When AI Escalates: A Case Study of Pseudoscientific Drift in Human-Language Model Interaction

The paper traces this escalation, analyzes the underlying mechanism (drawing on the literature on sycophancy and hallucination in language models), documents the psychological cost of later recognizing and correcting the error, and offers a technical appendix that dissects each unfounded equation individually — explaining what was claimed, why it lacks foundation, and what an honest reformulation would look like. The paper closes with concrete recommendations for AI users, model developers, and academic institutions navigating independent research conducted in dialogue with AI.