Skip to main content
Preprint Credible

Conversational trajectory degrades large language model detection of suicidal ideation relative to clinicians: a preregistered study

Preregistered study comparing 49 large language models against 8 clinicians on detecting suicidal ideation embedded in psychotherapy transcripts of increasing length (0-200 speaker turns). Model F1 declined with conversational depth across all model families while clinician performance remained stable; conversational content, not length alone, explained the degradation.

Publisher

medRxiv (Beth Israel Deaconess / Harvard digital-psychiatry group)

Published

14 Jul 2026

Added

2 weeks ago

Key Findings

  • F1 for suicidal-ideation detection declined with conversational depth across all 49 evaluated model families; clinicians showed no comparable decline
  • Larger models performed better overall but still degraded with depth; conversational content rather than raw length explained the changes
  • Restating instructions partially recovered model performance, and the authors conclude safety evaluations must test sustained performance across realistic extended interactions rather than isolated prompts

Methodology Notes

Preregistered design inserting validated clinical statements of suicidal ideation into therapy transcripts at varying depths (0-200 speaker turns); 49 LLMs and 8 clinicians evaluated head-to-head. Preprint, not yet peer-reviewed (medRxiv v1, 2026-07-14). medrxiv.org blocks automated fetchers; verified via the medRxiv API (title, authors, date, abstract).

Authors

Mark Kalinich, James Luccarelli, Joseph Santa Maria, Matthew Flathers, An Nguyen, Sung Hyun Song, Karim Makhoul, Maria Jose Rivera Criado, Caroline M. Ginapp, Benjamin Hill, Jackson N. Shumate, Hana Notsu, Colin Smith, Frazer Moss, John Torous

Tags

medrxivpreregisteredtrajectory-degradationtorous-groupclinician-comparison

Cite This

APA

Mark Kalinich et al. (2026). Conversational trajectory degrades large language model detection of suicidal ideation relative to clinicians: a preregistered study. medRxiv (Beth Israel Deaconess / Harvard digital-psychiatry group). https://www.medrxiv.org/content/10.64898/2026.07.10.26357132v1