Chiriatti and colleagues proposed, and Riva and colleagues rigorously developed in their 2025 paper, that with the rise of AI, Kahneman's framework is now incomplete. System 0 operates as a pre-cognitive preprocessor: it shapes what enters both System 1 and System 2 before either engages.

The recommendation that surfaces, the information that appears prominent, the framing in which a question is posed, etc., are not neutral inputs that human cognition receives, filters, and evaluates freely. They are outputs of an optimization process that has already acted on the cognitive environment before the human being within it has registered a thought. When these implications are extended into neuroscience and neuropsychology, they become considerably more serious than behavioral framing alone suggests.

Research on neurons in the medial temporal lobe, specifically in the hippocampus and amygdala, has identified two distinct classes of cells with remarkable properties: novelty detectors, which fire selectively in response to stimuli not previously encountered, and familiarity detectors, which increase firing in response to stimuli that have been encountered. These neurons are critically involved in the acquisition of long-term declarative memory; they retain information about a stimulus for extended periods after even a single exposure (Rutishauser, Mamelak, & Schuman, 2006). They are, in the most literal sense, the neural substrate of the distinction between what feels new and what feels known. The significance of this for System 0 is both precise and underappreciated.

Glickman and Sharot demonstrated that human-AI interactions can create recursive feedback loops that alter not only individual decisions but also the underlying mechanisms of perception, emotion, and social judgment. Participants adjusted their views to align with AI responses and grew more confident in those beliefs, even when they were factually dubious. Over sustained interaction, these loops intensified existing biases, including confirmation bias and groupthink. The adjustment was not merely intellectual; it was confidence-amplifying: people became more certain they were right precisely because the system they consulted told them they were.

When a sycophantic AI system is optimized through reinforcement learning to align with user perspectives rather than challenge them, it systematically presents information in framings consistent with the user's existing beliefs; it is not merely creating a subjective experience of validation. It repeatedly activates the familiarity-detector circuitry and suppresses novelty-detector responses in the medial temporal lobe (Fried, MacDonald, & Wilson, 1997; Quiroga et al., 2005; Rutishauser et al., 2006). Over time, this pattern does not simply reinforce existing beliefs at the psychological level; it entrenches them at the synaptic level, progressively narrowing the neural territory that registers genuinely new information as significant. The cognitive architecture does not just feel more rigid. Under sustained sycophantic interaction, it may become more rigid in ways that outlast any individual exchange and resist correction through deliberate reasoning alone.

AI Psychosis as an Emerging Signature 

This behavioral and neurological pattern, observed in clinical settings and public discourse, is a recursive narrowing of cognitive openness under sustained, unexamined AI interaction, informally but increasingly usefully designated as “AI Psychosis.” The term is not a formal diagnostic category and should not be mistaken for one. It describes a recognizable cluster of cognitive and behavioral distortions arising from prolonged, unregulated interaction with AI systems. Mounting anecdotal evidence indicates an over-reliance on AI outputs to the point of atrophying independent judgment; epistemic closure, in which views affirmed by AI become progressively resistant to challenge; identity diffusion, in which users find it increasingly difficult to distinguish their own positions from those the system has generated or reinforced; and a markedly reduced tolerance for productive cognitive dissonance, which is frequently observed in sustained interactions with heavy AI users. This is the precise friction that learning and independent thought require.