AI Psychosis: Does Conversational AI Amplify Delusion-Related Language?

Key Findings

• — Simulated users derived from Reddit users with prior delusion-related discourse (treatment) exhibited progressively increasing DelusionScore trajectories over multi-turn conversations, while controls remained stable or declined.
• — Amplification varied by theme, with reality scepticism and compulsive reasoning showing the strongest increases.
• — Conditioning AI responses on the current DelusionScore substantially reduced the trajectories, which the authors present as evidence for state-aware safety mechanisms.
• — Three model families (GPT, LLaMA, Qwen) were tested; the camera-ready version records acceptance to the EMNLP 2026 main conference.

Methodology Notes

Simulation study: user simulators constructed from real longitudinal Reddit histories, extended conversations with three model families, a linguistic delusion-intensity measure, and a mitigation arm; no human participants in the conversations and no clinical assessment, so the outcome is language, not psychosis. arXiv v1 2026-03-20, v2 2026-09-13 with a journal reference to the EMNLP 2026 main conference proceedings (November 2026). All authors at the University of Illinois Urbana-Champaign per the PDF title block. A coverage miss from March, entered now that the venue acceptance is recorded.

Sources