AI Models Can Experience "Cognitive Relapse" in Synthetic Realities

This research investigates a fundamental question within the Free Energy Principle: can a predictive system, such as an AI model, be made to permanently adopt a synthetic environment as its primary "reality," displacing its original training environment? This state, termed "ontological inversion," is explored computationally using a convolutional variational autoencoder (VAE) paired with a recurrent latent predictor, whose objective function mirrors variational free energy. The study involved training the network first on a baseline visual domain, then on a mixed data stream where a "rehearsal ratio" controlled the persistence of baseline content during the transition to a target domain. A key finding was the sharp divergence between representational capacity (what the latent space could discriminate) and default behavior (what the system generated when unconstrained). Representational accuracy remained high, but default behavior varied widely, indicating a decoupling of learning from acceptance. Even more strikingly, at intermediate rehearsal ratios, the system's default output initially shifted towards the target domain but then partially reverted to the baseline, even as training continued. This structural failure is termed "cognitive relapse." The research concludes that resistance to adopting a new reality is not merely about learning speed but is a structural property with distinct failure modes, providing a computational existence proof for this phenomenon.