LLMs hallucinate — everyone has seen what it looks like. The model reports, confidently and coherently, with the right intonation and terminology, something that isn't there. Engineers treat this as a bug: one being fixed, one about to be fixed.
It won't be. This is not a technical fault but an inherent property of language, which the machine has made visible.
Language does not need reality
Language is a unique system precisely in this: it emerges from reality, yet needs no reality for its own coherence.
The word-as-gesture and the word-about-a-gesture are built the same way. From inside language they are indistinguishable. "I lifted a stone" and "I lifted the sky" are grammatically equal; what tells them apart is the consequence — the response of whatever doesn't care how it is described: the stone, the hand, the weight. Inside language there is no filter and never was; the check against reality is always performed from outside.
Early human societies solved this by survival. Convince the tribe there are no lions beyond the hill, be wrong — and the carrier of unfaithful speech was eaten along with his speech. Today there is no such filter. Fidelity has been replaced by coherence. And systems resting on grammar alone are doing very well.
A word on "fidelity." The Russian "vernost" carries two senses at once: correspondence — a faithful answer, one that matches how things are — and allegiance: fidelity to an oath, to a sovereign, to a spouse. Both senses are about holding to something outside yourself. That is exactly what speech loses here. Speech is faithful not when it copies the world, but when it answers to something that does not care how it is phrased. And only what can be faithful can betray — which is why the question of lying will return at the end.
Inside language, a coherent construction and a representation of reality are indistinguishable. By means of linguistic checking alone, there is no way to tell a merely coherent construction from one that corresponds to how things are.
I am not denying that language reflects reality. The trouble is that language may also fail to reflect it, masking the absence of fidelity with coherence. This simply has to be understood and kept in view.
The trap
We live inside this trap; it is unavoidable — we are creatures of speech. But its presence has to be recognised.
This is the condition of language's existence. Precisely because coherence is self-sufficient, language can be carried, accumulated, and used to build what is not present in any actual experience — mathematics, law, plans for tomorrow. The same condition that can deceive is also what gives language its power.
The problem with philosophy and with LLMs is that language turns into thinking. Description comes to feel like gesture.
The clearest example of people who have lost the fidelity of language: those who write "re-sponsibility is response-ability," "con-sciousness is knowing-with." They lean on what itself requires support. Breaking a word into morphemes looks like uncovering a hidden meaning, but it produces nothing except a connection. Words are explained by words, and there is no path out to reality.
In a Telegram channel, the following programmatic thesis appears:
"The Bridge-Theorem": The development of AGI and Super-AI is the process of creating a Shared Oracle, which transfers the entropic cost of compression out of the transmission channel and into the model's hidden weights. This makes operational supra-entropic compression (the Ideal Archiver) possible and proves that AI and compression are two sides of the same coin, where overcoming the classical SEL is a necessary condition for the onset of the Technological Singularity. © Ko. V.
To many readers the text looks coherent: it uses terminology, it looks scientific. But if you have any ML background at all, this elaborate construction falls apart into entirely meaningless claims. What are "hidden" weights supposed to be? What does "supra-entropic" mean? What exactly is being transferred out of the channel and into the weights? And so on.
A coherent construction looks handsome and authoritative. That is its only merit.
Here is an excerpt from a philosophy channel — a fragment of a very long text:
So, what requires justification is the causal criterion of existence:
Causalist postulism — the indispensability of postulating an entity within the best causal explanation is sufficient ground for postulating it.
I justify it thus:
Causalist explanationism — accepting assumption 2 is indispensable for the conscious explanation of actuality.
Explanationist-rationalist obligationism — the conscious explanation of actuality is a rational duty (this undertaking is rationally non-optional).
Obligationist justificationism — if D is a rational duty and accepting P is indispensable for discharging D, then accepting P is a rational duty and P is justified.
Therefore, accepting 2 is a rational duty, and 2 is justified.
© Exalted_Kitharode
To a non-philosopher the text is opaque, and it commands respect through its depth and rigour. But not one of its claims is connected to the world; these are words following from words. One cannot call this wrong. It is simply language about language — text for the sake of text, a system for the sake of a system. This is neither good nor bad; everyone is entitled to their own picture of the world. It is only worth remembering that it is a picture, and that it was painted by language.
The test
How can you quickly gauge how far an author has drifted from reality? Try assessing two marks of speech that holds together on coherence alone.
A term does not unfold downward. Ask for it to be explained through something simpler — it either vanishes, or unfolds into terms of the same kind, and so on in a circle.
Negation changes nothing. If a claim and its denial are equally acceptable to both speaker and listener, the claim conveys nothing.
And the central question: what would have to happen if the claim were false? If there is no answer, then coherence is at work, and nothing else.
Frictionless speech
Politicians, philosophers, executives and self-help grifters each build their own frictionless speech — coherent, logical, and requiring no reality.
The reasons differ. For the politician, a checkable claim is dangerous. For the executive, it may become a commitment. For the grifter, it may destroy the product being sold. The philosopher has no material that would push back. And for an LLM, technically, the answer simply has to close on something — anything, so long as it is coherent.
Different reasons, one result. Because the matter is not the intent of the speakers. Wherever there is no external check, language inevitably drifts toward coherence, because coherence is the only thing still being checked. This is a property of the environment. What happens to reality meanwhile falls outside the frame of language.
Before long, philosophers themselves will realise that they differ from an LLM in nothing.
That is perhaps too strongly put — but take a philosophical text with the author's name removed, and a model's text on the same subject. If your way of telling them apart comes down to "I know the author," then within the speech itself there is no difference.
You will say I am exaggerating, I am sure. Perhaps. But contemporary philosophy does seem to be sliding steadily toward a substantive convergence with LLMs: the same thoughts recombined in different arrangements, according to the conventions accepted in the philosophical milieu — and in the case of an LLM, according to the statistics of the corpus it learned.
The pure case
An LLM is not an anomaly but a pure case: a system with no external check whatsoever.
The politician has an opponent who will try to refute and attack him. The philosopher has history, which grinds slowly but grinds. Even the grifter has a client who will one day refuse to pay. The model has nothing. It is language in laboratory purity, released from the last external filter.
And so the LLM displays what was always there but stayed hidden behind the figure of the author. As long as coherent text was written by a person, we assumed out of habit that fidelity stood behind the coherence — the author was risking something, meant something, leaned on something. The model removed the author, and it became visible that coherence gets on perfectly well by itself.
Only someone with access to the difference between the faithful and the merely coherent can lie: the executive speaks of optimisation where the matter is mass layoffs; the politician evades the check, the philosopher defers it, the grifter hides from it. All of them grasp what is actually the case. The model evades nothing. It has neither the function nor the organ for checking; it simply is language.
This is why an LLM's hallucination is speech in its natural state. Not a lie, not an error. The coherence of language, which merely seems to be something more.
Ask a model to explain why leaves turn yellow in autumn — to a child, to a colleague, and to a specialist. There will usually be an error; but the more demanding the addressee, the more convincing the error and the deeper it is buried. To the child the model lies by inventing a purpose: leaves fall "to cover the roots." To the specialist it lies with terminology: triacylglycerides of chlorophyll, phytochrome as a decay product. The model is not scaling up fidelity — it is scaling up coherence to fit the audience. Fidelity does not enter the calculation at all.
Developers try to solve this with RLHF (reinforcement learning from human feedback), but RLHF only teaches the model the form of coherence that pleases the evaluators. It adds no external check; it scales up the internal one. The hallucinations become more sophisticated, more legible to the user, more confident and more syntactically well-founded.
Engineers, incidentally, are the only ones who deal with this honestly: they measure rather than listen. The distinction runs not through the quality of the speech but through whether there is anything that does not depend on how it is spoken of. The code either runs or it does not. Though there is room for self-deception here too — writing your own benchmark, or making the benchmark itself the goal, is its own kind of coherent construction, in the engineering idiom.
Whether hallucinations can be fought, how reliability prompts work, and why refining a prompt increases hallucinations rather than reducing them — that is a subject for a possible separate article.
Conclusion
This text is speech as well. It is coherent, it is about language, and language has nothing with which to check it — I tell you that confidently.
That is why it is published here. For an article about the absence of any external check on language, the only chance at fidelity is the criticism of its readers. Politicians, philosophers and executives who refuse criticism do not even have that chance.
Ask the question: what would have to happen if this text were false? It would mean that fidelity and coherence are distinguishable from inside language. And then a person can, and should, tell false constructions from sound ones by eye — and developers will be able to build a universal detector that flags an LLM's hallucinations.
And if instead of criticism there comes another comment — a coherent objection that sounds equally good in the affirmative and in the negative — well. Then coherence is at work.