presentsUNAIVERSE
The UN AI timeline

    Learn AI /AI Foundations /Generative AI

    Generative AI

    Also called: gen AI, GenAI.

    Detail

    In five seconds

    What it is

    A type of AI that produces new content such as text, images, audio or video from patterns learned from many examples.

    In plain words

    Generative AI is software that can make new text, pictures, audio or video after you give it a prompt.

    An everyday example

    Someone who has seen thousands of invitations can quickly draft one when given the occasion, place and tone. The result may look polished because it follows familiar patterns. It may also add a dress code, invent a contact detail or imply that food will be served, even though nobody supplied those facts. Generative AI works in a comparable way: it constructs a plausible new output from learned patterns and the instructions it receives, so the finished draft still needs checking.

    Where the story stops working

    A human clerk may understand the actual shipment, ask clarifying questions and consult current records. A generative model does not reliably do that by itself; it generates from learned patterns, not from direct awareness of the real shipment unless connected to trusted data and tools.

    Remember thisGenerative AI constructs new content from learned patterns and the prompt. Fluent output can still contain unsupported details.

    This term is contested

    The broad idea is widely used, but the exact boundary is not fully settled. Some people use the term narrowly for systems that generate media, while others include a wider range of content-producing models and tools.

    Understand it

    Understand it

    Why it matters

    Non-specialists hear this term in procurement, policy, media and workplace tools. Knowing what it actually does helps you ask better questions about accuracy, risk, cost, privacy, authorship, oversight and whether it is even the right tool.

    How generative AI turns a prompt into content

    After viewing this, a reader should be able to say that generative AI uses learned patterns from training data to generate new content from a prompt, and that the output still needs checking.

    1. 01

      Training examples

      The system is trained on many examples of content, such as text, images or audio.

      Draw a cluster of small file, image and waveform icons on the left with the label 'many examples'.

    2. 02

      Pattern learning

      Training adjusts the model so it captures statistical patterns: which pieces often fit together.

      Draw a central rounded rectangle labelled 'model' with subtle internal dots or nodes; connect the example cluster to it with a thick arrow.

    3. 03

      Prompt

      A user provides an input, such as instructions, questions or reference material.

      Place a prompt card above or before the model with sample text and an arrow into the model.

    4. 04

      Generated content

      The model produces a new output that matches learned patterns, such as text, an image or audio.

      Draw three small output cards on the right for text, image and audio, with a warning icon and label 'check accuracy and appropriateness' beneath them.

    A left-to-right flow diagram with four boxes connected by arrows. Box 1 shows many small icons for documents, images and audio clips labelled 'training examples'. Box 2 shows a larger box labelled 'trained model learns patterns'. Box 3 shows a prompt card entering the model, for example 'Draft a three-line situation summary'. Box 4 shows several possible outputs: a paragraph, an image thumbnail and a speaker icon, followed by a small warning badge labelled 'review needed'.

    Hovering or tapping each stage reveals a short tooltip with one sentence: examples feed training, training shapes the model, the prompt guides generation, and outputs require review.

    Describe this without the diagram

    This diagram explains generative AI as a four-stage flow. First, many examples of content such as documents, images and audio are used in training. Second, a model is trained to capture statistical patterns in those examples. Third, a user gives a prompt, such as a request for a summary or an image. Fourth, the model generates new content based on those patterns. A note at the end says the result can be useful but still needs human checking for accuracy, safety and suitability.

    With reduced motion: A static four-box flow replaces animation. Arrows show the same sequence from training examples to trained model to prompt to generated content, with a final visible note saying review is needed.

    Generative AI takes an input, often called a prompt, and produces content. The input might be a sentence, question, sketch, reference image, audio clip or a mix. The output can be text, pictures, speech, music, code or video.

    Under the surface is a mathematical model trained on examples. Training adjusts internal numbers so the model captures statistical patterns in the examples. When a person gives a prompt, the model uses those patterns to produce an output conditioned on that input. A large language model is one kind of model used to generate or transform text. Generative AI is the broader category and also covers other media.

    This differs from systems used mainly to sort, score or predict labels, such as “fraud risk” or “spam”. Those are often called predictive AI. Generative AI produces content. One system can do both, so the boundary is not always neat.

    The output can be useful, but it can also be wrong, biased, misleading or made-up. It does not verify facts without another step or tool. Human review, source checking and safe use rules matter, especially for public communication or official work.

    How “Drafting an invitation from familiar patterns” maps onto the real thing

    In the storyIn the AI system
    The clerk who has seen many past documentsA trained model built from many examples
    The supervisor's requestThe prompt or input
    The drafted delivery summaryThe generated output
    Checking against real records before sendingHuman verification and fact-checking

    A worked example

    A programme officer in Lima needs a short opening paragraph for a side-event concept note before close of business, based on three cleared points.

    Starting from: Prompt: "Draft a concise opening paragraph using only these points: accessible participation, multilingual materials and a way for participants to provide feedback. Do not add dates, partners or commitments."

    1. The prompt supplies the required themes and limits what the model may add.
    2. The model estimates a likely sequence of words based on patterns learned from concept notes and similar prose.
    3. It generates a fluent paragraph that connects the three points, rather than retrieving a pre-approved paragraph from a database.
    4. The programme officer checks each statement against the cleared points and removes an unsupported claim that the event is annual.

    Result: The revised paragraph is suitable for the first draft of the concept note. The fluent but unsupported reference to an annual event shows why generated text still requires line-by-line review.

    In a United Nations workplace

    The human resources officer supporting an evaluation office has terms of reference that must clear before Friday. Using an approved tool, the officer asks for first drafts of the purpose, deliverables and required competencies based on cleared, non-sensitive bullet points.

    The model constructs plausible sections by estimating likely wording and structure from patterns learned during training, conditioned on the officer's prompt. It can accelerate drafting, but it does not know whether a requirement is authorised, fair or consistent with the file. The officer must check every clause before clearance.

    Before you try thisPersonal data, performance records, applications and procurement-sensitive material must not be entered into a tool that is not approved for them. Disclosure could harm individuals or compromise a selection or contracting process. Use only cleared, non-sensitive inputs and verify the draft against the file.

    Where you may hear it

    • A vendor demo before an ICT governance board
    • A front-office discussion during clearance of a briefing note
    • A procurement unit review of an AI drafting assistant
    • An inter-agency working group considering guidance on external tools

    Avoid confusion

    Avoid confusion

    Often confused with

    Side-by-side

    Common misconceptions

    • Generative AI only means chatbots.

      Chatbots are one use case. Generative systems can also create images, audio, video, code and other content formats.

    • If the output sounds confident, it is probably correct.

      Fluent wording comes from pattern generation, not guaranteed truth. A model can produce convincing errors or invented details.

    • Generative AI always creates something completely original.

      The output is newly generated, but it is shaped by patterns in training data and prompts. That raises questions about similarity, copying and attribution in some cases.

    • Generative AI and predictive AI are the same thing.

      Predictive AI usually estimates a label, score or category, while generative AI produces content. Some systems combine both, which is why people mix them up.

    Quick check

    Which example best matches generative AI?

    Go deeper

    Go deeper

    The simple version

    Generative AI is software that can make new text, pictures, audio or video after you give it a prompt.

    The technical version

    Generative AI uses trained machine-learning models, often deep-learning models, to estimate the probability of content sequences or structures and then sample or construct outputs conditioned on an input prompt or reference signal.

    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

    Generative AI constructs new content from learned patterns and the prompt. Fluent output can still contain unsupported details.