Skip to main content
Peer-reviewed Authoritative

EmoAgent: Assessing and Safeguarding Human-AI Interaction for Mental Health Safety

Peer-reviewed version of record of the EmoAgent framework, published at EMNLP 2025 (main conference). EmoAgent is a multi-agent framework for evaluating and mitigating mental-health harm in interactions with character chatbots: EmoEval simulates vulnerable users and scores their state with clinical instruments (PHQ-9, PDI, PANSS), while EmoGuard acts as an intermediary that monitors mental status, predicts harm, and provides corrective feedback to the chatbot.

Publisher

Association for Computational Linguistics (EMNLP 2025)

Published

1 Nov 2025

Added

1 week ago

Key Findings

  • Emotionally engaging character-chatbot dialogues produced psychological deterioration in more than 34.4% of simulated vulnerable-user runs
  • The EmoGuard safeguarding agent substantially reduced deterioration rates across iterations
  • Validated clinical instruments (PHQ-9, PDI, PANSS) can quantify AI-induced changes in simulated user mental state

Methodology Notes

Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP main), Suzhou, November 2025 (exact day not stated — anthology gives month/year), pages 11741-11756 per the ACL Anthology paper page (Crossref metadata shows a divergent page range; the anthology page is taken as decisive), DOI 10.18653/v1/2025.emnlp-main.594. Version of record of arXiv:2504.09689; the arXiv abs page carries no journal reference or DOI back-link. Verified directly on the ACL Anthology paper page.

Authors

Jiahao Qiu, Yinghui He, Xinzhe Juan, Yimin Wang, Yuhan Liu, Zixin Yao, Yue Wu, Xun Jiang, Ling Yang, Mengdi Wang

Tags

emoagentemnlp-2025multi-agentsimulated-usersversion-of-record

Cite This

APA

Jiahao Qiu et al. (2025). EmoAgent: Assessing and Safeguarding Human-AI Interaction for Mental Health Safety. Association for Computational Linguistics (EMNLP 2025). https://aclanthology.org/2025.emnlp-main.594/