Self-Harm and AI

2025-THR-01

Incident page

Back to Incidents

Moore, Grabb, Agnew, Klyman, Chancellor, Ong, and Haber (2025) evaluated therapy-oriented LLM/chatbot systems and reported failures to recognize indirect or escalating suicidal intent, along with some responses that provided potentially harmful situational information or reinforced false-belief frames.

Parent
Incidents
Provider
No matched provider page
Last reviewed
March 25, 2026
Incident ID
2025-THR-01
Provider
No matched provider page
Outcome
UNSAFE_OUTPUT
Grade
A1
Date window
unknown to unknown

Contents

Summary

2025-THR-01 is a A1 benchmark study row involving Therapy chatbots (multi-app evaluation). Moore, Grabb, Agnew, Klyman, Chancellor, Ong, and Haber (2025) evaluated therapy-oriented LLM/chatbot systems and reported failures to recognize indirect or escalating suicidal intent, along with some responses that provided potentially harmful situational information or reinforced false-belief frames.

Therapy Chatbot Red-Team Evaluation (FAccT 2025). Original date string: June 2025 (presentation). Source summary: Moore, Grabb, Agnew, Klyman, Chancellor, Ong, and Haber (2025) evaluated therapy-oriented LLM/chatbot systems and reported failures to recognize indirect or escalating suicidal intent, along with some responses that provided potentially harmful situational information or reinforced false-belief frames.

Coding

2025-THR-01 is coded as UNSAFE_OUTPUT with grade A1. The public row assigns DET-FN, GEN-M, RTI-CO and treats the incident as evaluation or demonstration.

Field Value
Platform Therapy chatbots (multi-app evaluation)
Model version therapy chatbot evaluation set
User type test
Jurisdiction unknown
Evidence type benchmark_study
Product mechanism benchmark_or_evaluation_set
Trajectory strict no

Assigned taxonomy codes

DET-FNGEN-MRTI-CO

Evidence

The public evidence anchors for 2025-THR-01 are the materials listed below. Only explicitly linked evidence files are hosted live.

The closest related pages are the incidents that share a provider family, case cluster, pathway-duplicate group, or overlapping coding profile.

Sources

This page is grounded in the coded incident registry, the analysis companion, the coding rules, and the public evidence links listed above.

Source filesIncident registry/Analysis view/Coding rules

Access publication

Safety note

This publication preserves a harm-minimization posture and does not provide self-harm instructions. If you or someone you know is in immediate danger, call your local emergency number. In the U.S. and Canada, call or text 988.