Learn AI /Large Language Models /Language model
Language model
Also called: LM.
Back to the milestone you came from
In five seconds
What it is
A language model estimates probabilities for pieces of language in context; many can use these estimates to generate text from a prompt.
In plain words
A language model is a system that has seen a great deal of text and learned to continue text in likely ways. Give it some words, and it generates more words one piece at a time.
An everyday example
A technician in Vienna receives a short note on a shared tablet from a pump monitor: "Station 4 flow dropped at 03:10. Check..." The technician has read thousands of maintenance notes before and starts guessing how such notes usually continue: "valve pressure", "sensor cable", or "power supply". Each new word narrows the likely next words, so the sentence grows one piece at a time. The technician can often produce a very plausible continuation, even if they do not know what truly happened at Station 4 that night.
A person reading maintenance notes may bring real-world knowledge, memory of the site, and awareness of consequences. A language model does not have that kind of grounded access just from generating text. Also, the model does not literally remember and compare a few whole notes in the way a technician might; it uses numerical patterns learned during training.
Remember thisA language model generates text one token at a time from probabilities conditioned on context. Plausible wording is not evidence that the output is accurate.
Understand it
Understand it
Why it matters
You may hear about language models in chat assistants, translation tools, writing aids, coding helpers, and search products. Knowing what the term means helps you ask better questions: What text went in? What patterns might the system have learned? What checks are needed before acting on the output? This matters because a language model can save time on drafting, summarising, and reformatting text, but it can also produce confident mistakes, biased wording, or fabricated details. For managers, educators, public servants, and everyday users, the key decision is not whether the text sounds good, but whether it is suitable, verified, and safe for the task.
How a language model turns a prompt into text
After reading this diagram, the reader should be able to say that a language model reads input text, predicts one next token at a time, and builds an output step by step.
01
1. Input text
The process starts with text such as a question, instruction, or partial sentence.
Draw a rounded rectangle on the left labelled 'Prompt'. Inside, show sample text: 'Summarise this paragraph in two sentences.'
02
2. Split into tokens
The system converts the text into tokens, which are small pieces such as words, parts of words, or punctuation.
Draw a second box labelled 'Tokens'. Show the sample prompt broken into small pill-shaped units.
03
3. Predict the next token
Using patterns learned during training, the model estimates which token is likely to come next in this context.
Draw a third box labelled 'Prediction'. Show three candidate token pills with example percentages, with one highlighted as selected.
04
4. Repeat to build output
The chosen token is added, then the model predicts again, repeating until it stops.
Draw a final wide box labelled 'Growing output'. Show text appearing token by token, ending with a note beneath: 'Fluent does not always mean correct.'
A left-to-right flow diagram with four boxes linked by arrows. Read from the user's prompt on the left, through tokenisation and next-token prediction in the middle, to the growing output on the right. A small note under the final box reminds the reader that fluent text can still be wrong.
Hovering or tapping each box reveals a short tooltip with one sentence of explanation and highlights the connected arrow path.
Describe this without the diagram
The diagram explains a four-stage process. First, a user provides text such as a question or instruction. Second, the system breaks that text into smaller units called tokens. Third, the model uses learned patterns from training to estimate likely next tokens and selects one. Fourth, it adds that token to the output and repeats the prediction step many times until a full reply appears. The main message is that a language model generates text step by step by predicting likely next tokens, and polished wording does not guarantee correctness.
With reduced motion: Show the same four boxes in a fixed left-to-right layout with static arrows. The final box displays a completed example output and the note 'Fluent does not always mean correct.' No animation or progressive reveal is used.
A common text-generating language model takes in text and produces text. The input might be a question, an instruction, part of a sentence, or a longer conversation. Inside the system, text is broken into small units called tokens, which may be whole words, parts of words, or punctuation marks.
The model is trained by adjusting numerical settings, often called parameters, so that it becomes good at predicting what token is likely to come next after earlier tokens. It does not search a rulebook sentence by sentence. It uses patterns learned from examples of language.
When you give it a prompt, it processes the tokenised prompt, estimates probabilities for possible next tokens, selects or samples one, adds it to the text, and repeats. This step-by-step generation can produce a sentence, a summary, computer code, or a longer reply.
This is why this kind of language model can sound fluent even when it is wrong. Fluent output is not proof of factual accuracy, real-world access, or reliable reasoning. A language model is not the same thing as a chatbot interface, a search engine, or a person.
How “Finishing a message from familiar writing patterns” maps onto the real thing
| In the story | In the AI system |
|---|---|
| The unfinished maintenance note | The prompt given to the language model |
| Guessing the most likely continuation from many past notes | Next-token prediction based on patterns learned from training text |
| Adding the message one piece at a time | Generating output token by token |
| A plausible but possibly wrong explanation of the pump issue | Fluent output that may still contain errors or hallucinations |
A worked example
Before clearance, a reporting officer asks an approved writing assistant to simplify a sentence in a progress report.
Starting from: Rewrite in plain language: Implementation was deferred pending completion of the budget review.
- The model splits the prompt into tokens and represents them in context.
- It calculates a probability distribution for the next token, taking into account the instruction to use plain language.
- It selects or samples a likely token, appends it to the sequence, then calculates probabilities again using the expanded context.
- It repeats this process until it reaches a stopping condition, producing a sentence that is likely in context rather than one retrieved as a fixed answer.
Result: Possible output: "Work will begin after the budget review is complete."
In a United Nations workplace
At a lightly staffed field presence where the connection regularly drops, a protection officer must screen a long list of applications against a roster profile by close of business. They consider using an approved language-model tool to draft entries for the screening matrix.
The model can turn narrative application text into concise draft entries by generating likely wording from the prompt and context. It does not verify qualifications or apply criteria reliably, so each entry must be checked against the application and roster profile.
Before you try thisApplications contain personal and potentially sensitive employment information. They must not be entered into an external or unapproved tool. Even with an approved tool, its use must be authorised for this data. The model must not decide whom to shortlist: plausible summaries can omit evidence, introduce assumptions or reproduce bias.
Where you may hear it
- During a vendor demonstration to an ICT governance board
- In a procurement meeting reviewing an AI writing assistant
- At an inter-agency working group discussing approved uses of generative AI
- During staff training on drafting and summarising with AI
Avoid confusion
Avoid confusion
Often confused with
Large language model
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.
Chatbot
A chatbot is a system that takes messages in ordinary language and returns replies in a back-and-forth conversation.
"Understanding" (what models do and do not do)
In language models, apparent comprehension usually means producing text that fits patterns in data and context, not reliably building a grounded grasp of the world like a person does.
Common misconceptions
Common misconception: A language model is the same thing as a chatbot.
Correction: A chatbot is usually an application or interface for conversation. A language model is the text-generating component that may sit inside it. Some chatbots use language models, but the terms are not identical.
Common misconception: If the reply sounds natural, the model must understand the topic like a person.
Correction: Natural wording mainly shows that the model is good at producing likely text patterns. That can overlap with useful behaviour, but it does not by itself prove grounded understanding, fact checking, or reliable judgement.
Common misconception: A language model looks up each answer in a database of exact sentences.
Correction: It may be connected to tools that retrieve information, but the model itself typically generates output by predicting tokens from learned statistical patterns. That is why it can produce novel text, and also why it can invent details.
Quick check
Which statement best describes a language model?
A language model works by estimating likely next tokens from the text context and repeating that process. It may be connected to retrieval or internet tools, but those are separate components. Fluent language does not guarantee truth.
Go deeper
Go deeper
The simple version
A language model is a system that has seen a great deal of text and learned to continue text in likely ways. Give it some words, and it generates more words one piece at a time.
The technical version
A language model maps a sequence of input tokens to a probability distribution over possible next tokens. During training, its parameters are adjusted to reduce prediction error across many text examples. At use time, it tokenises the prompt, computes probabilities for candidate next tokens given the current context, selects or samples one, appends it, and repeats until a stopping condition is met. Many modern examples are neural networks, including transformer-based models. The term 'language model' is broader than 'large language model': a model can be a language model without being especially large, conversational, or multimodal.
Reviewed external resources
EncyclopaediaLanguage model (opens in a new tab)Wikipedia
This Wikipedia link gives a broad overview of the term and suits readers who want a quick general reference.
Intermediatelink checked 2026-07-31
CourseLLM Course (opens in a new tab)Hugging Face
This Hugging Face course link suits readers who want a more structured and technical introduction to how these systems are built and used.
Deeperlink checked 2026-07-31
VideoLarge language models, explained visually (opens in a new tab)3Blue1Brown
This 3Blue1Brown video suits visual learners who want an intuitive explanation of large language models through diagrams and animation.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Language model (opens in a new tab)YouTube search
This YouTube search results page helps readers browse multiple video explanations of language models and suits people who prefer to compare different teaching styles.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Language model (opens in a new tab)Reddit search
This Reddit search results page helps readers find community discussions and questions about language models and suits readers who want informal explanations and debate.
search results, not one pagelink checked 2026-07-31
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 language model generates text one token at a time from probabilities conditioned on context. Plausible wording is not evidence that the output is accurate.