What Is Generative AI?

What is being generated?

Different model families generate in different ways. An autoregressive language model produces a sequence by predicting successive tokens. Diffusion models offer another approach: learning to reverse a process that adds noise, then using that learned process to generate samples. Generative AI is broader than either method.

In a writing task, a prompt and the available context influence what comes next. In an image task, a description may guide the visual output. These are ways of steering generation, not detailed specifications that every result will satisfy.

Evidence & context: Brown and colleagues, 2020 · Ho, Jain and Abbeel, 2020

Generating an answer is not the same as finding a source

A search system can return an existing document. A generative system can compose a new sentence. A product may do both, but the activities remain distinct: finding a document does not prove that the generated sentence represents it accurately.

If your question concerns a deadline, quotation or eligibility rule, you need a source you can inspect. Read how to verify AI-generated information before treating a polished response as evidence.

Example: designing a learning activity

Suppose you are preparing a discussion about water use. You could ask a model to propose three classroom activities using household objects. Treat those suggestions as draft designs: check materials, age suitability, local constraints and what each activity would actually teach.

Then ask which observation would distinguish two competing explanations. This moves the task beyond ‘produce an attractive worksheet’ towards ‘help a learner investigate an idea’. The educational purpose remains yours to define.

Original-looking is not the same as original thinking

Generative output can be inaccurate, biased or inappropriate. NIST treats these as risks to manage, rather than problems solved merely by producing fluent content.

For a creative task, ask whether the result meets your intention and respects the people it represents. For a factual task, check its claims. For a learning task, ask what intellectual work the learner is still doing. The same tool can be useful for one purpose and counterproductive for another.

Evidence & context: NIST

Sources & further reading

1. Language Models are Few-Shot Learners

   Brown and colleagues, 2020. Primary research on an autoregressive language model and learning from examples in context; not a guarantee about all models.

2. Denoising Diffusion Probabilistic Models

   Ho, Jain and Abbeel, 2020. An example of a generative modelling approach beyond language-model token prediction.

3. Generative Artificial Intelligence Profile (NIST AI 600-1)

   NIST. Risk-management guidance, including confabulation. It does not establish a universal error rate.

Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.