What is generative AI?
Patterns turned into drafts.

A model learns patterns, then produces a new sample.

NIST describes generative AI as a class of models that learns the structure and characteristics of input data and uses them to create derived content. That content can include text, images, audio, video, and other digital material. The term describes a kind of model and task; it does not name one architecture, product, or level of capability.

A classifier might label a picture as a cat. A generative system might write a description of the picture or produce a new image from a prompt. Some AI products combine both kinds of work. The distinction is about what the system is asked to produce, not a ranking of intelligence.

“Synthetic” means that a model generated the content. It does not guarantee that the result is accurate, original, harmless, or free of rights concerns. For the larger map of AI terms, start with what artificial intelligence means ↗.

The prompt steers generation; it does not rewrite the model.

During training, a model adjusts its parameters using examples and a learning objective. At use time, the prompt and any supplied context influence the output the model produces from those learned parameters. A prompt changes the current request; it does not automatically retrain the model or give it new facts.

For many language models, generation proceeds as a sequence of token predictions. The model repeatedly produces the next part of an answer until it reaches a stopping condition. Sampling settings and the model’s learned probabilities can make wording vary between runs. See how language models generate text ↗ and what tokens and context windows do ↗.

Other generative models use different representations and procedures. An image model may refine a noisy signal over several steps, while another model can generate a sequence of image tokens. The word “generative” alone does not tell you how a system works internally.

Content type and generation method are separate choices.

These are useful examples, not an exhaustive list. Generative adversarial networks, autoencoders, diffusion models, and autoregressive models learn and produce data in different ways. Read more about how image generation works ↗ and what multimodal AI means ↗.

The surrounding software determines what the model can use and do.

A generative application may add instructions, retrieve documents, filter or format an answer, call tools, store conversation history, and ask a person to review a result. These are system features. A model’s ability to generate text does not by itself give it permission to send an email, alter a file, or look up current information.

Retrieval can supply selected source material for an answer; tool calling can pass a structured request to an application. Each step needs its own checks. Explore retrieval-augmented generation ↗ and tool calling ↗.

Use generation where a draft or variation is useful.

Generation may be the wrong tool when a task needs an exact lookup, a guaranteed calculation, or a decision that must follow fixed rules. Search, databases, ordinary software, and human expertise can be better fits. Compare the full workflow, including review and correction, rather than judging a polished sample.

A fluent result is not evidence that the result is true.

A model can produce a plausible answer without having retrieved or checked a source. It may omit important context, reflect gaps or patterns in its training data, or change its response when the prompt changes. NIST’s Generative AI Profile discusses risks across system design, use, and evaluation; the relevant checks depend on the product and people affected.

Generated material also does not establish authorship, permission, or legal status. Keep source and review records where they matter, and read what AI output means for copyright and authorship ↗. For methods to test whether an image is AI-generated, see the image verification guide ↗.

1. Define the job. Write down what a useful, acceptable result must do.
2. Check representative cases. Include common requests, edge cases, missing information, and harmful or misleading inputs.
3. Verify against evidence. Use tests, source documents, calculations, or expert review that match the task.
4. Control the workflow. Protect private inputs, limit tool permissions, and keep a way to correct or reverse consequential actions.

Follow one part of the system further.

For a broader map, read what AI is ↗. For specific methods, continue with image generation ↗ or language models ↗. For evidence and trust, explore how to read AI benchmarks ↗.