JMIR Publications

ABSTRACT

Background:

Large language models (LLMs) are increasingly used in academic writing, often for grammar refinement. However, LLMs can also generate linguistically polished yet factually incorrect content, which may distort evidence appraisal and, if translated to practice, compromise patient care

Objective:

To estimate the likelihood that physicians (i) misperceive false information embedded in LLM-generated abstracts as true and (ii) report intent to use that information in future practice; and to describe cues physicians use when judging credibility.

Methods:

We conducted a cross-sectional, vignette-based survey of physicians working in acute care (emergency medicine and trauma surgery) across four hospitals in Germany. Each participant received a questionnaire containing four abstracts produced by an LLM (GPT-5 “Thinking” mode): some intentionally included false claims framed with plausible, logical explanations; others were accurate. To minimize expectancy effects, participants were told that the study concerned the influence of grammar on the perceived truth of manuscripts. Primary outcomes were the proportion endorsing (a) false claims as likely true and (b) intent to use those claims clinically. Secondary outcomes included the proportion rejecting accurate claims and ratings of grammatical/linguistic quality. Data were summarized using descriptive statistics. Any optional free-text justifications were subjected to directed content analysis to identify credibility heuristics (e.g., grammar/fluency, plausibility, citations).

Results:

Twenty physicians provided 80 ratings of four abstracts (two true, two false). Overall, they indicated they would act on 54/80 (67.5%) abstracts. Using the true status of the abstract as reference, the decision to act showed a sensitivity of 77.5% and a specificity of 42.5%, with a positive predictive value of 57.4% and a negative predictive value of 65.4%. Perceived accuracy and precision were only slightly higher for true than for false abstracts, with no statistically significant differences. In contrast, abstracts on which physicians would act were rated markedly higher for accuracy, precision and structure than those on which they would not act, while grammatical ratings were similar.

Conclusions:

In this sample, a substantial proportion of physicians accepted—and some intended to act upon—false yet well-written LLM-generated claims. Linguistic polish appears to function as a credibility heuristic. Journals and institutions should restrict LLM use to language editing, prohibit generative use for creating or embellishing scientific claims, and encourage transparent authorship disclosures. Clinicians should be trained to critically appraise content beyond surface form to mitigate the risk of practice decisions based on fabricated or inaccurate information