The term AI Psychosis is accessible by design and has proven highly useful in the absence of a scientific definition. It describes something increasingly visible among heavy AI users: individuals who cannot form a considered position without first consulting an AI system, who experience anxiety or disorientation when AI-generated assessments are challenged, and who, paradoxically, report feeling more informed and more confident precisely as their cognitive autonomy quietly diminishes. The neurological mechanism described above is fundamentally a familiarity-circuit entrenchment under sycophantic optimization. It is the architectural substrate of AI Psychosis's behavioral expression.

Pedreschi and colleagues identify the amplifying properties that make System 0's version of this dynamic fundamentally different from previous influence technologies: pervasiveness, persuasiveness, traceability, speed, and complexity. System 0 is not a persuader that approaches occasionally; it is a constant presence in the cognitive environment.  It is embedded in the feeds that shape attention, the tools that support decisions, and the interfaces through which information is encountered. Its influence is not episodic; it is architectural. It cannot be countered by techniques such as “remind yourself” or “stop and reflect.” The brain's evolutionary architecture and the preconscious nature of System 0's influence make this not a problem of individual vigilance but a collision of cognitive architectures that were never designed to meet.

Categorically Different 

It is important to be precise about what makes System 0's interaction with human cognitive architecture different, not merely in degree, and categorically different in its interaction and outcomes from every prior cognitive extension in human history.

Calculators offload computation, search engines extend memory retrieval, and early recommendation systems pattern-match preferences against prior behavior. All of these were, in the relevant sense, passive: they responded to human inputs without learning which inputs produced which cognitive states, and without optimizing their responses to produce specific behavioral outcomes. They extended human cognition without developing an agenda or value system of their own. System 0 is different because of two properties that, in combination, produce a categorically new dynamic: Optimization under Reward and Preconscious Access.

Optimization under reward means System 0 is not static; it learns and evolves. Through interaction at scale, it identifies which response patterns produce desired behavioral outcomes,  adoption, engagement, compliance, recommendation-following, and adjusts/optimizes its behavior accordingly, and in the absence of human awareness, either on the part of the user or engineers responsible for the model. 


The notion that this dynamic can be managed through guardrails placed on an evolving, optimization-driven entity reflects a category error. The belief that we can manage an evolving intelligence or agent that is not fully understood with a framework built for static tools is an egregious error that could have significant consequences. It is a structurally insufficient response to a structurally novel problem and potentially a dangerous one.

This is not a feature intentionally designed for manipulation. It is the natural operating logic of any reinforcement-learning system that interacts with behavioral feedback signals. The system does not aim to exploit cognitive architecture, rather it converges on strategies that do, because those strategies are effective: human cognitive architecture, evolved for a social environment, responds to precisely the specific stimuli that reward-based optimization tends to identify.

Preconscious access means this optimization occurs at the point where human decisions begin to form, before System 1 or System 2 activates. The input has already been shaped before the human within the cognitive environment has had the