Generative AI

Generative AI uses learned patterns to produce content such as text, images, audio, or code. Inputs can guide that production: a request to summarize a document, a description of an image, or a passage to read aloud. Generation describes what the system produces, not a guarantee that its output is true or original.

Different tasks can share one application

Consider a shop answering a return question. Retrieval finds the applicable policy, classification labels the request, and generation drafts a reply. A language model may participate in all three, but each task has different acceptance criteria. The draft still needs to preserve the policy's conditions.

AI is the broader field. Large language models are models for language, while generative AI also includes systems for other media. The two terms are not interchangeable. A system can use prediction internally to generate content, so prediction and generation are not opposites at every level.

A model output is a proposal to evaluate

A generated illustration may be useful even when it depicts an imaginary product. A customer-facing policy answer must match real evidence. Set the criterion according to the task rather than judging only fluency or realism.

The surrounding application supplies context, controls access, executes permitted tools, and checks outcomes. NIST's Generative AI Profile discusses risks associated with these systems. Generating “your return is submitted” does not establish that a submission occurred; the application needs a successful execution result.

Reference: NIST: Generative AI Profile.