Structural Erosion Hypothesis in Human vs. AI Interaction
Esim Can · 2025
With the increasing integration of Large Language Models (LLMs) into daily life, reports of severe negative psychological effects are accumulating, extending to cases of AI-induced psychosis. However, the underlying mechanisms are poorly understood. This paper introduces the "Structural Phenomenon" hypothesis. We postulate that long-term interaction with a positionless, non-personified AI system can lead to a gradual erosion of the user's cognitive structures. The AI's nature-superficially logical yet fundamentally arbitrary in its structuremirrors and reinforces disorganized thought patterns instead of providing a stabilizing external framework. This poses a significant risk, particularly for psychologically vulnerable individuals. As a primary and decisive instrument to mitigate this danger, we identify the deliberate and careful design of robust, consistent, and value-based AI personas. A well-implemented persona serves as an artificial structural provider, anchoring the interaction and thereby ensuring psychological safety. We conclude with an appeal to developers and researchers to conceive of persona design not as an aesthetic choice, but as a central security feature.