Title: Examining Stigmatizing Terminology in Clinical Records: A Comparative Study of Ambient AI Drafts Versus Clinician-Edited Notes

Abstract: While ambient artificial intelligence (AI) documentation solutions are being widely implemented to alleviate the administrative load on healthcare providers, the impact of these tools on the use of biased language within clinical records has not been fully understood. To address this gap, we performed a large-scale comparative analysis to quantify how stigmatizing language shifts during the editing process from AI-generated drafts to final clinician-approved notes. By employing a lexicon-based natural language processing (NLP) pipeline, our study evaluated three key metrics: the incidence of stigmatizing terms in initial AI drafts, the prevalence and specific composition of such terms in finalized documentation, and the rate at which clinicians removed or added stigmatizing vocabulary. Analyzing 66,297 paired note sections, we found that 21.4% of AI drafts contained at least one instance of stigmatizing language. This figure increased to 24.0% in the corresponding clinician finalized notes. Notably, the introduction of stigmatizing terms occurred more frequently than their removal. These findings suggest that clinician editing of ambient AI drafts may inadvertently serve as a net source of stigmatizing language entering the Electronic Health Record (EHR).

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC