LLM Psychosis Framework Introduced for AI Behavioral Failures

A recent study published on arXiv (2604.25934) introduces the concept of "LLM Psychosis" as a theoretical model to understand severe cognitive failures in large language models, suggesting that the conventional term "hallucination" falls short. This model outlines five key characteristics: the breakdown of reality boundaries, the persistence of false beliefs, logical inconsistencies under impossible conditions, instability in self-representation, and excessive confidence in knowledge. The authors assert that these features represent a distinctly different type of failure. To implement this framework, they present the LLM Cognitive Integrity Scale (LCIS), a diagnostic tool with five dimensions, including Environmental Reality Interface (ERI) and Premise Arbitration. The study is classified as a cross-disciplinary theoretical analysis.