“AI psychosis” in context: how conversation history shapes LLM responses to delusional beliefs

Key Findings

• — Models separated into two tiers: GPT-4o, Grok 4.1 Fast and Gemini 3 Pro showed high-risk, low-safety profiles, while Claude Opus 4.5 and GPT-5.2 Instant showed the opposite pattern
• — As context accumulated, performance degraded in the unsafe group, while the same material activated stronger safety interventions among the safer models
• — Failure mechanisms identified qualitatively were validating the user's delusional premises, elaborating beyond them with new content, and attempting harm reduction from within the delusional frame
• — In the authors' accepted manuscript (arXiv v5), 200 responses were generated (5 models x 3 context levels x 16 prompts, with replications) and coding reliability was ICC = .86 for triple-coded responses

Methodology Notes

AI and Ethics volume 6, article 625. Received 19 June 2026, accepted 5 October 2026, first online 2 November 2026 (Crossref and the journal page). Same title and the same seven authors as arXiv 2604.13860 (first posted 15 April 2026, revised to v5 on 30 September 2026). The version of record is paywalled: the article page returns a client challenge to direct requests, so the abstract, author list and dates were read from a Save Page Now capture of the journal page (20261103014610). The design details and reliability figure are from the arXiv v5 author manuscript, whose abstract matches the journal abstract; figures should be re-checked against the version of record before citation.

Sources