Human–AI Self-Excited Loop Simulation

This repository presents a minimal dynamical systems model exploring the possibility of self-excited dependence loops in Human–AI interaction. The model investigates whether certain AI support structures can induce a feedback cycle in which: AI usage → short-term satisfaction → increased reliance → reduced autonomy → further AI usage The framework is intentionally abstract and focuses on structural dynamics rather than empirical prediction. Core state variables include: - Autonomy A(t)- Reliance R(t)- External difficulty D(t)- AI usage U(t)- Successful experience proxy E(t) Short-term satisfaction S(t) is modeled as a sigmoid-transformed proxy reward incorporating: - Immediate convenience from AI usage- Comfort associated with dependent state- Risk-related short-term stimulation- External difficulty as a negative factor A key modeling assumption is that risk-containing satisfaction may be positively reinforced, reflecting tendencies reported in public analyses of AI usage behavior. The simulation evaluates phase structure across parameter space under three AI policy designs: - Delegating (AI substitutes effort)- Feedback (AI provides guidance)- Empower (AI enhances capability) Phase diagram analysis indicates that delegation-centered support structures are more likely to generate self-reinforcing disempowerment loops, whereas empowerment-oriented designs significantly reduce such structural risk. The results suggest that the topology of support (delegation vs empowerment) is more decisive than the amount of support in determining long-term stability. This model does not claim empirical validation. It serves as a structural exploration of potential Human–AI closed-loop dynamics. All simulations and figures are generated using the implementation provided in this repository.