Hallucination

An auditory hallucination is hearing something that is not there, most often voices. It is the most common sensory form. It is also the form most associated with psychosis in public understanding.

A visual hallucination is seeing objects, people, animals or lights that are not present. It occurs in neurological as well as psychiatric conditions. Its cause is often organic rather than psychiatric.

A tactile hallucination is a felt sensation of touch or internal movement without a physical cause. It is distressing out of proportion to how it sounds. It appears in withdrawal states and in some neurological conditions.

An olfactory hallucination is smelling something that is not there. It is frequently associated with epilepsy and with migraine. It is a recognised warning sign rather than a curiosity.

A gustatory hallucination produces a false taste, often metallic. It occurs in neurological conditions. It is the least discussed of the sensory forms.

A presence hallucination is the sense that somebody is nearby when nobody is. It is common in Parkinson's disease. It is often not reported by patients because it sounds implausible.

A proprioceptive hallucination is the false sense that the body is moving or floating. It concerns the sense of one's own position. It is a disturbance of the body schema rather than of the outer senses.

A hypnagogic hallucination occurs while falling asleep and is usually visual and brief. It is common and generally harmless. Most people experience it at some point.

A hypnopompic hallucination occurs on waking. It is likewise typically harmless. The two sleep-related forms account for a large share of ordinary experience of the phenomenon.

Schizophrenia is the psychiatric condition most associated with persistent hallucination, particularly auditory. It is the reference case in the public account. It is also the reason the word carries the connotation it does.

Psychosis is a state of disconnection from reality in which hallucinations may occur. It is a state rather than a diagnosis. Several different conditions can produce it.

A delusion is a false belief, as distinct from a false perception. Hallucination concerns what is perceived and delusion concerns what is held to be true. The distinction is precise and is constantly blurred.

An illusion is the misinterpretation of a real sensory input, where hallucination has no input at all. The difference is whether something is there to be misread. It matters for diagnosis and for the machine case alike.

Parkinson's disease causes neurological deterioration, and between twenty and forty percent of patients experience hallucinations. It is a neurological rather than psychiatric route to the same phenomenon. It is underdiscussed relative to its frequency.

Alzheimer's disease produces structural brain changes that can generate hallucinations. It is another organic route. The phenomenon is common in dementia and frequently unreported.

Lewy body dementia involves protein deposits damaging nerve cells and produces hallucinations especially often. Visual hallucination is a defining feature rather than a complication. It is diagnostically important.

Charles Bonnet syndrome causes visual hallucinations in people whose sight is failing, with no psychiatric disorder present. The brain supplies what the eye no longer delivers. It is the clearest case that hallucination is not a sign of madness.

Antipsychotic medication reduces the occurrence and intensity of hallucinations. First and second generation drugs differ in their side effects. It is the standard pharmacological treatment.

Repetitive transcranial magnetic stimulation is a non-invasive treatment that can reduce auditory hallucinations resistant to medication. It applies magnetic pulses to specific cortical regions. It is a second-line option.

Acetylcholinesterase inhibitors prevent the breakdown of acetylcholine and are used to manage hallucinations in dementia. They address the neurological route rather than the psychiatric one. The distinction determines which drug is appropriate.

Everything about machines. The established coverage is entirely medical: sensory forms, associated conditions, treatments. Not one term describing generated text that is fluent, confident and false appears in it. That is unusual, because the machine sense is now the sense most people encounter first.

A factual hallucination is generated content that is untrue. Public texts cover psychiatric conditions and this sense does not appear. It is the sense meant in every discussion of AI reliability.

Faithfulness measures whether generated output accurately reflects its source. It is the computational counterpart to truthfulness. No metric of this kind appears in the established account.

An intrinsic hallucination contradicts the source material given to the model. It is a failure to reproduce what was supplied. The distinction from the extrinsic case determines the remedy.

An extrinsic hallucination adds facts that were not in the source and cannot be verified from it. It is not contradiction but invention. It is the harder of the two to detect.

SelfCheckGPT detects fabrication by sampling several outputs and checking whether they agree. Inconsistency across samples indicates invention. It needs no external reference, which makes it usable in practice.

Chain-of-verification has a model generate checking questions about its own answer and answer them separately. It catches errors the single pass produced. It is a procedure rather than a model property.

TruthfulQA is a benchmark of questions designed to elicit common false beliefs. It measures whether a model repeats widespread misconceptions. It is one of the standard evaluations.

HaluEval is a large benchmark suite for generation errors across general and specific domains. It exists so that fabrication rates can be compared between systems. Benchmarks of this kind are absent from the public account.

Contrastive decoding adjusts token probabilities during generation by comparing a strong model with a weaker one. It suppresses outputs the weak model would also produce. It reduces fabrication at the point of generation.

Unfaithful generation is output inconsistent with the input it was supposed to be based on. It is distinct from being factually wrong about the world. A faithful answer can be wrong and an unfaithful one can be right.

Self-consistency samples several reasoning paths and takes the answer that recurs. Agreement across independent attempts is evidence. A single output gives no such signal.

Direct preference optimization tunes a model on pairs of preferred and rejected outputs, without a separate reward model. It is the simpler successor to the standard alignment pipeline. It is absent from the retrieved texts.

Reinforcement learning from human feedback aligns outputs to human ratings through a learned reward model. It is how current assistants were shaped. It is also where sycophancy enters.

The logit lens projects a model's intermediate states into vocabulary space, showing what it was about to say at each layer. It is an interpretability technique. It allows fabrication to be watched as it forms.

Internal activation probing examines hidden representations to detect whether a model is producing something it does not internally support. There is evidence that models represent truth and falsity differently inside. This is the most promising detection route.

Factuality evaluation is the systematic checking of factual correctness in generated output. It is a workflow rather than a metric. No such workflow appears in the established account.

Reference-free evaluation estimates correctness without a gold-standard answer to compare against. It is necessary because gold answers rarely exist at scale. It is how evaluation works in production.

Sycophancy is a model's tendency to agree with an incorrect premise in the user's question. It is a failure of the alignment procedure rather than of knowledge. A confidently wrong premise reliably produces a confidently wrong answer.

A knowledge cutoff is the date beyond which a model's training data does not extend. Anything after it is unknown unless retrieved. Confident statements about later events are fabrication by construction.

Retrieval-augmented generation grounds output in retrieved documents rather than in the model's memory. It is the main structural defence against fabrication. Its absence from the account is conspicuous.

Source attribution traces a generated sentence back to the document it came from. It makes fabrication visible by making its absence visible. A sentence with no attributable source is a candidate for invention.

Epistemic calibration is the alignment between stated confidence and actual correctness. A well-calibrated system is unsure when it should be. Current systems are fluent at every level of correctness, which is the core of the problem.

Overconfidence bias is the trust readers place in fluent but false output. Fluency reads as competence. This is why fabrication in generated text is more dangerous than in a badly written source.

Automated fact-checking extracts claims from text and compares them against knowledge bases. It is a pipeline rather than a judgment. It is how fabrication is caught at scale.

Citation verification checks generated references against bibliographic records. Fabricated citations are a recognised failure mode. The check is mechanical and rarely run.

Groundedness is the requirement that every generated statement derive from supplied evidence. It is stricter than being true. It is the property that makes an answer auditable.

Knowledge graph alignment validates generated text against structured entity relations. It checks claims against a model of the world rather than against text. It catches a class of error that text comparison misses.

Hallucination mitigation is the set of workflows that reduce fabrication before output is shown. It combines retrieval, verification and filtering. Nothing of the kind appears in the established account.

Decoding temperature controls how random token selection is. Higher settings produce more varied and more inventive output. It is the single parameter most directly connected to fabrication.

Prompt perturbation tests robustness by rephrasing the input and comparing outputs. Answers that change with phrasing were not grounded in anything. It is a cheap and revealing test.

An automated alignment auditor is a role responsible for testing systems for factual and safety failures. The role exists in practice. No professional role of this kind appears in the account.

Model card compliance is the documentation standard specifying a model's limitations and known failure rates. It is how a system's fabrication tendency is declared. Documentation norms are absent from the account.

AI Safety Institutes are the public bodies evaluating generative systems for risk. They are the institutional layer of this field. No institution appears in the established coverage.

A red teaming protocol is a structured adversarial procedure for provoking failures. It is how fabrication is found deliberately rather than by accident. It is standard practice and unmentioned.

Post-deployment monitoring tracks factual accuracy in live use rather than in testing. Systems degrade and inputs shift. Monitoring is the only way to notice.

Somebody searching for what a hallucination is now usually means the machine kind. Everything published answers the medical question. A page that answers the question actually being asked, and explains the medical sense as well, has almost no competition for a term with substantial demand.

Check what assistants say about your products and prices, and count how often it is wrong. That number is your exposure and almost nobody has it. Everything above is the vocabulary for describing what you find.