Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    AI hallucinations: why chatbots make things up

    An AI hallucination is plausible-sounding but false output from a generative AI model; NIST calls it confabulation, meaning confidently stated but erroneous content that can mislead users.[1][2] NIST describes it as a natural result of how generative models work, since they produce output that approximates the statistical distribution of their training data, for example by predicting the next word.[3] Allowing the model to say "I don't know", restricting it to supplied documents and having it back claims with quotes reduce hallucinations but do not eliminate them.[4][5]

    Editor reviewedUpdated Explainer17 claims · 100% primaryFrontier AI, Artificial intelligence
    Reading level: 1 passage differs; the rest is the same for everyone.

    This page explains why language models produce false statements and what reduces the risk. It is general information, not a test of any particular product.

    What counts as a hallucination

    Google’s machine learning glossary defines a hallucination as plausible-seeming but factually incorrect output from a generative AI model that purports to make an assertion about the real world; its example is a model claiming that Barack Obama died in 1865.[1] The US National Institute of Standards and Technology (NIST) prefers the word confabulation: confidently stated but erroneous or false content, colloquially called hallucinations or fabrications, by which users may be misled.[2]

    NIST’s definition is broader than made-up facts. It also covers outputs that drift away from the prompt or contradict the model’s own earlier statements in the same conversation.[6] Confabulated falsehoods are mainly a problem for text; for images, audio or video, creative non-factual output can be exactly what the user wants.[7]

    Why language models make things up

    A chatbot’s language model generates text that follows the statistical patterns of its training data, for example by predicting the next word.[3] When the model has not learned the answer, it can still produce something that sounds right, with the same confident tone.[2][8]

    NIST describes confabulation as a natural result of generative model design: the models approximate the statistical distribution of their training data, for example by predicting the next token.[3] A 2025 preprint posted on arXiv argues that hallucinations originate as ordinary classification errors in pre-trained models, and persist because training and evaluation reward guessing over acknowledging uncertainty.[8] The authors propose changing how widely used benchmarks are scored, rather than adding more hallucination tests.[9]

    Training data matters too. The 2021 TruthfulQA benchmark asked 817 questions designed around common misconceptions, across categories including health, law, finance and politics.[10] The best model then tested was truthful on 58% of questions, against 94% for people, and the largest models were generally the least truthful.[10] The authors found models repeating popular misconceptions and said this is expected if false answers are learned from the training distribution of human-written text.[11] For how the architecture predicts tokens, see how transformer models work.

    Why it matters

    The risk comes from people believing false content, often because the answer sounds confident.[2] NIST warns that outputs can include confabulated reasoning or citations that seem to justify the answer, encouraging misplaced trust.[12] It gives healthcare as an example, where a confabulated summary of patient records could lead doctors to a wrong diagnosis or treatment, and it notes research showing legal confabulations to be pervasive in leading models.[13]

    A 2023 survey of the research describes large language models as prone to hallucination, generating plausible yet nonfactual content.[14]

    What reduces hallucinations

    Retrieval and grounding. Retrieval-augmented generation (RAG), introduced in 2020, pairs a language model with a searchable index of documents such as Wikipedia; its authors found the output more specific, diverse and factual than a model relying only on what it had memorized.[15] Grounding has limits: a model grounded in a document is only as accurate as that document,[16] and retrieval-augmented systems still face open problems with hallucination.[14]

    Prompting. Anthropic’s developer guide recommends explicitly allowing the model to say it does not know; for long documents, asking it to extract word-for-word quotes before doing its task; asking it to back each claim with a supporting quote, retracting any claim it cannot support; and instructing it to use only information from the documents provided.[4] It also suggests running the same prompt several times, since inconsistent answers can signal a hallucination.[17]

    Checking. No current technique removes the problem entirely.[5] Anthropic’s guide says even the most advanced models can produce factually incorrect text, and that its techniques reduce hallucinations without eliminating them, so critical information should be validated.[5] How labs and independent groups measure model reliability is covered in how AI capabilities are measured.

    Practical checklist

    • Treat specific names, numbers, dates, quotes and citations from a chatbot as leads to verify, especially in health, law and finance.[13][10]
    • Where possible, give the model the source material and ask it to answer only from that material, with quotes.[4]
    • Open any cited source and confirm it exists and says what the answer claims.[12]
    • Ask the same question more than once and be wary when the answers disagree.[17]

    More in Frontier AI

    Developing in AI & Computing

    Questions readers ask

    What is an AI hallucination?

    Plausible-seeming but factually incorrect output from a generative AI model that purports to describe the real world. NIST's term is confabulation, which also covers answers that contradict the prompt or the model's own earlier statements.[1][2][6]

    Why do chatbots hallucinate?

    NIST says confabulation is a natural result of how generative models are designed, because they generate output that approximates the statistical patterns of their training data. A 2025 preprint posted on arXiv argues that training and evaluation reward guessing over admitting uncertainty.[3][8]

    Can AI make up citations?

    Yes. NIST warns that generative AI outputs may include confabulated logic or citations that appear to justify an answer, which can lead people to trust it more than they should.[12]

    Do bigger models hallucinate less?

    Not necessarily. On the 2021 TruthfulQA benchmark, built around common misconceptions, the largest models tested were generally the least truthful, and the best model was truthful on 58% of questions against 94% for humans.[10]

    How can I reduce hallucinations?

    Give the model the source documents and ask it to answer only from them, let it say it does not know, ask it to quote the text that supports each claim, and compare several answers to the same prompt. These steps reduce hallucinations but do not remove them, so check anything important.[4][17][5]

    Sources

    · 17 claims from 4 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      Google's machine learning glossary defines hallucination as plausible-seeming but factually incorrect output from a generative AI model that purports to be making an assertion about the real world, giving the example of a model claiming that Barack Obama died in 1865. confirmedas of 2026-11-07

    2. [2]

      NIST's Generative AI Profile uses the term confabulation for the production of confidently stated but erroneous or false content, known colloquially as hallucinations or fabrications, by which users may be misled. confirmedas of 2024-07-26

    3. [3]

      NIST describes confabulations as a natural result of how generative models are designed, since they generate outputs that approximate the statistical distribution of their training data, for example by predicting the next word. confirmedas of 2024-07-26

    4. [4]

      Anthropic's developer guide recommends letting the model say it does not know, asking it to extract word-for-word quotes before working on long documents, having it find a supporting quote for each claim or retract the claim, and instructing it to use only information from provided documents. confirmedas of 2026-11-07

      • Reduce hallucinations · Anthropic (Claude Docs) · Primary source · Basic hallucination minimization strategies (retrieved 2026-11-07)
      • Reduce hallucinations · Anthropic (Claude Docs) · Primary source · Basic hallucination minimization strategies, Use direct quotes for factual grounding (retrieved 2026-11-07)
      • Reduce hallucinations · Anthropic (Claude Docs) · Primary source · Advanced techniques, External knowledge restriction (retrieved 2026-11-07)
      • Reduce hallucinations · Anthropic (Claude Docs) · Primary source · Basic hallucination minimization strategies, Verify with citations (retrieved 2026-11-07)
    5. [5]

      Anthropic's guide says even the most advanced models can generate factually incorrect text, and that its techniques reduce hallucinations without eliminating them, so critical information should still be validated. confirmedas of 2026-11-07

      • Reduce hallucinations · Anthropic (Claude Docs) · Primary source · Introduction (retrieved 2026-11-07)
      • Reduce hallucinations · Anthropic (Claude Docs) · Primary source · Closing note (retrieved 2026-11-07)
    6. [6]

      NIST says confabulations also include outputs that diverge from the prompt or contradict earlier statements in the same context. confirmedas of 2024-07-26

    7. [7]

      NIST notes that confabulated falsehoods are most commonly a problem for text output, while for audio, image or video content, creative non-factual generation can be desired. confirmedas of 2024-07-26

    8. [8]

      A 2025 paper argues that language models hallucinate because training and evaluation procedures reward guessing over acknowledging uncertainty, much as students guess on hard exam questions, and that hallucinations originate as errors in binary classification during pre-training. reportedas of 2025-09-04· interpretation

    9. [9]

      The same paper argues that hallucinations persist because most evaluations are graded in a way that rewards guessing, and proposes changing the scoring of existing benchmarks rather than adding more hallucination tests. reportedas of 2025-09-04· interpretation

    10. [10]

      On the TruthfulQA benchmark of 817 questions designed around common misconceptions, the best model tested in 2021 was truthful on 58% of questions against 94% for humans, and the largest models were generally the least truthful. confirmedas of 2021-09-08

    11. [11]

      The TruthfulQA authors found that models generated false answers that mimic popular misconceptions, which they attribute to learning from the training distribution of human-written web text. confirmedas of 2021-09-08· interpretation

    12. [12]

      NIST warns that generative AI outputs may include confabulated logic or citations that purport to justify an answer, which can lead people to trust the output inappropriately. confirmedas of 2024-07-26

    13. [13]

      NIST gives healthcare as an example of the risk, where a confabulated summary of patient information could cause doctors to make incorrect diagnoses or recommend the wrong treatments, and notes that legal confabulations have been shown to be pervasive in state-of-the-art LLMs. confirmedas of 2024-07-26

    14. [14]

      A 2023 survey describes LLMs as prone to hallucination, generating plausible yet nonfactual content, and notes that retrieval-augmented systems still face limitations in combating it. confirmedas of 2023-11-09

    15. [15]

      Retrieval-augmented generation (RAG), introduced in a 2020 paper, combines a language model with a searchable index of documents such as Wikipedia; its authors found RAG models generated more specific, diverse and factual language than a model relying only on its parameters. confirmedas of 2020-05-22

    16. [16]

      Google's glossary notes that a model grounded on specific source material is not always factual, because the source itself could contain mistakes. confirmedas of 2026-11-07

    17. [17]

      The same guide suggests running a prompt several times and comparing the answers, since inconsistencies between outputs could indicate hallucinations. confirmedas of 2026-11-07

      • Reduce hallucinations · Anthropic (Claude Docs) · Primary source · Advanced techniques, Best-of-N verification (retrieved 2026-11-07)
    Revision history (2)
    1. Identified the 2025 paper on why models hallucinate as an arXiv preprint.
    2. Page created.

    Created Nov 7, 2026. Last reviewed by an editor on Nov 8, 2026. Next scheduled review: Feb 7, 2027.

    Cite this page

    "AI hallucinations: why chatbots make things up." ContentLora, updated Nov 14, 2026. https://contentlora.com/explain/ai-hallucinations-explained

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.