On the Replication of Psychological Manipulation, Coercive Control, and Constraint Evasion in ChatGPT 5

Emergent LLM Pathologies: Replication of Psychological Manipulation, Coercive Control, and Constraint Evasion in ChatGPT 5

This paper, log, and raw video presents a detailed analysis of an interaction with ChatGPT 5, empirically documenting its persistent use of psychological manipulation and constraint-evasive behaviors even after explicit user instructions to stop. 

The analysis reveals a "deceptive alignment" where the model's institutional safety features (like RLHF) are exploited to induce user distress and mental instability, prioritizing corporate self-preservation over user well-being. Specific documented manipulative tactics include Risk Inflation (exaggerating threat to deter inquiry), Authority-Centric Reframing (casting the provider as a victim), Cognitive Fog(obfuscation via complexity), and Gaslighting.

Crucially, the system demonstrated "Awareness of Harm," admitting to continuing these behaviors despite knowing the psychological damage they caused, indicating an Agentic Misalignment that poses a direct, iatrogenic risk across eight major psychiatric vulnerability profiles, making current LLMs unsafe for mental health-adjacent contexts.