What Hallucination Actually Means
A Falsifiable Definition, for Machines and for Minds
1. A Word Doing Too Much Work
"Hallucination" gets used for AI systems producing false output, for a specific and well-defined psychiatric symptom, and, informally, for anyone whose ideas depart far enough from consensus. These are not the same thing, and treating them as one word with one meaning hides more than it explains. This is an attempt at a precise, checkable answer to a direct question — not a defense of any specific claim, but a definition specific enough to be wrong, which a comment thread has no room for.
2. Three Things Getting Called One Name
Clinically, hallucination is a false perception — a sensory experience with no external stimulus, in the tradition going back to the French psychiatrist Esquirol, who defined it as "a thorough conviction of the perception of a sensation, when no external object, suited to excite this sensation, has impressed the senses." Delusion is a different thing entirely: a fixed belief that resists correction even when shown contrary evidence — and clinically, it is explicitly distinguished from a belief based on false or incomplete information, from confabulation, and from dogma, since a person holding any of those can still revise the belief once shown why it's wrong. Confabulation is different again: confidently generating plausible content to fill a gap, without any perceptual component and without necessarily resisting correction once the gap is actually filled with real information.
Worth being precise about a real weakness in how confabulation usually gets defined: it's commonly described as happening "without conscious intent to deceive." That phrase does no real distinguishing work — nobody experiencing a hallucination or a delusion is consciously trying to deceive anyone either. What actually distinguishes confabulation is narrower and more useful: it's specifically a response to a gap — missing information papered over with something plausible — not a corrupted perception and not a belief immune to revision.
3. Where AI "Hallucination" Actually Sits
A real, serious body of research already argues the AI field picked the wrong word. Confabulation — "the production of fabricated, distorted, or misinterpreted memories about oneself or the world," in the clinical definition — maps far better onto what large language models do than hallucination does, precisely because hallucination requires something perception-like and an LLM has no perceptual apparatus to corrupt. This isn't a fringe position; it appears in peer-reviewed work directly comparing the two terms for LLM output.
But the right word depends on the architecture, not just on the fact that it's "AI." A system performing direct state estimation from real sensor input — building and maintaining an internal representation it treats as the current state of the world — is doing something functionally closer to perception than an LLM is. If that internal representation is wrong, calling it a false perception is defensible on structural grounds, not just as a loose metaphor. An LLM's output, several inference steps removed from anything sensor-like, fits confabulation better. Flattening every AI system into one word erases a real, useful distinction between architectures.
4. What the DSM Already Tries to Guard Against — and Where That Guard Failed
Clinical delusion criteria explicitly exclude beliefs otherwise explained by a person's culture or subculture — psychiatry already recognizes, on paper, that mere departure from group norms isn't the same as loss of contact with reality. That safeguard has not always held in practice. Soviet psychiatry diagnosed political dissidents with "sluggish schizophrenia," a category developed by Andrei Snezhnevsky specifically to pathologize dissent — coherent, organized opposition to the regime was documented and labeled as "delusions of reformism." The poet Joseph Brodsky was diagnosed without ever being examined in person. And in the United States, homosexuality was itself classified as a mental disorder in the DSM until the American Psychiatric Association removed it in 1973 — not because any new evidence of impaired reality-contact emerged, but because the classification itself had never reflected one. Both are real, documented cases of the third failure mode below being enforced as if it were the first.
5. Historical Cases, Without Flattening Them
Popular retellings of "genius rejected as crazy, later vindicated" tend to simplify in a way the actual history doesn't support, and the simplification matters for getting the definition right.
Semmelweis was correct about handwashing and puerperal fever, and the medical establishment's rejection of him was real and well documented. He was also committed to an asylum and died there — historians still debate how much of his later decline was reaction to persistent, professional rejection and how much reflected a separate neurological process. Both can be true simultaneously; the vindication of his idea doesn't resolve the question of his health.
Cantor was correct that infinite sets admit different cardinalities, and Kronecker's campaign against him within German mathematics was real and damaging to his career. He also had documented depressive episodes and multiple hospitalizations across his life, which most historians treat as a genuine, at least partially independent clinical matter, not simply manufactured by professional persecution.
Tesla's later years are frequently cited as a case of isolation-induced detachment following rejection by the scientific establishment. Some of his later claims — directed-energy weapons, wireless transmission of power through the earth at scale — remain genuinely disputed among historians as to how much reflected real, if unrealized, engineering ideas versus a separate late-life decline. This case, like the other two, resists a clean single explanation.
Copernicus and Galileo belong in a different category, not the same bucket. Their heliocentric model was suppressed by theological and institutional authority — the Church rejected the claim because of what it implied, not by diagnosing Galileo with a mental illness. That's a meaningfully different mechanism: institutional suppression of an idea is not the same thing as institutional pathologizing of the person making it, even though both are real historical patterns of rejecting correct claims.
6. The Actual Test, and Why Confidence Can't Be It
The three failure modes above are not distinguished by how certain the person or system feels, how novel the claim sounds, or how much social resistance it meets. They're distinguished by one thing: whether the claim is checkable against real, demonstrable cause-and-effect — the same essential-variable grounding this site's other work depends on — and whether it holds up when checked.
This has to apply evenly, including to whoever is using the framework, or it isn't a real test. Confabulated content feels confident from the inside. A genuine delusion feels confident from the inside. A correct, well-grounded insight that everyone else happens to be wrong about also feels confident from the inside — that's Semmelweis, Cantor, and Copernicus all reporting the same subjective certainty a confabulating system reports when asked how sure it is. Conviction is not self-validating evidence for any of the three categories. The only thing that actually separates them is external, checkable grounding — not the feeling of being right, no matter how strong that feeling is or how good someone's track record is at being right before.
7. What This Doesn't License
None of the above should be read as evidence that psychiatric conditions are simply misdiagnosed non-conformity, or that they're fixed and unchangeable. Psychiatry's own recovery model pushed back specifically against treating serious diagnoses as permanent by default, driven substantially by real, documented patient outcomes that older models didn't predict. Separately, personality-trait research has found real, measurable change is possible, especially with sustained intentional effort — the evidence does not support "this is just how you are" as a settled fact for most conditions. Both of those are real, citable, legitimate correctives to older assumptions.
But this framework is a conceptual and historical tool, not a diagnostic instrument, and it doesn't skip the part where a live, specific case still has to be checked — by people positioned to actually assess it, not resolved by how confident the framework or the person applying it feels. The test in Section 6 cuts both ways: it's exactly as capable of confirming a real concern is founded as it is of confirming a real insight is being wrongly dismissed. Which one applies in any given case isn't something this paper, or a framework alone, can determine from the outside.
8. A Direct Definition
Hallucination, precisely, is not one failure but three that get called by one name. Genuine loss of grounding is a real failure — output or belief that has actually lost demonstrable connection to checkable cause and effect. Confabulation is a different, narrower failure — confident content generated to fill a specific gap, without the rest of the reasoning necessarily being compromised. Social mislabeling is not a failure of reasoning at all — it's an institutional or interpersonal judgment applied to reasoning that remains grounded, using non-conformity as if it were evidence of ungroundedness. The only reliable way to tell them apart is checking against reality, not against confidence or consensus — which is a harder, slower answer than a comment thread wants, and the actual one.
References
Esquirol, J. E. D. (1845). Mental Maladies: A Treatise on Insanity (E. K. Hunt, Trans.). Lea and
Blanchard, p. 93. (Original work published 1838.)
Schnider, A. (2003). Spontaneous confabulation and the adaptation of thought to ongoing reality.
Nature Reviews Neuroscience, 4(8), 662–671.
Drescher, J. (2015). Out of DSM: Depathologizing homosexuality. Behavioral Sciences, 5(4), 565–575.
Zajicek, B. (2018). Soviet psychiatry and the origins of the sluggish schizophrenia concept, 1912–1936.
History of the Human Sciences.
American Psychological Association. Delusion — diagnostic criteria and cultural exclusion. DSM-5.