Preprint Finds Incidental Speech Can Contaminate AI-Generated Medical Notes
Published in Artificial Intelligence, Journals, Tools.
October 7, 2026
A preprint submitted to arXiv on October 6 examines how incidental conversation and background speech affect language models used for clinical documentation and reasoning.
In 576 patient-clinician dialogues, the researchers found that frontier models included small-talk exchanges in 35% of generated notes. In 3.7% of notes, models misattributed these asides or used them clinically. Average quality scores changed by no more than 0.20 points on five-point scales, illustrating that aggregate scores can obscure specific contamination errors.
In a separate set of 57 simulated recorded consultations, speech from another patient encounter, introduced at −10 dB, appeared in 48.2% of transcripts. Contamination was detected in 5.3% of downstream notes produced by four open-weight models.
These are experimental findings from a preprint, not established real-world error rates or evidence of patient harm. The work supports testing resistance to irrelevant speech and cross-encounter contamination before deploying documentation systems in clinical workflows.
Source: Vishwanath and colleagues, arXiv preprint submitted October 6, 2026.