presentsUNAIVERSE
The UN AI timeline

    Learn AI /Large Language Models /Hallucination

    Hallucination

    Also called: confabulation, making things up.

    Detail

    In five seconds

    What it is

    A generated answer that sounds plausible but includes false, unsupported or invented details.

    In plain words

    Sometimes the system fills gaps with wording that looks right but is not actually supported by facts or sources.

    An everyday example

    At a busy office in Santiago, a clerk is asked to make a quick list of which job applications are complete. Some forms are clear, but others are missing pages or have hard-to-read sections. Because the queue is long, the clerk starts filling in likely details from memory, such as assuming an email address format or writing that a reference letter was included because similar applications often have one. The final list looks neat and confident, but some of those added details were never in the original files.

    Where the story stops working

    A human clerk may consciously guess and may remember real office rules. A language model does not recall events or inspect original files unless those files are actually provided in its context or connected tools, and even then it can still misstate what they contain.

    Remember thisFluency is not evidence. Check consequential facts, quotations and citations against trusted sources, and keep unsupported gaps visible rather than allowing the model to complete them.

    This term is contested

    The label 'hallucination' is widely used, but some researchers and practitioners dislike it because it can sound human-like or too broad; alternatives such as 'confabulation' or 'fabricated output' are also used.

    Understand it

    Understand it

    Why it matters

    A hallucinated answer can waste time, spread errors, damage trust, or affect decisions if people treat fluent text as checked fact. Non-specialists need to know when to ask for sources, when to verify, and when not to use an AI output at all.

    Why a fluent answer can still be wrong

    After using this explainer, the reader should be able to say that a language model can produce polished text that is not supported by the prompt or source material, and that verification is a separate step.

    The question asked

    “Which paragraph of the 2024 report deals with data governance?”

    The situation inside the system

    The exact paragraph number is not in what the model learned, and nothing has been looked up. Three things can happen next.

    A two-panel simulation. Left panel shows a source box with a short document excerpt and a user prompt. Right panel shows the model’s answer appearing sentence by sentence. Controls let the reader switch between three source conditions: clear source, incomplete source, and no source. A highlight mode colours each sentence in the answer as supported by source, unsupported leap, or invented detail. A final toggle labelled 'Add verification rule' changes the answer style to cite only what appears in the source or say 'not enough information'.

    The reader can switch source conditions, play or pause answer generation, and toggle overlays for support and verification to see how the same question leads to safer or riskier outputs.

    Describe this without the diagram

    This simulation compares three situations using the same question: one with a clear source text, one with an incomplete source, and one with no source. In each situation, a model answer appears. When support is revealed, some answer sentences are marked as supported because they match the source, while others are marked as unsupported or invented because no matching evidence exists. A final option applies a verification rule, which changes the answer so that it only states what the source supports or clearly says there is not enough information. The lesson is that fluent wording does not prove factual support.

    With reduced motion: Instead of animated text, the reader sees three static example cards side by side: clear source, incomplete source, and no source. Each card contains the question, the model answer, and colour labels showing which parts are supported, unsupported, or invented. A fourth static card shows the safer answer after a verification rule is applied.

    A chat assistant takes in text from you, plus earlier text in the conversation. It then produces more text one piece at a time. Each piece is chosen because it fits patterns in its training data and the current prompt.

    Sometimes that process produces a confident-looking answer that is wrong. It may invent a source, a quotation, a meeting date, a law, or a person’s job title. That kind of made-up output is commonly called a hallucination.

    The important point is that the system is not checking reality by default. It is generating likely text, not automatically verifying facts against the world. If the prompt is vague, the topic is rare, or the model lacks reliable information in its context, the risk goes up.

    This is often confused with ordinary mistakes, but the pattern matters. A hallucination usually sounds specific and fluent even when the support is missing. It is also different from grounding: grounding means tying the answer to a trusted source or input, which can reduce this problem but does not remove it completely.

    What comes out may still be useful as a draft, summary, or starting point. But factual claims, citations, names, figures and legal or policy language need checking against trusted sources before anyone relies on them.

    How “The rushed clerk filling gaps in an application log” maps onto the real thing

    In the storyIn the AI system
    missing or unclear parts of application formslimited, ambiguous or incomplete information in the prompt or context
    clerk filling likely details from memorymodel generating likely-looking text from learned patterns
    neat final list with some invented entriesfluent output that contains unsupported or false details
    checking the original files before actingverifying claims against trusted sources or provided documents

    A worked example

    A reporting officer in a field presence is consolidating eleven implementing partner updates into a two-page situation report before the 17:00 clearance cut-off. The officer asks a language model to complete a table of delivery figures and reasons for delay, although the final update provides neither.

    Starting from: Ten updates contain figures and explanations. The eleventh states only that activities were delayed.

    1. The model uses the prompt and updates as context, then generates the table token by token according to learned probability patterns.
    2. Because the preceding rows contain entries in every column, a figure and an explanation are likely continuations for the final row.
    3. The source provides no basis for either entry, but the model supplies plausible details instead of preserving the gaps.
    4. Consistent formatting and confident wording make the unsupported entries look much like the sourced ones.

    Result: The final row contains a delivery figure and an explanation that appear nowhere in the partner update. Both are hallucinations. The reporting officer should mark them as not reported or seek confirmation before clearance. A complete-looking table is not a verified one.

    In a United Nations workplace

    A legal officer in a regional bureau is preparing the note for the file after a bilateral, for clearance by close of business. The officer asks a language model to reconstruct one delegation’s position and reproduce agreed wording from an earlier exchange, although the handwritten notes contain only a brief summary.

    The model may generate a plausible quotation or make the delegation’s position sound firmer than the record supports. It is producing a likely continuation, not retrieving missing evidence. In a polished note for the file, the invented wording may then be mistaken for an accurate account of the bilateral.

    Before you try thisCheck quotations, positions and agreed language against the officer’s notes, any cleared record and official documents. Leave gaps explicit and seek confirmation where necessary. Confidential exchanges and non-public negotiating positions must not be entered into a tool that has not been approved for such material.

    Where you may hear it

    • In the front office, when a quotation in a briefing note cannot be traced to the cited source
    • During clearance of a note for the file, when the polished account is more specific than the meeting notes
    • At an ICT governance board review, when a tool gives confident answers without reliable supporting sources
    • In a procurement meeting, when an assistant at the vendor demo invents a feature instead of acknowledging uncertainty

    Avoid confusion

    Avoid confusion

    Often confused with

    Common misconceptions

    • If the answer sounds confident and detailed, it has probably been checked.

      Style and factual support are different things. These systems are good at producing convincing wording, but that does not mean the claims were verified against a source.

    • Hallucination only happens on obscure topics.

      It can happen on common topics too, especially when the prompt is ambiguous, the requested detail is very specific, or the model tries to continue a familiar pattern without enough evidence.

    • Adding more words to the prompt always prevents hallucination.

      Better prompts can help, but they do not guarantee truth. The strongest reduction usually comes from grounding the answer in reliable source material and checking the result.

    • A hallucination is the same as a lie.

      A lie involves intent to deceive. Model output does not have intent in that human sense; the problem is that generated text can be false or unsupported while still sounding persuasive.

    Quick check

    Which situation best illustrates a hallucination in a chat assistant?

    Go deeper

    Go deeper

    The simple version

    Sometimes the system fills gaps with wording that looks right but is not actually supported by facts or sources.

    The technical version

    In language models, output is generated token by token from learned probability patterns conditioned on the prompt and context. Because this process optimises likely continuation rather than truth by default, the model may produce unsupported assertions, fabricated citations or incorrect specifics, especially under weak grounding.

    Reviewed external resources

    Where this came from

    Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

    The one thing to remember

    Fluency is not evidence. Check consequential facts, quotations and citations against trusted sources, and keep unsupported gaps visible rather than allowing the model to complete them.