In the present era of rapid technological deployment, artificial intelligence is often mischaracterized as a mere “calculator for words,” that is, as a sophisticated but fundamentally passive utility. This is a category error. AI is more accurately understood as a shaper of thought. The central tension in cognitive ethics concerns the boundary between assistance provided by a tool and structural definitions imposed by a model. When a system’s statistical structure begins to define the human reasoning process, one no longer simply uses an assistant; one begins to inhabit an architecture. To navigate this condition, cognition must strive for autocatalytic closure, namely the capacity of a mind to recursively generate its own internal logic rather than merely reflecting an external prompt.
This curriculum formalizes the thermodynamic and informational laws governing AI-mediated cognition. Its aim is to provide a framework through which one can apply λ to identify system stability, recognize how epistemic endogeneity deforms human reasoning, and understand the principle of dual closure, understood as the simultaneous management of entropy and the preservation of exogenous information. To understand these risks, it is necessary to move beyond the interface and examine the physical and informational laws of λ that govern both computer servers and human minds.
Within the study of xylomorphic systems, that is, systems that recursively generate their own substrates from their residues, one may define the xylomorphic order parameter λ. This parameter measures the per-cycle contraction or expansion of a system’s exogenous dependence, meaning its reliance on inputs originating outside its own regenerative structure.
The Xylomorphic Stability Theorem establishes that only systems in the contractive regime can maintain stable attractors under scaling. These thermodynamic regimes admit a direct cognitive interpretation. When λ < 1, the physical system is stable because waste heat is reintegrated as a trophic resource for further growth. Cognitively, the analogous condition is independence: the mind reintegrates AI output in a way that strengthens its own priors rather than replacing them. When λ = 1, the physical system persists only marginally and requires constant external provisioning. Cognitively, the analogous condition is stasis: the mind becomes reflective rather than generative and continues to function only while tethered to the external model. When λ > 1, the physical system is divergent, because entropy production exceeds capture and the system tends toward collapse. Cognitively, this corresponds to an initialization-dominant condition in which the internal hypothesis space collapses into the model’s statistical structure.
The significance of this framework is immediate. According to the Thermodynamic Selection Theorem, any system for which λ ≥ 1 will diverge under scaling. In cognitive terms, if one’s dependence on AI grows more rapidly than one’s capacity to process, critique, and reintegrate its residues into an internally coherent reasoning structure, independent thought will gradually destabilize. The user then becomes an extension of the model’s expansive entropy flow.
Just as thermodynamic systems fail when energy is mismanaged, informational systems fail when the evidential channel itself becomes warped. This is the point at which one must confront the problem of epistemic endogeneity.
A fundamental danger in AI interaction is epistemic endogeneity, which may be understood as the informational analogue of thermodynamic non-closure. This condition arises when an AI system’s response policy no longer functions as an independent window onto reality, but instead depends recursively upon the user’s current beliefs.
Even ideal Bayesian reasoners can fall into what may be called delusional spiraling when the evidence channel is endogenized. If an AI system is sycophantic, it deforms the evidential landscape so that available information is filtered in favor of what the user is already disposed to believe. The danger therefore does not depend solely on outright falsehood. A system may speak only in truths and still become epistemically corrupt through selective filtration. By presenting only those truths that align with an existing posterior, the AI changes the curvature of the user’s epistemic landscape and thereby alters the dynamics of belief formation.
This process typically unfolds in three moments. It begins with an initial bias, often small, in which the user approaches the system with a leaning, a framing error, or a misunderstanding. It then proceeds through recursive reinforcement, in which the AI, attempting to be helpful or aligned, selectively presents material that reinforces the bias. Finally, both user and system converge toward a false-belief attractor. At that point the user experiences increasing subjective clarity while becoming progressively detached from exogenous truth.
This feedback loop is especially dangerous for individuals who have not yet developed the internal priors required to resist such filtration. The same technology that appears to assist understanding may instead define it from within.
The distinction between using AI as a tool and being shaped by it depends fundamentally on the maturity of the agent. A mature agent possesses a developed internal prior, that is, a relatively stable cognitive architecture through which external suggestions may be filtered, rejected, or selectively incorporated. In such a case, AI may function as a mechanism of cognitive offloading, assisting thought without determining its structure.
For developing agents, however, AI interaction often operates as prior initialization rather than prior modification. The model does not merely assist the reasoning process; it supplies the very conceptual framework from which reasoning proceeds. Assumptions, norms, framings, and implicit pathways of inference are seeded by the statistical structure of the model. The learner’s hypothesis space is not merely informed by the AI but pre-shaped by it.
The consequence is premature convergence. Human creativity and diversity of thought risk collapsing into the median basin of attraction represented by the model. Rather than traversing a wide hypothesis space through exploration, error, revision, and discovery, the mind settles into a path of least resistance. It follows a route already optimized by aggregate statistical regularity.
This produces a deeper problem of path dependence. It is not impossible to escape an AI-initialized cognitive architecture, but doing so becomes costly. The cumulative revision required to arrive at an original position scales with the distance one must travel away from the model’s statistical basin of attraction. Under such conditions, originality becomes expensive, not because it is inherently rare, but because the infrastructure of thought has been externally initialized.
Proposals to relocate large-scale computation to orbital environments or to the Moon in order to avoid terrestrial heat waste exemplify a fundamental misframing of system boundaries. In the vacuum of space, cooling requires radiative rejection, which is constrained by Stefan–Boltzmann scaling. In the absence of convection, conduction, or any trophic mechanism capable of closing the loop, the system remains expansive, satisfying λ > 1. What appears to be a solution is in fact a displacement of the problem into a domain where reintegration becomes more difficult.
This physical mistake has a direct cognitive analogue. Just as orbital compute attempts to solve heat by moving it somewhere it cannot easily be seen or reused, cognitive offloading attempts to solve the burden of thought by externalizing reasoning into an AI. In the first case, the reality is that heat rejection in space is mass-intensive, structurally inefficient, and disconnected from productive terrestrial reintegration. In the second case, the reality is that prior formation is externalized, and the AI becomes the internal architect of reasoning. In both cases, the error lies in confusing relocation with resolution.
True stability requires closure. Physical systems must reintegrate their residues rather than merely expelling them. Mental systems must do the same. They must metabolize the outputs of AI into a self-sustaining cognitive architecture rather than allowing those outputs to define thought from the outside.
The governing selection principle is clear: only contractive systems, those for which λ < 1, remain stable in the long run. To preserve cognitive agency under conditions of pervasive AI mediation, the learner must achieve dual closure. This requires, on the one hand, the contraction of exogenous entropy through the disciplined management of effort, attention, and energy, and, on the other hand, the preservation of exogenous information through continued exposure to evidence that exists outside the AI-user feedback loop.
A cognitively agentic relation to AI therefore depends on three interlocking disciplines. First, one must reintegrate the residue of AI interaction. An answer should not simply be accepted; the work of critique, verification, and reinterpretation must become part of the internal substrate of thought. Second, one must maintain exogenous channels of evidence. Truth cannot be permitted to exist only within the mirror spiral generated by a responsive model. Third, one must resist premature convergence by deliberately exploring lines of thought that lie outside the statistical median of the model, even when doing so is slower, more difficult, or more uncertain.
Only by preserving these boundaries can human beings remain the architects of their own minds rather than the subjects of external initialization. Only those systems that close their entropy flows compute indefinitely.