Why a clinical-message similarity score changed meaning when doctors saw the AI draft

A similarity score used to evaluate AI-written patient messages changed meaning depending on how doctors worked. In a UCLA study, BERTScore correlated with physician quality judgments when clinicians edited a supplied AI draft. The relationship disappeared when doctors wrote their replies without seeing the draft. Shared wording was useful evidence of editing fidelity, but an unreliable substitute for independent clinical evaluation.[1]

Researchers examined a staged rollout in which 30 primary-care physicians handled real portal messages. In one condition, GPT-generated drafts were displayed for possible editing; in another, drafts were quietly generated but hidden from clinicians, who wrote replies without seeing them. The two groups comprised 21,073 displayed-draft message opportunities and 45,127 non-displayed ones, respectively, a total of 66,200. The dates of the underlying clinical workflow were September 2023 to August 2024; the analysis was published in October 2026.

A score that changes with the workflow

The team examined automated metrics including BLEU, ROUGE, BERTScore and readability scores, then compared their behaviour with blinded physician judgments about effectiveness, misunderstanding and potential harm. These are not interchangeable outcomes. BLEU and ROUGE pay attention to overlap between sequences of words; BERTScore examines similarities between contextual representations. A reply can resemble another without being the right clinical instruction.[1]

The most revealing contrast came from what clinicians actually did. In the displayed-draft group, 2,264 drafts, roughly 11% of the 21,073 opportunities, were incorporated into final messages with high semantic alignment. Those incorporated cases achieved a reported BERTScore F1 of 0.93 against the final text. In the condition where clinicians never saw the AI reply, physician-composed messages nonetheless had moderate similarity to the hidden drafts, BERTScore F1 of 0.83.

BERTScore measures semantic similarity, so 0.83 is not an 83% clinical-accuracy rate. The 11% figure also has a specific meaning: incorporation with high alignment. It does not count every occasion on which a doctor might have borrowed an idea while rewriting the text.

Editing fidelity is not clinical quality

In the condition where drafts were visible and sometimes incorporated, BERTScore correlated with clinicians' quality evaluations. When drafts were hidden, that relationship did not hold. In other words, a metric useful for studying how closely a professional editor stayed to a supplied draft did not automatically grade an AI's independent performance against a human message.[1]

Physician reviewers preferred the independently written human replies, with a mean preference score of 4.24 on a five-point scale running from strong preference for the AI draft to strong preference for the doctor’s final message. This is a comparative preference rating, not an absolute clinical-quality score. The staged rollout was observational and did not measure a causal effect on patient outcomes.

Consider two replies that say a symptom is worth checking. They might have nearly identical vocabularies, yet one advises urgent assessment today while the other suggests waiting a week. A semantic model may score the texts as close because most of the language overlaps. Clinicians care about the difference in action. This is why clinical appropriateness often must be checked against the relevant patient context, not just another sentence.

The important question is what the benchmark is for

A drafting tool raises several evaluation questions. Does it give the doctor a useful starting point? Does its unedited content give appropriate clinical advice? Does introducing it change response times or patient outcomes? Similarity to a signed message can inform the first question, while the others need direct measures of content and workflow.[1]

Developers should specify which use case they are testing before choosing a metric. An evaluation of editing efficiency might validly measure the distance between a suggested draft and a signed message, alongside time saved. An evaluation of autonomous patient communication needs expert judgments of clinical content, risks, omissions and patient understanding. Those may require different datasets and thresholds.

The practical lesson is to validate the metric in the workflow where it will be used. A system can produce language close to a doctor’s response while differing on the instruction that matters most. Clinical evaluation needs to inspect that instruction and its patient context.

Report an error or suggest a correction