.