# ambient AI scribes

This study constructed a multilingual clinical dictation corpus (five clinical dictation scripts spanning a complexity gradient, translated into 99 languages, rendered to synthetic speech under three acoustic conditions, and transcribed by a production ambient scribe), computed six frequency metrics, and had three independent large language model raters assess clinically meaningful error patterns using a Severity x Likelihood framework; across 59,819 genuine transcription-error occurrences, 58,329 (97.5%) were LOW risk and 251 (0.42%) CRITICAL or HIGH, none of the six frequency metrics showed a detectable association with serious clinical risk (absolute Spearman rho < 0.16), a Severity x Likelihood sum remained strongly correlated with WER (rho=0.

This study constructed a multilingual clinical dictation corpus (five clinical dictation scripts spanning a complexity gradient, translated into 99 languages, rendered to synthetic speech under three acoustic conditions, and transcribed by a production ambient scribe), computed six frequency metrics, and had three independent large language model raters assess clinically meaningful error patterns using a Severity x Likelihood framework; across 59,819 genuine transcription-error occurrences, 58,329 (97.5%) were LOW risk and 251 (0.42%) CRITICAL or HIGH, none of the six frequency metrics showed a detectable association with serious clinical risk (absolute Spearman rho < 0.16), a Severity x Likelihood sum remained strongly correlated with WER (rho=0.

This paper identifies and names the phenomenon of "inference impersonation"—ambient AI scribes generate rather than transcribe clinical reasoning in the Assessment and Plan, producing text indistinguishable from transcription in the final note, so trainees may edit AI drafts instead of reasoning independently and risk never developing the judgment training exists to build; it proposes vendor-side learner-specific configurations and section-level transparency plus training-program responses such as reasoning-before-note, competency gating, and oral assessment.

This paper identifies and names the phenomenon of "inference impersonation"—ambient AI scribes generate rather than transcribe clinical reasoning in the Assessment and Plan, producing text indistinguishable from transcription in the final note, so trainees may edit AI drafts instead of reasoning independently and risk never developing the judgment training exists to build; it proposes vendor-side learner-specific configurations and section-level transparency plus training-program responses such as reasoning-before-note, competency gating, and oral assessment.