CrisisBench offers privacy-preserving tests for crisis counseling AI
Researchers introduce an evaluation suite built from de-identified hotline conversations, testing how well large language models handle risk assessment and de-escalation against expert counselors.
In plain words
Crisis helplines are understaffed, and AI models might help counselors, but real hotline transcripts are hard to use because of privacy rules. CrisisBench is a test suite built from de-identified conversations that scores how well AI systems handle crisis conversations. In the authors' tests, the systems they checked did not match human counselors on some response qualities.
What was announced
The suite, described in a paper by the CrisisBench authors on arXiv, contains 1,085 single-turn tasks drawn from 200 de-identified conversations, plus 1,185 simulated patient models built from hotline metadata. A 23-item judge for scoring responses was calibrated against expert counselors, and the patient models were calibrated the same way. The authors say this design allows evaluation without exposing real caller transcripts.
In the reported results, five GPT models along with draft-then-edit and retrieval-augmented generation approaches all fell short of human counselors on excessive empathy and questioning in single-turn evaluation. In multi-turn evaluation, retrieval-augmented generation scored highest at 0.961, with all systems within 0.058 of that mark. Microskills and style or tone scores were near ceiling, while productivity and safety or risk drove most differences between systems. Counselors rated the simulated patient models as realistic in 86% of cases and useful in 92%. The team has released the rubrics, judge prompts, and evaluation tasks.
What it does not show
- The evidence is the paper's abstract only, so methods, caveats, and full results cannot be verified from the supplied passages.
- All reported scores and counselor ratings come from the authors themselves; no independent confirmation is available.
- The abstract does not state which specific failure modes persist in multi-turn settings or how the single-turn gaps affect real helpline use.
For security and IT teams
- Scope
- CrisisBench covers single-turn and multi-turn crisis counseling evaluation and includes released rubrics, judge prompts, and evaluation tasks, per the paper's abstract.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.