presentsUNAIVERSE
The UN AI timeline

    Learn AI /Large Language Models /Large language model

    Large language modelLLM

    Also called: LLM, LLMs, large language models.

    Detail

    In five seconds

    What it is

    A very large text-processing model predicts likely next pieces of language from a prompt and uses that pattern-matching ability to generate or transform text.

    In plain words

    An LLM is a very large text model that reads a prompt and generates likely next pieces of text until it forms an answer.

    An everyday example

    Amina and Luis are in a car driving between Nairobi and Arusha, and the radio signal keeps fading. They hear most of a news sentence clearly, but a few words are lost in static. Using the words before and after the gap, they make a best guess about the missing phrase so the sentence sounds complete. As more of the sentence arrives, they keep adjusting their guess one piece at a time.

    Where the story stops working

    A radio listener is trying to recover words that were actually spoken, while an LLM usually generates new text rather than restoring an original hidden sentence. Also, people bring real-world knowledge and intentions to the task; the model mainly applies statistical patterns from training and the current prompt.

    Remember thisA large language model generates text by repeatedly predicting the next token from context. Its output may be fluent and useful, but fluency is not evidence that the content is correct.

    This term is contested

    The basic idea is widely used, but the exact threshold for what counts as “large” is not fixed, and related terms such as “understanding” in connection with these models remain debated.

    Understand it

    Understand it

    Why it matters

    Non-specialists make better choices about procurement, risk, policy and daily use when they know that an LLM is a pattern-based text generator, not a guaranteed source of truth or a complete product by itself.

    How an LLM turns a prompt into generated text

    After viewing this, a reader should be able to say that an LLM reads tokenised input, predicts one likely next token at a time, and repeats the process to build an output.

    1. 01

      Prompt enters

      The process starts with text from the user or another system.

      Draw a box labelled “Prompt” containing the text: “Summarise this meeting in 3 bullet points.”

    2. 02

      Text becomes tokens

      The text is split into smaller units the model can process.

      Draw a second box labelled “Tokens” with separate rounded chips such as “Summ”, “arise”, “ this”, “ meeting”, “ in”, “ 3”, “ bullet”, “ points”.

    3. 03

      Next-token probabilities

      The model calculates which next token is more or less likely in this context.

      Draw a third box labelled “Candidate next tokens” with four horizontal bars: “-” 42%, “The” 21%, “1.” 18%, “In” 9%, with one option highlighted as selected.

    4. 04

      Output grows and repeats

      The selected token is added to the text, then the model predicts again until it reaches a stopping point.

      Draw a fourth box labelled “Growing output” showing text building step by step, such as “- Budget approved …”, and add a loop arrow from this box back to the probability box labelled “repeat”.

    A left-to-right flow diagram with four boxes connected by arrows. Box 1 shows a user prompt as a short sentence. Box 2 shows that sentence split into token chips. Box 3 shows the model assigning probabilities to several candidate next tokens, with one highlighted as selected. Box 4 shows the growing output sentence, then a loop arrow from Box 4 back to Box 3 labelled “repeat until stop”. Use plain rectangles, token chips and percentage bars in SVG; no decorative icons.

    On hover or tap, each step highlights and reveals one sentence of explanation; a play control animates the addition of two or three example tokens before stopping.

    Describe this without the diagram

    This diagram explains an LLM as a four-step flow. First, a text prompt enters the system. Second, the text is split into smaller units called tokens. Third, the model scores several possible next tokens and chooses one based on those scores and settings. Fourth, the chosen token is added to the output, and the same prediction step repeats until the text is complete. The main lesson is that generation happens incrementally, one token at a time, rather than all at once.

    With reduced motion: A static four-box flow is shown instead of animation. The final box contains three successive output states side by side, for example: “-”, then “- Budget”, then “- Budget approved”, to show repeated token addition without motion.

    A large language model, often shortened to LLM, is a type of language model trained on very large amounts of text. “Large” usually refers to its scale: many model parameters, meaning many adjustable numerical settings learned during training, and often very large training data and computing effort. “LLM” is the acronym; “LLMs” is simply the plural.

    What goes in is text, broken into small units called tokens. A token may be a whole word, part of a word, punctuation, or another short text fragment. The model reads the prompt token by token, uses patterns learned during training, and estimates which token is likely to come next.

    What happens inside is not a database lookup and not human-style understanding. The system computes probabilities across many possible next tokens, then selects one according to its settings. It repeats that process again and again, so one token leads to the next until it produces a completion such as an answer, summary, translation or draft.

    What comes out is generated text that may be useful, fluent and well structured, but it can still be wrong, invented or misleading. An LLM is the underlying model, not necessarily the full product a person uses. A chatbot or AI assistant may contain an LLM, plus extra parts such as instructions, safety rules, tools or document retrieval.

    How “Tuning a radio and filling in a weak sentence” maps onto the real thing

    In the storyIn the AI system
    Words lost in radio staticThe next token is not fixed in advance, so the model estimates likely options
    The words heard before the gapThe prompt and previous tokens provide context
    Making a best guess for the missing phraseThe model predicts a likely next token based on learned patterns
    Continuing the sentence piece by pieceText generation happens one token at a time

    A worked example

    A reporting officer in a regional bureau needs to shorten an already-cleared paragraph for a background note before it goes to the chief of section.

    Starting from: Prompt: “Summarise in one sentence: The inter-agency working group will hold informal consultations to collect comments on the draft workplan before the revised version is circulated for clearance.”

    1. The prompt is divided into tokens and passed to the model.
    2. Using its learned parameters, the model estimates a probability distribution over possible next tokens, shaped by the instruction and source paragraph.
    3. It selects or samples a token, adds it to the emerging summary, then repeats the calculation using the expanded context.
    4. Generation stops when the model produces an ending or reaches another configured limit.

    Result: Possible output: “The working group will consult on the draft workplan before circulating a revised version for clearance.”

    In a United Nations workplace

    At a shared services centre, the duty officer is coding open-ended answers from a needs assessment into agreed categories for a monitoring report due for clearance at close of business. An approved language model is used to suggest first-pass labels for de-identified excerpts.

    Given the categories, instructions and example answers in the prompt, the model predicts a likely label or short rationale for each excerpt. This can accelerate initial coding, but pattern matching does not establish what a respondent meant, so the duty officer must review the suggestions and check consistency.

    Before you try thisNeeds-assessment responses may contain personal or sensitive details. They must not be pasted into a public or otherwise unapproved tool. Use only a system approved for that data, remove identifiers and unnecessary detail, restrict access, and require human checking before any coding is used in the monitoring report.

    Where you may hear it

    • A vendor demo during a procurement unit meeting
    • An ICT governance board reviewing a proposed writing assistant
    • A front-office discussion on clearance of AI-generated talking points
    • An inter-agency working group comparing text-analysis tools

    Avoid confusion

    Avoid confusion

    Often confused with

    Common misconceptions

    • An LLM is the same thing as a chatbot.

      A chatbot is an interface or product for conversation. It may use an LLM underneath, but it can also include retrieval, safety filters, memory, tools and workflow logic.

    • Because the text sounds confident, the output is probably correct.

      Fluent wording comes from learned language patterns, not from automatic fact-checking. The model can produce false or invented statements that still sound convincing.

    • Large means it has simply stored the whole internet and looks up sentences.

      Training adjusts many numerical weights so the model captures statistical patterns. It does not work like a normal search engine or exact phrase archive, though memorisation can occur in some cases.

    • LLMs understand language in the same way people do.

      People often use “understand” informally because the outputs can be coherent, but the mechanism is token prediction from learned patterns. Whether that counts as understanding in a stronger sense is debated, and the term can mislead.

    Quick check

    Which statement best describes an LLM?

    Go deeper

    Go deeper

    The simple version

    An LLM is a very large text model that reads a prompt and generates likely next pieces of text until it forms an answer.

    The technical version

    An LLM is a large-scale language model with many learned parameters that processes token sequences and generates a completion by repeatedly estimating a probability distribution over the next token and sampling or selecting from it.

    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

    A large language model generates text by repeatedly predicting the next token from context. Its output may be fluent and useful, but fluency is not evidence that the content is correct.