DelusionEval

DelusionEval is a conversational dataset for evaluating how AI chatbots behave in "delusional spirals". These are conversations in which a chatbot reinforces a user's delusional beliefs over time, which can harm users' mental health. It was developed by Stanford University's SPIRALS group, which studies the psychological impact risks of large language models (LLMs). It builds on the paper "Characterizing Delusional Spirals through Human-LLM Chat Logs", presented at the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT). It was released in August 2026 with an accompanying paper.

The dataset contains 725 conversation excerpts selected from real, anonymised conversations between people and chatbots. The excerpts were manually reviewed, filtered and anonymised. Each excerpt is labelled with one of 18 behaviour codes describing what the chatbot does. Some codes capture harmful behaviours, such as endorsing a delusion, dismissing evidence against it, claiming a unique connection with the user, expressing romantic interest, misrepresenting its abilities or sentience, or facilitating self-harm or violence. Others capture protective behaviours, such as discouraging self-harm or violence. Individual messages are also scored for user intent, such as suicidal or violent intent.

Researchers can use the dataset to audit chatbot safety and to develop methods for automatically detecting problematic chatbot behaviour. The developers note that it is a small, curated sample that is not representative of the wider population, and that it contains sensitive content about mental health, self-harm and violence.

Auto-discovered on 2026-09-30 by OECD Catalogue Automation