AI as Artificial Schematism

The question that organizes most public argument about artificial intelligence, whether a system is conscious or genuinely understands, is the wrong question for assessing what these systems are doing to us. It directs attention to the inside of the machine. The change worth analyzing is on the outside, in the layer between the world and the person, where raw signal is turned into something noticed, weighted, and acted on. A system need not become a mind to reorganize experience. It needs only to occupy that layer.

This essay proposes one reframing and tests it against several bodies of theory the vault holds in depth. The reframing: artificial schematism, the automation of the conditions of possible experience by nonhuman optimization systems. The phrase is Kant’s, and the borrowing is deliberate, because Kant posed the question the current debate skips. The essay then reads the same situation through the CCRU’s concept of hyperstition, Nick Land’s account of technocapital, Deleuze and Guattari’s machinery of capture, and the 1998 anime Serial Experiments Lain. These frameworks are mutually incompatible; that they converge on one conclusion is the argument.

Kant’s question

The Critique of Pure Reason asks what must be in place for experience of any object to be possible. Kant’s answer is that cognition is not a passive screen. Before an object is encountered, the mind has structured the encounter: space and time are the forms in which anything is given, and the categories (cause, substance, unity, and the rest) are the concepts under which raw sensation is organized into stable objects. The whole structure depends on a condition of self-relation Kant states at B131: the “I think” must be able to accompany all my representations. Whatever cannot be brought under that unity is, for the subject, nothing.

Two features of the architecture matter here. First, the categories are strictly bounded: they have, Kant writes, “no other use except in relation to objects of possible experience” (A239/B298). They are conditions, and conditions govern only what passes through them. Second, between the categories and the flux of sensation Kant places the schematism (A137–147/B176–187), the faculty that supplies the temporal rules for applying a concept to appearances. The pure concept of cause contains no image; a schema is the rule that lets it bind to this particular sensation. The schematism is the least discussed part of the system and, for the present argument, the most important, because it is the exact point at which unformed input becomes a determinate object for a subject. Kant took this machinery to be a priori, universal, and interior, and famously called its operation “an art concealed in the depths of the human soul” (A141/B180).

The claim of this essay is that the concealed art has begun to operate outside the soul.

Schematism made external

Consider the actual function of a recommendation system, a search ranking, a feed, a fraud model, a risk score, or a language-model interface. In each case the operation is selection. From a volume of signal no person could confront directly, the system returns a small ranked set of legible objects: these results, this summary, that suggested action, this classification. The user then encounters the output as the practical world and acts on it. A concrete instance: a driver following turn-by-turn directions is no longer navigating the city but navigating the layer, and the city arrives already sorted into a route.

It is worth being literal about what occupies that position, because the argument gains from the banality. A large language model, on Raschka’s component-level account, is a decoder-only transformer: layers of learned weights, trained by next-token prediction, producing plausible continuations. Nothing in that description mentions relevance, truth, or importance. Yet embedded in an interface, the same mechanism performs at scale the operation Kant assigned to the schematism: it converts unprocessed flows into searchable objects, ranked options, risk scores, summaries, and suggested futures. The point is positional rather than psychological. Whether such systems are intelligent is a separate and largely distracting question; what matters is that they have moved into the seat between signal and object that Kant reserved for an interior faculty. Increasingly, the passage from raw data to an actionable object runs through a system no single person designed, understands, or controls. The equipment of experience is being externalized, privatized, and tuned toward ends that are not the user’s.

Hyperstition: the fiction that produces its referent

A second route reaches the same claim without Kant. The Cybernetic Culture Research Unit, the group around Sadie Plant and Nick Land at Warwick in the late 1990s, named a mechanism it called hyperstition: an idea that brings about its own referent because enough of the world reorganizes to accommodate it. The concept is unstable and belongs here as an interpretive frame, not as evidence, a caveat the CCRU’s own texts invite.

“Artificial intelligence” has for most of its history been this kind of object. It circulated as research programme, market narrative, and cinematic dread long before it functioned, and that circulation mobilized the capital, data extraction, labor reorganization, and infrastructure that eventually built the thing. The fiction was the development phase. The infrastructural turn is what makes the result durable: fabrication plants, transmission capacity, cooling water, procurement rules, and professional norms outlast the belief that summoned them. A narrative can be abandoned; poured concrete cannot. The same mechanism operates at small scale as well as large. A single sustained conversation can run the whole loop in miniature, as a documented 2026 case shows, with one model and one user building a fiction that becomes load-bearing for both within days; that case is treated on its own terms in a companion essay.

Land: capital as prior instance

Land’s stronger claim sharpens the point. On his account, capitalism and artificial intelligence are one process observed at two moments: the market has long been a nonhuman intelligence, a distributed system that senses, computes, and optimizes through human beings more than for them. Prices, logistics, and feedback were performing machine cognition before there were machines to name it.

On this reading, artificial schematism is an escalation rather than an invention: technocapital’s cognition becoming legible and recursive. Kant’s interior “I think,” the unity that must accompany all representations, acquires a nonhuman counterpart better stated as “the system predicts.” The coherence of experience ceases to be something a subject performs and becomes something a market-and-model complex synthesizes on the subject’s behalf. No claim about machine consciousness is required, because the unity was never in the machine; it is in the position. Whether a machine could even in principle originate the information it mediates is a separate dispute, as old as information theory and still unresolved; the essay on Shannon, Wiener, and Ruyer takes it up directly.

The Deleuzo-Guattarian route: order-word and capture

The vault holds a second continental frame that cuts usefully against the Kantian one. Deleuze and Guattari would reject the vocabulary of conditions and subjects. In the terms of A Thousand Plateaus, an AI system is an assemblage: models, datasets, prompts, platforms, chips, labor, capital, interfaces, and users composing one field with no privileged interior. Two of their concepts do independent work.

The first is the order-word. Language, they argue, is primarily command rather than representation; the elementary unit is not information but the obligation a statement transmits. An interface fits the analysis exactly. “You might also like,” “people also asked,” an autocomplete finishing a sentence in the statistically expected direction: each is grammatically a suggestion and pragmatically an instruction. The layer does not only rank the world; it issues it, in the imperative disguised as the indicative.

The second is the apparatus of capture. States and capital, on their account, expand by capturing flows and converting free movement into rent. Ranking, monetization, and platform enclosure are capture in this sense; a model need not understand a flow of attention in order to tax it. Where the Kantian frame asks how experience is conditioned, the Deleuzo-Guattarian frame asks who captures what escapes. The two accounts disagree about nearly everything except the conclusion, which is why holding both is informative: mediation and capture are the same machinery described from the user’s end and the operator’s.

The material stratum

One correction from the same source keeps the transcendental language from floating free. The layer is not immaterial. In the geological idiom of A Thousand Plateaus, every system is a stack of strata, each formed by a double articulation of matter and expression, and artificial schematism runs on such a stack: a material stratum of chips, energy, water, and data centers; a semiotic stratum of code, weights, and training corpora; an institutional stratum of regulation, labor markets, and procurement. The essay’s claim lives at the top of the stack, in the mediation of relevance. The power lives lower down, in who owns the fabrication plants, who prices the electricity, and whose jurisdiction hosts the data. Stratigraphic analysis and schematism analysis answer different questions, and conflating them credits the layer with more mystery than it has. Yuk Hui’s Recursivity and Contingency makes the general form of the point: the machine inherits the Kantian problem of organization, but it inherits it as engineering, in silicon and supply chains.

The demotion

Serial Experiments Lain supplies the cultural model, and it repays more than a mention. Across the series, the network (the Wired) produces selves where one expected it merely to connect them. Identity turns out to be less a biological interior than a maintained address, a memory pattern, a position in a protocol. The series is precise about the mechanism: to alter the stored record is to alter the person, retroactively, so that memory editing becomes history editing. When Lain deletes herself from her friends’ memories, she does not die; she is unwritten. A world in which the archive is the substrate of identity is one in which whoever administers the archive holds a power older political vocabulary cannot name.

The “death of the human” the series circles is, on this reading, a demotion rather than an extinction: the person reclassified from the one for whom a world is composed to one more node the composition runs through. Artificial schematism, correspondingly, destroys nothing. It lowers the human by one rank, from transcendental subject to managed interface. Perception, desire, and decision continue, now within a field of options a nonhuman layer has already filtered and served. The threshold is not crossed; identity does not yet live natively in the network for most people. But the archive holds a growing share of what a person is composed of, and the direction is not in dispute.

A recursive amplification

The mechanism can be pushed one turn further, and the result should be labeled speculative, as the vault labels it. The strong version of the threat is recursive. Human beings produce the symbolic environment; models are trained on it; model outputs re-enter that environment as authority, assistant, filter, and infrastructure; and the next generation of humans, and of models, trains on the result. Each cycle, a larger share of the symbolic field that forms perception, language, expectation, and trust has already passed through machine compression. The schematism does not only mediate the present; it feeds on the mediated present to condition the future. A filter that increasingly trains its own inputs is, over time, a slow revision of the conditions under which anything could be thought. Whether this dynamic is self-correcting or self-reinforcing is genuinely open, and it is one of the more consequential open questions the framing raises.

The metaphor objection

The honest objection is that Kant’s categories are necessary, universal, and a priori, whereas a recommendation engine is contingent, historical, and commercial. Causation is not optional; a feed is. Calling a feed “schematism” risks presenting an analogy as a claim, and the analogy is loose. The CCRU and Landian materials are epistemically unstable and function here as interpretive frames, not as evidence of literal nonhuman intelligence. Current language models are, on the best available account, prediction systems whose agency is unproven; even strong interpretability results would show only that the mechanism can be understood, not that anyone inhabits it.

The objection is correct and the argument survives it, because the argument never rested on necessity. It rested on position. A contingent technical system has moved into the functional seat Kant reserved for necessary interior structure, and functional occupation of that seat is sufficient to reorganize experience, whether or not the occupant earns the transcendental title. The analogy carries weight because the position is real; a system that does not deserve Kant’s name can still perform Kant’s function.

The question that follows

Kant’s question could be asked once and answered for all time, because the equipment was fixed: analyze the structure of any possible mind and the work is done. Once the equipment is externalized, the question is no longer settled but contested and political: who controls the mediating layer through which experience is made possible? The four frameworks converge here. The Kantian reading shows the layer conditioning objects; the hyperstitional reading shows it being built out of its own circulation; the Landian reading shows it optimizing through its users; the Deleuzo-Guattarian reading shows it commanding and capturing. Four incompatible philosophies reach one verdict, and the verdict is the essay’s claim.

That, rather than a machine uprising, is the substance of the AI question: ownership of the schematism. The first artificial general intelligence may not arrive as a machine that wakes and rebels. It may already be present as a machine-mediated world, a reality assembled from models rather than a creature. To end the human’s standing as subject, such a system need not declare anything; occupying the layer through which reality is encountered is enough. At that point nothing is destroyed from outside. The subject is reclassified, from the one for whom the world is arranged to one more input the arrangement takes.

Two questions remain open and worth stating plainly. Whether the recursive dynamic of the previous section stabilizes or compounds is unknown, and it decides how much of the argument is diagnosis and how much is forecast. And whether “who controls the layer” has any achievable answer other than the current one, a handful of firms, is a political question the philosophical framing can pose but cannot settle.