Learn AI /Generative Media /Text generation
Text generation
- Starter
- Generative Media
- Communications teams
Also called: drafting, writing with AI.
Back to the milestone you came from
In five seconds
What it is
A system produces new written words from a prompt, patterning its output on examples seen during training.
In plain words
You give the system instructions, and it drafts text that seems to fit. It can be useful for first versions, but people still need to check facts, tone and safety.
An everyday example
A community centre manager in Dakar gives a communications assistant four bullet points about a vaccination event: date, place, opening hours and what people should bring. The assistant turns those notes into a clear public notice with a headline, short introduction and practical details. If the manager asks for a friendlier tone or a version for parents, the assistant rewrites it. If one bullet point is missing or unclear, the notice may still sound polished but include a wrong assumption.
A human assistant can ask follow-up questions, notice contradictions from real-world experience and take responsibility for mistakes. A text generation system does not have that kind of grounded judgement. It produces text by patterning from training data and the prompt, so it may sound confident without having verified the facts.
Remember thisText generation can produce useful drafts quickly, but fluent writing is not the same as verified, safe or appropriate content.
Understand it
Understand it
Why it matters
Non-specialists should care because generated text can save time but also change how decisions are made. A fast draft may help with routine writing, accessibility and multilingual communication support. But the same fluency can hide mistakes, invented citations, unsafe wording or misleading claims. If you rely on it without review, you may publish errors, breach trust, mishandle sensitive topics or give the impression that text came from a person when it did not. Understanding text generation helps you ask better questions, set review rules, protect sensitive information and decide when AI drafting is helpful and when it is risky.
How text generation turns a prompt into a draft
After viewing this, a reader should be able to say that text generation takes a prompt, predicts likely next text step by step, and produces a draft that must be reviewed.
01
1. Prompt enters
The user provides a request, such as bullet points, a question, a target audience or a writing style.
Draw a box labelled 'Prompt' with example chips: 'write an email', '150 words', 'for the public', 'use plain language'.
02
2. Context shapes the task
The wording of the prompt tells the system what kind of text to produce, how long it should be and what tone to use.
Draw a second box labelled 'Instructions and context' with tags for tone, format, length and audience.
03
3. The model predicts text
The system generates the output by predicting likely next word pieces repeatedly until it reaches a stopping point.
Draw a larger middle box labelled 'Prediction loop' with a short sequence such as 'Dear' → 'colleagues' → ',' to show step-by-step generation.
04
4. Draft comes out
The result may be useful and well-structured, but it can still contain errors, bias or invented details.
Draw an output box labelled 'Draft text' with a yellow warning icon and note: 'fluent does not mean verified'.
05
5. Human review
A person checks facts, tone, risks and suitability before sharing or relying on the text.
Draw a final box labelled 'Review and edit' with checklist items: facts, sources, sensitivity, audience, approval.
A left-to-right flow diagram with four boxes connected by arrows. Read from the user prompt, to instructions and context, to word-by-word prediction inside the model, to the draft output and human review. Small note labels show that better prompts shape the result and that fluent text is not automatically accurate.
Hovering over each box reveals one short example of the same task rewritten with different prompt quality, from vague to specific.
Describe this without the diagram
This diagram explains text generation as a five-stage flow. First, a user gives a prompt with instructions such as the topic, audience, length and tone. Second, those instructions shape the task. Third, inside the model, text is produced by repeatedly predicting likely next word pieces. Fourth, the system outputs a draft. A warning notes that smooth writing can still be inaccurate or invented. Fifth, a human reviews and edits the draft for facts, risks, tone and approval. The key point is that prompt quality affects the draft, but human checking remains necessary.
With reduced motion: A static left-to-right diagram shows the same five boxes and arrows without animation. The prediction loop is shown as a simple labelled box rather than moving text.
Text generation means using an AI system to produce new text after you give it a prompt. The prompt can be a question, instructions, a few bullet points, or existing text to continue. The output might be an email draft, a summary, a paragraph, a headline, or a longer report.
In simple terms, the system looks at your prompt and predicts what words or pieces of words are likely to come next, one step at a time. A “model” here means a trained mathematical system that has found patterns in large collections of text. It does not check truth by itself, and it may produce fluent wording that is wrong, biased or invented.
What goes in matters a lot. If the prompt is vague, the result is often vague. If the prompt asks for a style, audience, length or structure, the output is more likely to fit. People often use text generation for drafting, brainstorming, rewriting, changing tone or producing first versions faster.
Text generation is often confused with summarisation and machine translation. Summarisation aims to shorten source text while keeping its main points. Machine translation aims to render meaning from one language into another. Text generation is broader: it can do those tasks, but it can also create brand new wording that was not in the source. That is useful, but it also means human review is essential.
How “Drafting a notice from bullet points” maps onto the real thing
| In the story | In the AI system |
|---|---|
| bullet points from the manager | the prompt or input instructions |
| communications assistant drafting full sentences | the model generating text from patterns |
| asking for a friendlier version | prompting for tone, style or audience |
| a polished notice with a possible wrong assumption | fluent output that still needs fact-checking |
A worked example
A programme officer in a regional bureau uses an approved drafting tool to prepare a paragraph for a monitoring report before close of business.
Starting from: Prompt: 'Using only these cleared points, draft a short paragraph for the monitoring report: the planned workshop was completed; the implementing partner submitted the attendance record; procurement remains pending technical review. Do not add causes, dates or figures. Use a neutral tone.'
- The system divides the prompt into tokens and converts them into internal numerical representations.
- For each position, the model estimates probabilities for possible next tokens. A decoding method selects one, then repeats the calculation using the growing sequence.
- The attendance record itself is not provided to the tool because it contains personal data and is unnecessary for this drafting task.
- The programme officer checks each generated claim against the cleared points and removes any unsupported explanation before sending the paragraph for clearance.
Result: Draft output: 'The planned workshop has been completed, and the implementing partner has submitted the attendance record. Procurement remains pending technical review. These points reflect the cleared information available for the current monitoring report; no causes, dates or figures have been added.'
In a United Nations workplace
Before close of business, a records officer supporting an inter-agency secretariat uses an approved writing tool to turn cleared, non-sensitive instructions into a draft note for the file recording changes to the concept note for a side event in the margins of the session.
The tool generates sentences one token at a time from the prompt and patterns learned during training. It can produce a serviceable first draft quickly, but it does not know which email is authoritative or why particular wording was cleared.
Before you try thisSource correspondence, personal data and restricted material must not be entered into a tool unless it is approved for that material and use. The records officer must compare the draft with the cleared instructions. Generated text cannot replace the authoritative record and may invent a rationale or overlook a qualification.
Where you may hear it
- During a vendor demo to an ICT governance board
- At front-office clearance of talking points before close of business
- While a country office tests an approved writing assistant for progress reports
- In a procurement meeting reviewing tools for drafting routine correspondence
Avoid confusion
Avoid confusion
Often confused with
Summarisation
Summarisation produces a shorter version of text or speech that keeps the main points and drops less important detail.
Machine translation
Machine translation uses computer systems to convert text from one language into another using learned patterns, statistical methods or hand-built rules.
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.
Common misconceptions
Common misconception: If the writing sounds confident and polished, it is probably correct.
Correction: Fluent wording mainly shows that the system is good at producing likely language patterns. It does not guarantee that names, dates, quotations, references or claims were checked against reliable evidence.
Common misconception: Text generation just copies and pastes from its training data.
Correction: These systems usually generate new combinations of words rather than retrieving one stored passage. That is why the output can be original in form but still reflect patterns, biases or occasional memorised fragments from training data.
Common misconception: It is the same as summarisation.
Correction: Summarisation tries to shorten source material while preserving its main points. Text generation is broader and can create fresh wording, new structure or extra detail not present in the source, which increases both usefulness and risk.
Quick check
Which statement best describes text generation?
Text generation creates new wording from patterns learned during training and the instructions in the prompt. A specific prompt can improve usefulness, but it does not guarantee truth or safety. Translation and summarisation are narrower tasks that may use text generation but are not identical to it.
Go deeper
Go deeper
The simple version
You give the system instructions, and it drafts text that seems to fit. It can be useful for first versions, but people still need to check facts, tone and safety.
The technical version
In many current systems, text generation is done by a large language model that has been trained on large text collections to predict the next token, where a token is a word or part of a word. At use time, the model turns the prompt into internal numerical representations and then repeatedly estimates probabilities for possible next tokens. A decoding method selects one token at a time, and the growing sequence becomes the output. Settings such as temperature and maximum length can affect how predictable, varied or long the text is. The model may also be guided by system instructions, examples or retrieved reference material. This mechanism is different from classic rule-based natural language generation, which used hand-written templates and grammar rules. It is also different from summarisation and machine translation as task labels: those are narrower tasks, while text generation is the broader capability of producing text.
Reviewed external resources
Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This link gives UNESCO's ethical framework for AI use and suits readers who need policy-oriented guidance on responsible use, rights and human oversight.
Intermediatelink checked 2026-07-31
Primary sourceContent Credentials (opens in a new tab)Content Authenticity Initiative
This link explains Content Credentials and suits readers who want to understand how provenance information can help indicate how digital content was created or edited.
Starterlink checked 2026-07-31
EncyclopaediaNatural language generation (opens in a new tab)Wikipedia
This link gives a broad introductory overview of natural language generation and suits readers who want background terminology and a starting point for further reading.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Text generation (opens in a new tab)YouTube search
This link is a YouTube search results page, so it helps readers browse different video explainers on text generation and suits people who prefer audiovisual introductions.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Text generation (opens in a new tab)Reddit search
This link is a Reddit search results page, so it helps readers sample community questions and experiences about text generation and suits readers who want informal discussion alongside more formal sources.
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
Text generation can produce useful drafts quickly, but fluent writing is not the same as verified, safe or appropriate content.