Hallucination

Hallucination audio micro-course (transcript)

Three short podcasts about “Hallucination”. Listen with the interactive player.

Podcast 1 — Introduction & meaning

Narrator: Welcome. Today's word is hallucination — spelled H-A-L-L-U-C-I-N-A-T-I-O-N, pronounced huh-loo-sih-NAY-shun, IPA slash hə-ˌluː-sɪ-ˈneɪ-ʃən slash.

Narrator: Hallucination is a noun. In psychology and medicine, it names the experience of perceiving something that is not actually present — seeing, hearing, smelling, tasting, or feeling something without a corresponding external stimulus. The hallucination feels completely real to the person experiencing it.

Narrator: The word comes from the Latin hallucinari — to wander in the mind, to talk idly — possibly from the Greek alyein (to be disturbed or beside oneself). In English since the seventeenth century.

Narrator: Hallucinations are distinguished from illusions — which are misperceptions of real external stimuli — and from delusions, which are false beliefs. A person who sees a snake where there is actually a rope is experiencing an illusion. A person who sees a snake where there is nothing at all is hallucinating. The experience of hearing voices — auditory verbal hallucinations — is one of the most characteristic features of psychosis and one of the most distressing.

Narrator: The causes of hallucinations are diverse: schizophrenia and other psychotic conditions, bipolar disorder, delirium (especially in older adults with acute illness), substance use and withdrawal, severe sleep deprivation, and certain neurological conditions including Parkinson's disease and Charles Bonnet syndrome (visual hallucinations in people with severely impaired vision).

Narrator: The AI hallucination — a term that emerged with the widespread adoption of large language models — names a specific and important failure mode: the generation of false, fabricated, or confabulated information presented with apparent confidence and no acknowledgement of uncertainty. An AI model hallucinating a legal citation, a scientific study, or a historical fact creates the same problem as a human witness hallucinating a memory: plausible, confident, wrong.

Narrator: Remember this: hallucination is the mind generating its own reality — and the risk that the mind generating the AI's responses may do the same.

Speaker A: The phenomenology of hallucinations — what they actually feel like from the inside — is one of the more complex areas in psychiatric understanding, not least because the first-person experience of hallucination is so difficult to communicate. The classic account of auditory verbal hallucination — hearing voices that comment on one's actions, issue commands, or engage in conversation — describes experiences that range from mildly intrusive to profoundly disturbing. The Hearing Voices Network, founded in the UK in 1988, was a pioneering effort to enable people with voice-hearing experiences to discuss them and find peer support.

Speaker B: Yes. And the distinction between pathological hallucination and culturally sanctioned perceptual experience is one of the more challenging areas in cross-cultural psychiatry. Many cultures have traditions in which hearing the voices of ancestors, spirit guides, or divine presences is a valued and meaningful experience. The question of when a voice-hearing experience constitutes a pathological hallucination requiring treatment and when it is a culturally normal perception is not always straightforward.

Speaker A: Right. And the neuroscience of hallucination has advanced considerably in recent decades. Neuroimaging studies of auditory hallucinations show activation of the same brain areas involved in processing real speech — suggesting that the brain is generating and experiencing an internal voice as if it were an external one. The predictive processing account of hallucination — the idea that hallucinations occur when the brain's internal predictions about sensory experience override the actual sensory input — is one of the more influential theoretical frameworks.

Speaker B: Yes. And the AI hallucination problem is one of the most practically significant limitations of current large language models. Unlike human hallucinations, AI hallucinations are not random or bizarre — they are typically plausible, fluent, and difficult to distinguish from accurate output without independent verification. The model has learned to generate text that sounds like reliable information; it does not have a mechanism to distinguish between what it knows and what it is plausibly confabulating. The confident citation of a non-existent academic paper is a canonical example.

Speaker A: Now, drug-induced hallucinations.

Speaker B: Yes. The hallucinogenic drug class — including LSD, psilocybin, mescaline, and DMT — produces hallucinations by acting on serotonin receptors, particularly the 5-HT2A receptor, in ways that disrupt the normal filtering of sensory input and alter the relationship between internal prediction and external perception. The therapeutic use of psilocybin in clinical trials for depression, PTSD, and addiction has generated significant research interest and has forced a reconsideration of the relationship between the hallucination experience and therapeutic benefit.

Speaker A: Right. And delirium — the acute confusional state that affects large numbers of hospitalised patients, particularly older adults — produces hallucinations that are often terrifying and whose management is one of the more challenging aspects of acute hospital medicine. The delirium patient who is convinced they are somewhere dangerous or being attacked is experiencing their hallucinations as completely real; communicating with them in that state requires both clinical skill and considerable patience.

Speaker B: Common note — hallucination, illusion, delusion, confabulation, and pseudohallucination describe related perceptual and cognitive errors. An illusion is a misperception of a real stimulus. A delusion is a false belief, not a false perception. Confabulation is the unconscious fabrication of false memories without deceptive intent. A pseudohallucination is a hallucination-like experience the person knows is not real. Hallucination specifically refers to perception without external stimulus, experienced as real.

Speaker A: The mind's private theatre — presenting to itself what is not there, whether through illness, chemistry, or the architecture of a language model.

Speaker B: Today's word is hallucination. In prompting, it's a psychiatric care, AI reliability, psychedelic therapy, delirium management, and neuroscience word — for AI hallucination detection tools, psychiatric hearing voice support platforms, and LLM output verification systems.

Speaker A: Hallucination spans psychiatric phenomenology, AI reliability, psychedelic therapy, delirium management, and neuroscience — tools built around it serve psychiatrists, AI engineers, therapists, hospital clinicians, and researchers.

Speaker B: Exactly. Watch this AI reliability brief.

Speaker B: Build a Next.js AI hallucination detection and output verification tool for an AI product manager — providing a systematic framework for testing large language model outputs for hallucinated facts, citations, names, and statistics, generating a hallucination risk assessment for specified model outputs and task types, providing verification workflows including source checking and confidence calibration, tracking hallucination rates across model versions and task categories, and producing an AI hallucination risk report.

Speaker A: AI hallucination rates measured with systematic verification workflows so product managers understand where their model confabulates confidently and deploy appropriate human review and source-checking processes.

Speaker B: Right. Now psychiatric support.

Speaker B: Design a Django hearing voices support and hallucination management tool for a psychiatric nurse — providing a structured assessment framework for auditory verbal hallucinations including frequency, content, emotional valence, and command hallucinations, generating personalised coping strategy recommendations including distraction, cognitive reframing, and engagement approaches, tracking hallucination patterns over time, and producing a hearing voices clinical management plan.

Speaker A: Auditory hallucinations assessed with personalised coping strategy recommendations so psychiatric nurses support patients in developing effective responses to voice-hearing experiences.

Speaker B: Now psychedelic therapy.

Speaker B: Create a Svelte psilocybin-assisted therapy and therapeutic hallucination preparation tool for a clinical researcher — documenting the mechanisms by which psilocybin produces hallucinations and altered consciousness via serotonin receptor agonism, providing preparation and integration frameworks for therapeutic psilocybin sessions, tracking the relationship between hallucination content and therapeutic outcome, and producing a psilocybin therapy clinical research analysis.

Speaker A: Psilocybin hallucination mechanisms documented with preparation and integration frameworks so clinical researchers design and conduct therapeutic psychedelic sessions that maximise safety and therapeutic benefit.

Speaker B: Now delirium management.

Speaker B: Build a Vue delirium-related hallucination assessment and non-pharmacological management tool for a geriatric nurse — providing a rapid delirium and hallucination assessment framework for acutely confused older patients, generating non-pharmacological orientation and reassurance protocols for hallucinating delirious patients, providing guidance on family communication and involvement in delirium management, and producing a delirium hallucination management guide.

Speaker A: Delirium hallucinations assessed with non-pharmacological management protocols so geriatric nurses reduce patient distress and orient confused older adults without unnecessary sedation.

Speaker B: Last one.

Speaker B: Develop a React Native LLM hallucination rate benchmarking and model selection tool for a technology buyer — comparing the hallucination rates of major large language models across specified task categories including fact retrieval, citation generation, medical information, and legal advice, providing a structured model evaluation methodology for hallucination testing, generating risk-weighted model selection recommendations for specified use cases, and producing an LLM hallucination benchmarking report.

Speaker A: So hallucination in prompts spans AI output verification, psychiatric hearing voice support, psilocybin therapy, delirium management, and LLM benchmarking — the false-perception word serving every domain where confident unreliable output creates risk.

Speaker B: Exactly. Whether it's a mind or a model, hallucination looks like reality. Build tools that tell the difference.