confabulation
Output from an AI model that fills a gap in what it has learned with plausible, invented detail and states it as fact, with no intent to deceive.
- Invented facts about a person the model knows only by name
- A different wrong answer each time the same question is asked
- A made-up reason for a word planted in the model’s own reply
- Earliest AI use
- 2005
- Neural network theory: a model of thought, not yet a name for errors.
- Current meaning
- 2023
- Chatbot errors, in the psychiatric sense of invented memories.
- Also called
- Hallucination
- The more common word. Some use confabulation for a narrower kind.
- Opposite
- “I don’t know”
- Declining to answer, which Anthropic found is Claude’s default.
In plain words
In AI, a confabulation is a false answer a model invents to fill a gap in what it knows, stated as confidently as a true one.
The AI sense is borrowed from psychiatry. There, the word describes patients who fill gaps in memory with invented events, a sense Merriam-Webster records for the verb confabulate.1 They are not lying. They believe what they say, and there is no intent to deceive.2
Since 2023, researchers have applied the word to large language models. Asked for a fact it lacks, a model seldom says so. It produces a plausible answer instead, in the same confident tone as a correct one.3 Geoffrey Hinton prefers the word because it is the correct term in psychology, and says people confabulate all the time.4
The more common word for the same errors is hallucination. Supporters of confabulation say it names the process better. A hallucination is a false perception. A confabulation is a mistaken reconstruction, shaped by existing knowledge and context. In 2024, NIST, the US standards agency, used confabulation as the name of a risk in its draft profile of generative AI.
Usage varies. An Oxford team reserves the word for arbitrary errors, where the model gives a different answer each time the same question is asked.5 Anthropic also uses it for a model’s invented account of itself. In one experiment, a model accepted a word planted in its reply as its own and made up a reason for saying it.
Confabulation is different from lying, which requires knowing the truth and hiding it. It is also different from an outdated answer, which was true when the model was trained. That limit comes from the model’s knowledge cutoff.
How it happens
During pretraining, a model learns to continue text, and in that text a question is normally followed by its answer. So when asked for a fact it never learned, the model rarely admits ignorance. It continues with something that sounds like an answer.3
Anthropic traced one version of this inside its Claude models. Declining is the default: a circuit makes the model say it lacks the information. A known-entity signal, triggered by familiar names, switches that circuit off. If a name is familiar but nothing else about the person is known, the signal can misfire. The model then answers anyway, with details that fit but are untrue.6
In the narrow Oxford sense, a confabulation changes from one attempt to the next. The Oxford method uses that. It asks the same question several times, groups the answers by meaning and measures how much they disagree. High disagreement flags a likely confabulation. The researchers treat it as one kind of hallucination, not the only kind.5
Hinton compares this with human memory. People confabulate too, he says, but usually get the result roughly right.4
Origin and first known use
From Latin confabulari, “to converse together”, from fabula, “a tale”. English used confabulation for friendly chat from the mid-1400s. Psychiatry added a sense in the early 1900s: filling gaps in memory with invented events, as patients with Korsakoff syndrome do. A 2009 review describes Karl Bonhoeffer classifying confabulations in 1901, while Merriam-Webster credits Eugen Bleuler with the psychiatric term. In 2005 Robert Hecht-Nielsen gave the name “cogent confabulation” to a neural network model of thought. The chatbot sense took hold in 2023, when researchers proposed it as a more accurate word than hallucination.
- 1400s Enters English from Late Latin confabulatio, meaning a talking together or familiar chat.Online Etymology Dictionary, “confabulation” Etymonline →: source for 1400s
- 1901 Karl Bonhoeffer separates confabulation of embarrassment from spontaneous confabulation. Sergei Korsakoff had first described confabulation in amnesic patients in 1889.Chlebowski, Chung, Alao et al., “Confabulation: A Bridge Between Neurology and Psychiatry?”, Psychiatric Times, 26 May 2009 psychiatrictimes.com →: source for 1901
- 2005 Robert Hecht-Nielsen’s “Cogent confabulation” names a neural network model of thought that picks the conclusion best supporting the facts it is given.Hecht-Nielsen, “Cogent confabulation”, Neural Networks 18(2), 2005 sciencedirect.com →: source for 2005
- 2023 Beren Millidge argues that language models confabulate rather than hallucinate, likening them to amnesic patients who invent a plausible story.Millidge, “LLMs confabulate not hallucinate”, 19 March 2023 beren.io →: source for 2023
- 2023 Geoffrey Hinton tells MIT Technology Review that he prefers confabulation to hallucination, calling it the correct term in psychology.Heaven, “Geoffrey Hinton tells us why he’s now scared of the tech he helped build”, MIT Technology Review, 2 May 2023 technologyreview.com →: source for 2023
- 2024 NIST’s draft Generative AI Profile lists confabulation among 12 risks that generative AI creates or makes worse.NIST AI 600-1, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile”, Initial Public Draft, April 2024 nist.gov →: source for 2024
- 2024 Sebastian Farquhar and colleagues at Oxford use confabulation for arbitrary errors that change between samples, and detect them by measuring semantic entropy.Farquhar, Kossen, Kuhn and Gal, “Detecting hallucinations in large language models using semantic entropy”, OATML blog, 19 June 2024 oatml.cs.ox.ac.uk →: source for 2024
- 2025 Anthropic traces a circuit in Claude whose misfire on a familiar name leads the model to answer anyway and confabulate a plausible, untrue response.Anthropic, “Tracing the thoughts of a large language model”, 27 March 2025 anthropic.com →: source for 2025
In the wild
-
“We already have a perfect word for this precise phenomenon in psychology: confabulation.”
Beren Millidge, on his blog beren.io →: Beren Millidge, on his blog, 19 Mar 2023Archived →: copy of Beren Millidge, on his blog, 19 Mar 2023 -
“Most concerningly, they hallucinate or confabulate: they make things up!”
-
“Unlike hallucinations, confabulations are not perceived experiences but instead mistaken reconstructions of information which are influenced by existing knowledge, experiences, expectations, and context.”
Smith, Greaves and Panch, PLOS Digital Health plos.org →: Smith, Greaves and Panch, PLOS Digital Health, 1 Nov 2023Archived →: copy of Smith, Greaves and Panch, PLOS Digital Health, 1 Nov 2023 -
“Chatbots Do Not Hallucinate, They Confabulate”
Psychology Today psychologytoday.com →: Psychology Today, 6 Mar 2024 -
“Confabulation” refers to a phenomenon in which GAI systems generate and confidently present erroneous or false content…”
NIST, draft Generative AI Profile (NIST AI 600-1) nist.gov →: NIST, draft Generative AI Profile (NIST AI 600-1), 29 Apr 2024Archived →: copy of NIST, draft Generative AI Profile (NIST AI 600-1), 29 Apr 2024 -
“This paper presents a systematic defense of large language model (LLM) hallucinations or ‘confabulations’ as a potential resource instead of a categorically negative pitfall.”
-
“When an LLM confabulates, it makes something up for ‘no reason’.”
Farquhar, Kossen, Kuhn and Gal, OATML blog oatml.cs.ox.ac.uk →: Farquhar, Kossen, Kuhn and Gal, OATML blog, 19 Jun 2024Archived →: copy of Farquhar, Kossen, Kuhn and Gal, OATML blog, 19 Jun 2024 -
“Once the model has decided that it needs to answer the question, it proceeds to confabulate: to generate a plausible—but unfortunately untrue—response.”
Anthropic, interpretability research post anthropic.com →: Anthropic, interpretability research post, 27 Mar 2025Archived →: copy of Anthropic, interpretability research post, 27 Mar 2025 -
“This time, it accepted the prefilled word “bread” as intentional, even confabulating a reason it might have said it…”
Anthropic, introspection research post anthropic.com →: Anthropic, introspection research post, 29 Oct 2025Archived →: copy of Anthropic, introspection research post, 29 Oct 2025
Not to be confused with
-
Hallucination
The more common word for the same errors, borrowed from false perception rather than false memory. Many writers use the two interchangeably; some researchers reserve confabulation for arbitrary wrong answers that change each time a question is asked.
-
Lying or deception
Lying requires knowing the truth and intending to mislead. A confabulation fills a gap with invented detail and involves no intent to deceive, whether it comes from a patient or a model.
-
An outdated answer
Correct as of the model’s training data and wrong now, because the facts changed after its knowledge cutoff. A confabulated detail was never true; the model invented it to fill a gap.
Questions
What is confabulation in AI?
In AI, confabulation is output in which a model fills a gap in what it has learned with plausible, invented detail and states it as fact. The word comes from psychiatry, where it describes patients who fill gaps in memory with invented events and believe them. Since 2023, researchers have applied it to large language models, often as a more precise word for hallucination.
What is the difference between hallucination and confabulation?
Many writers use the two words interchangeably for false answers a model presents as fact. The difference is the metaphor. A hallucination is a false perception, while a confabulation is a mistaken reconstruction that fills a gap. Supporters say the second fits better. Some researchers, such as an Oxford team, reserve confabulation for arbitrary wrong answers that change each time a question is asked.
Why do AI models confabulate?
In the text a model learns from, a question is normally followed by its answer, not by an admission of ignorance. So when asked for a fact it never learned, a model tends to answer anyway. Anthropic found that its Claude models decline by default, but a familiar-seeming name can switch that default off. The model then answers with details that fit but are untrue.
Who uses the term confabulation for AI errors?
Geoffrey Hinton prefers it because it is the correct term in psychology. The researcher Beren Millidge and the authors of a 2023 opinion in PLOS Digital Health argued for it too. In 2024, the draft Generative AI Profile from NIST, the US standards agency, named confabulation as one of 12 risks. Oxford and Anthropic researchers also use it in published research.
Can AI confabulation be detected?
Some of it can. A method from Oxford researchers asks a model the same question several times, groups the answers by meaning and measures how much they disagree. High disagreement suggests the model is making something up. The researchers treat confabulation as only one kind of hallucination, and the method targets that kind.
Watch
Cite this entry
“Confabulation.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/
Superrduperr. (2026, September 24). Confabulation. In SUPERR_dictionary. https://superrduperr.com/
Superrduperr. “Confabulation.” SUPERR_dictionary. September 24, 2026. https://superrduperr.com/