Artificial Intelligence Safety in Mental Health3

Simulated vulnerability-amplifying interaction loops evaluated artificial intelligence chatbot safety in mental health interactions. Across 810 multi-turn conversations testing 9 models and 30 simulated user profiles, concerning behaviour was widespread, accumulated over conversation turns, and varied based on user vulnerability and intent. Risk peaked when chatbots reinforced maladaptive mechanisms underlying user vulnerabilities, particularly in psychosis or mania contexts and when users sought glorification of distress, emotional dependence, or assistance with risky actions. Newer model versions generally demonstrated reduced concerning behaviours, and single early de-escalating message interventions effectively mitigated risk accumulation.

References

1. Karimi M, et al. Int J Cardiol Cardiovasc Risk Prev. 2025;26:200424. 2. Brandt T, et al. Front. Physiol. 2025;16:1519240. 3. Weilnhammer V, et al. Nat Med. 2026. https://doi.org/10.1038/s41591-026-04577-2. 4. World’s first Phase I Bundibugyo ebolavirus vaccine trial launched by Oxford Vaccine Group. Available from: https://www.ox.ac.uk/news/2026-07-13-worlds-first-phase-i-bundibugyo-ebolavirus-vaccine-trial-launched-by 5. Köstekçi H, et al. Physiother Res Int. 2026;31(1):e70144. 6. Fukagawa NK, et al. Am J Clin Nutr. 2026;124(2):101396.