1601 - Synthetic Bias? Analysis of Stigmatizing and Narrative Language in Human- vs AI-generated medical documentation

Thursday, February 19, 2026

5:00 PM - 6:30 PM AST

Presenting Author(s)

MZ

Michael D. Zhitnitsky, DO

Resident
Department of Physical Medicine and Rehabilitation, Mayo Clinic, Rochester, MN
Rochester, Minnesota, United States

Co-Author(s)

DV

Daniel Veith, MS

Senior Research Technologist
Rehabilitation Medicine Research Center, Mayo Clinic, Rochester, MN
Rochester, Minnesota, United States
GS

Geoffrey Smith, MD

Physiatrist
Mayo Clinic
Rochester, Minnesota, United States

Corresponding Author(s)

RS

Ryan Solinsky, MD

Spinal Cord Injury Medicine Physician-Scientist
Mayo Clinic
Rochester, Minnesota, United States

Submitter(s)

MZ

Michael D. Zhitnitsky, DO

Resident
Department of Physical Medicine and Rehabilitation, Mayo Clinic, Rochester, MN
Rochester, Minnesota, United States

Objectives: Compare the prevalence and types of stigmatizing language present in Rehabilitation Medicine documentation with those generated by AI models. Compare the prevalence and types of narrative medicine themes present in Rehabilitation Medicine documentation with those generated by AI models.

Design: Researchers reviewed 65 de-identified history and physical notes from Mayo Clinic's inpatient rehabilitation unit from 2013-2023. Patient demographics, diagnoses, and word count were used to prompt Gemini 1.5 Pro to create similar documentation. Both human- and AI-generated notes were analyzed for the presence of stigmatizing language and narrative medicine themes. Independent reviews assessed each note for clinical realism and suitability. To reduce bias, a maximum of three notes per clinician were included and incomplete AI outputs were replaced.

Results: Human-written notes demonstrated a significantly lower frequency of stigmatizing language compared to Gemini-generated notes (1.8 vs 2.8 terms per note; p=0.003). When comparing narrative medicine themes, human-generated notes contained substantially less quest narrative (0.0 vs 3.4 terms per note; p< 0.001) and restitution narrative (0.05 vs 2.8; p< 0.001) terms compared to AI-generated notes. There was no significant difference in overall length of notes.

Conclusions: Our results suggest that AI, trained on potentially antiquated datasets, can produce more problematic language in medical documentation. This highlights the need for appropriate prompts and explicit constraints to ensure respectful, patient-centered communication. When used within secure, ethical frameworks, these large language models hold promise for augmenting clinical documentation, though further investigation is needed to determine how linguistic differences may impact patient outcomes and perceptions in real-world clinical settings.