The study investigated how incidental, non‑clinical information affects large language models (LLMs) when they generate ambient clinical notes or perform reasoning tasks. Researchers first analyzed 576 real‑world patient‑clinician dialogues, prompting frontier‑scale LLMs to produce visit summaries. They observed that the models inserted unrelated small‑talk into 35 % of the generated notes, yet overall note quality, measured on a five‑point scale, shifted by no more than 0.20 points on average. In a smaller subset—3.7 % of the frontier model outputs—the incidental remarks were either misattributed to the patient or mistakenly used for clinical decision‑making. To test auditory leakage, the team conducted 57 mock consultations where background speech from a separate encounter was played at –10 dB relative to the primary conversation. Speech‑to‑text transcription of these recordings showed that the background utterances appeared in 48.2 % of the transcripts, and downstream notes produced by four open‑weight models contained detectable contamination in 5.3 % of cases. The authors propose a dual‑encoding hypothesis, suggesting that the same LLM components vulnerable to distraction also underlie clinical reasoning, and they recommend evaluating robustness to incidental information before deploying LLMs in clinical settings, alongside safeguards that block contamination while preserving diagnostic performance.