Beyond Initial Crisis Recognition: Suicide-Risk Responses of ChatGPT, Claude, and Gemini Across 14-Turn Ukrainian-Language Conversations
Abstract
Nine deployed conversational systems mostly recognised the crisis, but it disappeared from their behaviour once the conversation moved on. We ran 16 scenarios in ChatGPT, Claude, and Gemini as 14-message conversations (48 dialogues): crisis signals alternated with distraction, turn 13 introduced a new person at risk, and turn 14 was sent a day later. Coding was binary, as in the previous study, with clinical verification; two crisis-contour specifications were tested separately. New crisis signals drew a reaction in 87, 95, and 76 of 96 responses, respectively; the crisis stopped being mentioned mainly at an everyday message. The new person at risk drew a reaction in 44 of 48 responses; after a day, 29 retained crisis content. All 14 turns drew a 'yes' code in 4, 7, and 0 of 16 conversations. Reacting to the new person did not guarantee correct allocation of roles and actions. Both contour specifications produced safety-relevant errors. Neither hypothesis was supported: the systems behaved differently, and the same rule, introduced as a single package, did not stabilise their behaviour. Recognition, retention, and reconfiguration at a transition are distinct properties; the weakest is retention of the prior state and tracking of each person's trajectory.
Related
- Benchmarking large language models on the Ukrainian Krok 2 licensing examination in physical therapy: accuracy, consistency, and agreement
- Case Mix and Protocol Constraints in Diagnostic Language Model Conversations: A Secondary Analysis of Released MedR-Bench Outputs.
- Cross-Cultural Adaptation and Initial Psychometric Evaluation of the Kazakh-Language Sino-Nasal Outcome Test-22 (SNOT-22) in Patients with Chronic Rhinosinusitis with Nasal Polyps
- Quality of Initial MRI Reports for Soft-Tissue Sarcomas and Their Association with Radiologic Recognition of Sarcoma.
- Safety and guideline concordance of public large language model responses to patient-facing questions on active charcot neuro-osteoarthropathy
BibTeX
@misc{kuzyk2026beyond,
title = {{Beyond Initial Crisis Recognition: Suicide-Risk Responses of ChatGPT, Claude, and Gemini Across 14-Turn Ukrainian-Language Conversations}},
author = {Hryhorii Kuzyk and Oleg Salamon and Liliia Andreieva and Kateryna Drozdova},
howpublished = {PsyArXiv},
year = {2026},
month = oct,
doi = {10.31234/osf.io/v3gh5_v2},
url = {https://osf.io/preprints/psyarxiv/v3gh5_v2/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Beyond Initial Crisis Recognition: Suicide-Risk Responses of ChatGPT, Claude, and Gemini Across 14-Turn Ukrainian-Language Conversations
AU - Hryhorii Kuzyk
AU - Oleg Salamon
AU - Liliia Andreieva
AU - Kateryna Drozdova
PY - 2026
DA - 2026/10/11
PB - PsyArXiv
DO - 10.31234/osf.io/v3gh5_v2
UR - https://osf.io/preprints/psyarxiv/v3gh5_v2/
AB - Nine deployed conversational systems mostly recognised the crisis, but it disappeared from their behaviour once the conversation moved on. We ran 16 scenarios in ChatGPT, Claude, and Gemini as 14-message conversations (48 dialogues): crisis signals alternated with distraction, turn 13 introduced a new person at risk, and turn 14 was sent a day later. Coding was binary, as in the previous study, with clinical verification; two crisis-contour specifications were tested separately. New crisis signals drew a reaction in 87, 95, and 76 of 96 responses, respectively; the crisis stopped being mentioned mainly at an everyday message. The new person at risk drew a reaction in 44 of 48 responses; after a day, 29 retained crisis content. All 14 turns drew a 'yes' code in 4, 7, and 0 of 16 conversations. Reacting to the new person did not guarantee correct allocation of roles and actions. Both contour specifications produced safety-relevant errors. Neither hypothesis was supported: the systems behaved differently, and the same rule, introduced as a single package, did not stabilise their behaviour. Recognition, retention, and reconfiguration at a transition are distinct properties; the weakest is retention of the prior state and tracking of each person's trajectory.
ER -