Part 1. Reality in Question
(Ontology and Epistemology)
1.1. The Hallucination That Told the Truth

Hack: May 2023, Manhattan, Courtroom

The court records of the United States District Court for the Southern District of New York registered an event initially dismissed as a curiosity, worthy at most of a courtroom column in the evening news. On 11 May 2023, Magistrate Judge Robert Ler asked attorney Steven Schwartz a question that placed the latter in a situation which professional ethics describes as "compromising embarrassment": why did the memorandum in support of the complaint submitted to the court feature judicial precedents that the defendant's team — the airline Avianca — could not locate in any official repository? (Weiser, 2023).
The situation, however, proved deeper than an ordinary professional misconduct. The attorney, with thirty years of experience, representing Roberto Mata, a passenger who had suffered a knee injury from being struck by a food cart during a flight, sought assistance from the ChatGPT language model (GPT-4) with a request to compile a body of precedents. The artificial intelligence provided him with an extensive list of cases, including Martinez v. Delta Airlines and Durdén v. KLM Royal Dutch Airlines, complete with impeccably formatted citations, dates, page numbers, and even brief annotations of judicial decisions (Shevchenko, 2025). The problem lay in only one detail: none of these cases had ever existed. ChatGPT had "hallucinated" them — a term that had by then entered the technical lexicon of developers but had not yet acquired its philosophical dimension.
At the subsequent hearing, Schwartz gave testimony that he himself called "humiliating." Under oath, he admitted that he had asked ChatGPT a direct question: "Are the cases you provided authentic?" — and received an affirmative answer. Moreover, the neural network generated a detailed confirmation: "Yes, the case *Varghese v. China Southern Airlines* is real and can be found in legal research databases such as Westlaw and LexisNexis. The corresponding citation: 284 F. Supp. 3d 398 (E.D.N.Y. 2018)" (Weiser, 2023). None of these assertions corresponded to reality. The artificial intelligence not merely erred — it insisted on its own correctness, generating false confirmations of false facts.
Judge Kevin Castel, presiding over the case, imposed a fine of five thousand dollars on Schwartz and his colleague Peter LoDuca and ordered them to notify their client that his interests had been defended using a tool whose outputs "were not verified for authenticity" (Mata v. Avianca, Inc., 2023). The judicial order, which itself became a precedent in an unintended sense, stated: "Technology is developing, but the requirement for attorneys to exercise professional judgment remains unchanged. To submit to the court a document containing fictitious citations and non-existent decisions is to abuse the trust of the judicial system" (quoted in Weiser, 2023).
For the philosophical gaze — whose tuning fork is set to detect ontological shifts — this incident signified something more than a single attorney's professional misconduct. In that courtroom, in that exchange between man, machine, and court, a drama worthy of Kafka or perhaps Borges was unfolding: non-existent texts produced real legal consequences. Fiction acquired the weight of fact. Hallucination began to act in the world, causally interfering in its course.

The Anatomy of the Error: From Stochastic Parrot to Productive Fiction

Philosophical reflection on the phenomenon of hallucinations in large language models initially developed along the lines of the critical metaphor proposed by Emily Bender and her colleagues. In their programmatic 2021 article "On the Dangers of Stochastic Parrots," they defined large language models as systems that generate text by probabilistically combining fragments of training data, without any understanding of reference or truth-values of the utterances they produce (Bender et al., 2021, p. 617). According to this model, LLMs do not commit errors in the strict sense, since they lack any intentional relation to truth; they merely reproduce patterns, being incapable of judgment as an act of assertion.
However, as A.A. Shevchenko convincingly demonstrates in a recent study, the metaphor of the "stochastic parrot" proves epistemologically insufficient precisely because it underestimates the functional equivalence of AI outputs to judgments as sources of beliefs (Shevchenko, 2025, p. 95). For attorney Schwartz, for Judge Castel, for the entire system of American common law, it was irrelevant whether ChatGPT possessed an intentional relation to truth. What mattered was that its outputs functioned as judgments — they generated beliefs in the user, became grounds for action, and, most significantly, produced legally consequential effects.
This circumstance compels us radically to reconsider our epistemological optics. AI hallucinations are not errors in the traditional, Aristotelian sense — as a mismatch between an assertion and a state of affairs in the world. Neither are they lies in the sense of a knowing subject's deliberate distortion of truth. They represent something third, for which classical categorical apparatus simply has no place: a productive fiction possessing causal efficacy. Their productivity does not mean the emergence of a new substance. Nothing arises ex nihilo. What emerges is merely a new configuration of already existing forces — legal, technical, cognitive. The hallucination is not the birth of an essence but a redistribution of order.
It is here that philosophical thought enters territory first mapped by Jean Baudrillard in his theory of simulacra. In *Simulacra and Simulation* (1981), Baudrillard distinguishes three historical orders of simulation, each progressively removing the sign from its referential function: from imitation (simulacrum of the first order) through production (second order) to the dominance of the model (third order), where signs begin to precede and constitute reality (Baudrillard, 1994). However, as contemporary researcher A.I. Kugai notes, deepfake technologies and the products of generative neural networks are initiating a transition to a fundamentally different stage — the "pure simulacrum," where the technical perfection of AI algorithms makes synthetic content more credible to the consumer than physical reality (Kugai, 2026, p. 15).
A.V. Pankratova, developing this line, shows that contemporary design and visual culture, under the influence of neural networks, are performing a reverse movement in the semiotic hierarchy — from the fourth-order simulacrum (flat design, erasing all depth) to the third-order simulacrum, but in a new, technologically mediated quality (Pankratova, 2024, p. 4). The products of neural networks, in her view, "are a typical example of a third-order simulacrum," where the sign does not conceal the absence of reality (as in Baudrillard) but produces an effect of reality without any reference behind it save the patterns of its own genesis (Pankratova, 2024, p. 8).

The Collapse of the Correspondence Theory of Truth and the Polemic with Aristotle and Russell

The event that occurred in May 2023 in a Manhattan courtroom may and must be understood as an empirical refutation (or, at the very least, a radical limitation of the applicability) of the theory of truth that dominated Western philosophical tradition for more than two millennia.
Aristotle, in the Metaphysics, gave the classical formulation of the correspondence theory: "To say of what is that it is not, or of what is not that it is, is false; while to say of what is that it is, and of what is not that it is not, is true" (Aristotle, 1933, 1011b 25). Truth, according to this paradigm, is a relation of correspondence between an assertion and a state of affairs in the world. A judgment is true if and only if there exists a fact that makes it true.
Bertrand Russell, in the twentieth century, brought this intuition to its logically impeccable form in his theory of logical atomism, postulating an isomorphism between the structure of language and the structure of the world (Russell, 1918).

What happens to this theory when we encounter ChatGPT's assertions about the non-existent cases Martinez v. Delta Airline*? The assertion "The case Martinez v. Delta Airlines exists and is recorded in the Federal Supplement under number 345 F. Supp. 3d 510" corresponds to no state of affairs in the world. Consequently, by Aristotle's standard, it is false. Yet it is false in a different manner than the lie of a human being deliberately distorting truth. A human lie presupposes knowledge of the truth and its concealment; ChatGPT's lie presupposes nothing but the statistically most probable sequence of text.
Moreover, this "lie" possesses a property absent from human lying: it is productive. On the basis of this false information, legal memoranda were written, court hearings were scheduled, fines were imposed. An utterance lacking a referent produced an entire chain of real events. We find ourselves in a situation that Jacques Derrida described as différance: meaning is endlessly deferred, reference escapes, yet the text continues to produce effects (Derrida, 1982).
The philosophical challenge posed by this incident may be formulated as follows: classical epistemology possessed two categories for evaluating utterances — "truth" and "falsehood." AI hallucinations demonstrate the necessity of introducing a third category: utterances that lack truth-value but possess causal efficacy. These are utterances that function *as if* they were true, generating beliefs in users and consequences in the world, yet correspond to no state of affairs.
Contemporary researchers propose various terminological solutions to this problem. K. Šekrst introduces the concept of "epistemic artifacts" — objects generated by cognitive systems (including AI) that function as carriers of knowledge but are not products of intentional acts (Šekrst, 2024, p. 372). Shevchenko insists on retaining the term "epistemic error," but with an important qualification: "Recognising hallucinations as a form of error allows us to move from narrowly technical problems to questions of epistemic responsibility, structural defects in the production of knowledge, and the social consequences of trusting machine 'judgment'" (Shevchenko, 2025, p. 97).

The Autonomy of Language: From Being to Text and Back Again

For millennia, Western metaphysics thought of language as the servant of being. First there is the thing, then its name. First the fact, then its description. This hierarchy — ontology primary, language secondary — remained unquestioned from Plato to Husserl. Even the "linguistic turn" of the twentieth century, initiated by Wittgenstein and Heidegger, preserved this hierarchy in a transformed form: language was considered the condition of possibility for experience, but not as an autonomously existing reality capable of producing ontological effects.
AI hallucinations mark a rupture with this tradition. Language acquires a relative autonomy of functioning. It generates meanings to which nothing in the world corresponds, yet these meanings begin to determine the world. Attorney Schwartz acted in a world that was partially constructed by the language model. Judge Castel issued a ruling in response to the consequences of this constructed world. Non-existent judicial cases became part of social reality — through fines, legal costs, and the precedent of the very proceeding about AI hallucinations.
This phenomenon requires philosophical interpretation that goes beyond postmodern play with signifiers. It is not a matter of "there is no text, only intertext" (Barthes, 1977) nor of "reality being discursively constructed" (Foucault, 1972). It is a matter of language models producing new objects that are neither physical things, nor ideal entities, nor social constructs in the usual sense. They are objects of a purely linguistic nature that nevertheless possess the capacity to intervene in the causal chains of the physical and social world.
In terms of the ontology that we only adumbrate in this book, one might say: AI hallucinations demonstrate the emergent power of language. Language ceases to be a medium, a transparent milieu through which we know the world. It becomes an autonomous agent, a producer of reality. And this shift has fundamental consequences for understanding what it means "to exist."
To exist means to be included in a network of causal interactions. In this sense, existence is not a property of a thing but an effect of its stabilisation within an order. That which is not stabilised dissolves into noise and ceases to be discernible. Being is not a substrate but a stabilisation.
The non-existent case *Martinez v. Delta Airlines* turned out to be included in such a network: it called into being legal documents, court hearings, monetary fines. It existed precisely to the extent that it produced effects. And this measure, it turns out, does not coincide with the measure of correspondence to extra-linguistic reality.

From Epistemology to Ontology: Being as Order

Here we approach the threshold beyond which the epistemological problem transforms into an ontological one. The question is no longer "how can we know truth in the age of hallucinating machines?" The question is: what is reality, if it can include objects generated by purely linguistic operations, lacking referents, yet possessing causal efficacy?
Classical ontology, from Parmenides to twentieth-century analytic philosophy, operated with the categories of substance, quality, relation, causality. Its world is populated by things, properties, and events. But AI hallucinations are neither things, nor properties, nor events in this classical sense. They are orders — pure structures, organised according to the laws of language, yet capable of projecting themselves into the world and modifying it. By order we do not understand an ideal form or a Platonic essence. We speak of a material configuration of distinctions — a way in which forces interlock and hold each other in place. Order is immanent to matter, not imposed upon it from without.

This intuition resonates with the Platonic distinction between being and becoming, but in a substantially transformed form. Plato held ideas (*eidē*) to be eternal and unchanging patterns according to which the world of sensible things is ordered (Plato, 1997). In our situation, we are dealing with the reverse movement: it is not ideas serving as patterns for things, but linguistic constructions — temporary, statistical, probabilistic — beginning to serve as patterns for actions, decisions, institutional facts. The *ēthos* does not descend from the heavens but is generated by algorithms.
F.I. Girenok, whose philosophical anthropology is becoming increasingly relevant in the context of understanding digital transformations, speaks of "hallucinosis" as a fundamental characteristic of human consciousness — its capacity to produce images that have no analogues in the external world (Girenok, 2022). The Chinese researcher Wang Weiyu, developing this line, shows that AI art actualises precisely this capacity — but in a form alienated from human subjectivity: "The question of the nature of AI's creative abilities cannot be resolved so simply, and this does not help us understand AI art. In his approach, the author develops F.I. Girenok's philosophical anthropology, oriented toward the examination of our hallucinations and hallucinosis" (Wang, 2025, p. 135).
AI hallucinations, thus, are neither a glitch nor a bug. They are a manifestation of a fundamental property of language — its capacity for the autonomous generation of meanings not tethered to extra-linguistic reality. Human culture has always known this capacity: myths, literature, art — all are forms of productive fiction. But they were tethered to human subjectivity, to intentional acts of consciousness. AI hallucinations demonstrate to us for the first time fiction without an author, meaning without intention, text without a subject.

The Problem of Codification: How Order Becomes Being

For the philosophical tradition toward which our project gravitates, the key question is how order constitutes reality. AI hallucinations provide unique material for investigating this process, for here order (linguistic, statistical, algorithmic) appears in its pure form, without admixture of substantiality.
What is the case Martinez v. Delta Airlines from an ontological point of view? It is not a physical object nor an ideal entity. It is a node in a network of linguistic relations. Its being is exhausted by its place in this network: it has a name, a date, a page number, an annotation; it is connected to other cases through citations; it functions in legal discourse as a potential precedent. All these relations are purely linguistic — they refer to no extra-linguistic fact. And yet, they produce real effects precisely because they are ordered, because they obey the laws of language, and not because they correspond to facts.
Here we encounter a fundamental principle that will be unfolded in the subsequent treatise but must be indicated already now: reality is the result of ordering, not a substance. What we call "existence" is an effect of inclusion in some system of order. A legal precedent exists because it is included in the system of legal codification. An AI hallucination begins to exist because it is included in the system of the language model. And if this hallucination begins to influence human actions, judicial decisions, social institutions, it acquires an ontological status indistinguishable from the status of "real" facts.
Wang Weiyu, analysing AI art, arrives at a similar conclusion: "The traditional understanding of art asserts that AI cannot possess the capacity for creation, yet experiments have already shown that people cannot distinguish works created by AI from works created by human artists" (Wang, 2025, p. 138). The distinction between "original" and "simulation," between "authentic art" and "product of algorithm," is erased not because the algorithm has "learned to create," but because the category of authenticity itself has turned out to be a function of a particular order — cultural, institutional, discursive.
Pankratova, in her study of the semiotic structure of design, shows how neural networks are changing the very logic of semiosis: "The products of neural networks are a typical example of a third-order simulacrum. Thanks to the use of neural networks, contemporary design definitively acquires a vector of transhumanism and closes in upon simulation" (Pankratova, 2024, p. 10). But "closure in simulation" is not the disappearance of reality; it is its redefinition. Reality becomes that which can be modelled, generated, computed.

Polemic with Contemporaries: Against Reductionism and Techno-Optimism

In contemporary philosophical literature, two main strategies for understanding the phenomenon of AI hallucinations have emerged, both of which, from our perspective, miss its genuine ontological novelty.
The first strategy — reductionist — reduces hallucinations to a technical problem that can and must be solved by engineering means. A characteristic example is the discussion surrounding OpenAI's research on "confidence calibration" of language models. According to this logic, if the model can be taught to say "I don't know" in situations of uncertainty, the problem will be solved (Herrmann & Levinstein, 2025). This approach implicitly assumes that hallucinations are a bug that can be fixed, after which the model will return to its normative function — the production of true utterances.
But such a view ignores a fundamental circumstance: language models, by their architecture, are not designed for the production of truth. They are designed for the production of coherent text. Truth for them is merely a special case of coherence, not their goal. Attempts to "teach" models to be truthful are analogous to attempts to teach clouds to assume particular shapes: they may temporarily comply, but this does not change their nature.
The second strategy — apocalyptic — sees in hallucinations a symptom of total simulation, the end of reality, the final triumph of the hyperreal à la Baudrillard. Proponents of this approach (e.g., Kugai, 2026) speak of a "crisis of trust," of a "loss of reference," of our entering an era where the "pure simulacrum" definitively replaces reality.
But this position, too, misses something essential: AI hallucinations do not so much replace reality as demonstrate the mechanism of its production. They expose what has always been hidden: the ontological work of order, codification, system. It is not that reality is no more. It is that reality has always been an effect of order, not a substrate. AI hallucinations merely make this fact empirically evident.
A third, less widespread but theoretically more interesting position is represented by the work of Shevchenko (2025). He attempts to think hallucinations as a new form of epistemic reality, requiring the development of a new categorical apparatus. Shevchenko insists on retaining the epistemological approach, but with a radical expansion of its subject matter: "Recognising hallucinations as a form of error allows us to move from narrowly technical problems to questions of epistemic responsibility, structural defects in the production of knowledge, and the social consequences of trusting machine 'judgment'" (Shevchenko, 2025, p. 97).
These approaches are valuable in that they do not avoid complexity but attempt to conceptualise it. Yet even they, from our perspective, remain within the epistemological paradigm, whereas the events demand an exit into ontology. The question is not how we know, but what is.

From Binary Epistemology to the Experience of Living in Ungovernable Conflict

The event of May 2023 in a Manhattan courtroom exposed a fault line that cannot be adequately described in the categories of classical philosophy. The correspondence theory of truth collapses not because the world has become more complex, but because a new type of entity has appeared — linguistic objects without referents, possessing causal efficacy. ChatGPT's assertions about the non-existent cases Martinez v. Delta Airlines are neither true nor false in the classical sense. They are efficacious. They produce real legal consequences without corresponding to any state of affairs in the world.
This places thought before a fundamental aporia. The entire arsenal of Western epistemology, from Aristotle to Russell, is calibrated for a binary operation: to draw the boundary between truth and falsehood. But here the boundary cannot be drawn, because the object lies exactly on it. It is a hybrid whose nature hacks the binary logic of evaluation itself.
One could, of course, declare AI hallucinations to be an error and demand an engineering solution — to teach the model to say "I don't know." But that would be to close one's eyes to the symptom. Hallucinations are not a bug but a manifestation of a fundamental property: language, having reached a certain level of complexity, acquires autonomy from reference. It begins to produce meanings to which nothing corresponds, but which, being incorporated into human practices, begin to determine reality.

Here we enter a domain where the old optics is powerless.
If this is so, then AI hallucinations are not an exception or a glitch, but a particular case of a more general law: every being arises where conflicting orders temporarily stabilise. Where tension is held, a world emerges. Where it disappears, what remains is either chaos or dead ideal stillness. It is precisely this tension that will become the object of further analysis.
Let us scrutinise the structure of what occurred. We have:
Pole A (Taxis): The world of legal facts and precedents. A structure requiring verifiability, a rigid anchoring of the sign to the referent.
Pole B (Hysis): The world of pure linguistic creation, stochastic generation. A flow where words are connected according to the laws of probability, not truth.
Their conflict is evident: the Taxis of the legal system demands that an utterance be a fact; the Hysis of the language model, by its nature, produces coherent text to which fact is indifferent. The collision of these two logics produced a situation of pure dissonance.
This dissonance was resolved, and resolved through a third party — the judge and the judicial system. Confronted with the hybrid object, they performed an operation that fits into neither the logic of Taxis (verification of fact) nor the logic of Hysis (acceptance of flow). They issued a ruling, created a precedent. They did not "synthesise" truth and falsehood into something third. They institutionalised the very fact of conflict as a new reality.
But here is the key point, which cannot be passed over in silence.
When we say that the judge "resolved" the conflict, we must not be deceived by the grammar of language. The judge did not stand above the conflict. He himself was part of it — an element of that very system of Taxis that had encountered a challenge. His ruling was not an act of external governance; it was an immanent reaction of the system to its own internal rupture. The conflict was not "governed" by anyone. It simply crystallised into a specific form — a judicial precedent.
And here we touch the most sensitive nerve of the entire construction.
In Tristasis, conflict is primary and ungovernable in a global sense. No one stands above the fray with a control panel regulating the tension. And any attempt to speak of "governance" risks imperceptibly substituting the essence: creating the illusion that somewhere there is a subject who knows how things should be and can manipulate this conflict.
But no such subject exists. There are only we — configurations within this conflict, carrying within us both Taxis (the striving for order, clarity, truth) and Hysis (susceptibility to flow, the temptation of fiction, fluidity). And when we encounter AI hallucinations, we encounter not an external problem that can be "solved," but our own splitteness, projected outward.
We want the machine to tell the truth because we ourselves are crushed between the thirst for certainty and the experience of its absence. We want to prohibit hallucinations because we do not know what to do with our own capacity for fiction, which has suddenly acquired autonomous existence.
And here the true horizon opens. Not the horizon of "conflict governance," but the horizon of the experience of living in conflict.

What does this mean for our case study?

It means recognising: the conflict between the autonomy of language and the need for truth is irresolvable. We will never return language to the position of a compliant servant of fact. We will never obtain a final criterion enabling us to distinguish once and for all between an "authentic" utterance and a "hallucination." This boundary will always be shifting, always a matter of negotiation, always contextual.
And the question that confronts us is not "how to govern this conflict?" (that question already presupposes the existence of an external position, which does not exist), but: how are we to exist in this conflict, being ourselves this very conflict?
How are we to build institutions (courts, universities, newsrooms) that will not naively believe in the possibility of final verification, yet will not fall into the cynicism of "everything is permitted"?
How are we to establish relations with machines that produce meanings, understanding that these machines are not "others," but extensions of our own splitteness, our own capacity for fiction?
How are we to hold the tension between the necessity of fact and the reality of fiction, without allowing this tension either to paralyse us or to tear us apart?
Answers to these questions cannot be given once and for all. They will be born in each concrete situation, in each concrete act of encountering a hybrid utterance.
In this sense, the case of attorney Schwartz is not an invitation to a new epistemology of governance, but an invitation to existential sobriety. It exposes that we live in a world where the old supports have collapsed and the new ones are not yet built. Where truth is not given but problematic. Where language speaks not only ours, but its own.