confabulation
In plain wordsWhen an AI fills a gap in its knowledge with an invented but plausible-sounding answer, much as some people with memory disorders sincerely 'remember' things that never happened.
- 1
Language models
The generation of fluent, confident content that is false or unsupported — fabricated facts, citations, quotations or details — proposed as a more accurate name than hallucinationhallucination, n.: When an AI states something false or made up — a fake fact, quotation or source — fluently and confidently, as if it were true. for the same failure, on the grounds that the model is not perceiving something absent but producing plausible material to fill a gap, without intent to deceive. In a narrower technical usage, Sebastian Farquhar and colleagues (2024) define confabulations as the subset of hallucinations that are arbitrary — answers that change from one sample to the next — as distinct from consistent errors learned from faulty data; such confabulations can be detected by measuring semantic entropysemantic entropy, n.: A way to tell when an AI is likely making something up: ask it the same question several times and check whether the answers mean the same thing. Answers that keep changing signal guessing. across sampled answers.
Asked for sources, the model confabulated a 2019 paper attributed to a real author.
- 2
Psychology and neurology
A disturbance of memory and self-report in which a person sincerely recounts invented or distorted events, or makes up reasons for their own behavior, unaware that the account is untrue. It is characteristic of amnesic conditions such as Korsakoff syndrome, also follows some brain injuries and, in milder forms, appears when healthy people explain their own choices. Confabulators are not lying: they believe what they say.
Origin
From Latin confabulari, 'to talk together'. In neuropsychiatry, Carl Wernicke's 1900 textbook defined it as the appearance of memories of events that never happened, a hallmark of Korsakoff syndrome. Applied to language models by 2023 — the researcher Beren Millidge argued in a March 2023 blog post that LLMs confabulate rather than hallucinate — and Geoffrey Hinton has said the failure 'ought to be called confabulation, not hallucination' (2024).
Usage note
Like 'hallucination', the word borrows from human psychopathology and is criticized as anthropomorphizing; 'fabrication' and 'error' are plainer alternatives. Many writers use confabulation and hallucination interchangeably.
Related
- hallucinationhallucination, n.: When an AI states something false or made up — a fake fact, quotation or source — fluently and confidently, as if it were true.
- semantic entropysemantic entropy, n.: A way to tell when an AI is likely making something up: ask it the same question several times and check whether the answers mean the same thing. Answers that keep changing signal guessing.
- factualityfactuality, n.: How far what an AI says is actually true about the world. A factual model gets facts, numbers, names and dates right and does not make things up.
- abstentionabstention, n.: When an AI declines to answer because it is not sure enough — saying 'I don't know' or asking for more information — instead of guessing and risking a confident mistake.
- calibrationcalibration, n.: How well a model's stated confidence matches reality: when a well-calibrated system says it is 80% sure, it turns out to be right about 80% of the time.
- anthropomorphismanthropomorphism, n.: Treating something that isn't human — a pet, a car, a chatbot — as if it had human thoughts, feelings or intentions. AI that talks fluently makes this temptation unusually strong.
- groundinggrounding, n.: Connecting an AI's words to something real — trusted documents, a live search, images or the physical world — so that its answers rest on evidence rather than only on what it absorbed in training.
- hallucinationhallucination, n.: When an AI states something false or made up — a fake fact, quotation or source — fluently and confidently, as if it were true.
Sources
- Farquhar, S., Kossen, J., Kuhn, L. & Gal, Y. (2024). Detecting hallucinations in large language models using semantic entropy. Nature 630, 625–630. doi.org/10.1038/s41586-024-07421-0
- Wernicke, C. (1900). Grundriss der Psychiatrie in klinischen Vorlesungen. Leipzig: Thieme.
Mentioned in
hallucinationhallucination, n.: When an AI states something false or made up — a fake fact, quotation or source — fluently and confidently, as if it were true., factualityfactuality, n.: How far what an AI says is actually true about the world. A factual model gets facts, numbers, names and dates right and does not make things up., introspectionintrospection, n.: Looking inward to observe your own thoughts and feelings. For AI, the question is whether a model can accurately report what is actually happening inside it, rather than guessing or making it up., botshitbotshit, n.: A blunt term for false or made-up chatbot output that people pass on and use as if it were true, especially at work., semantic entropysemantic entropy, n.: A way to tell when an AI is likely making something up: ask it the same question several times and check whether the answers mean the same thing. Answers that keep changing signal guessing.