Generative AI

Generative AI refers to AI systems whose primary output is novel content — text, images, audio, video, code, or structured data — produced by sampling from a distribution learned during training, rather than a classification, score, or decision applied to existing input. Large language models producing prose, diffusion models producing images, and code-completion models are all generative; a spam classifier or a diagnostic-imaging triage model is not.

How it differs from its siblings on this site

Generative AI is the broad category; several more specific terms sit underneath it and are easy to blur together. Hallucination is a specific failure mode generative systems exhibit (confident, fabricated output), not a synonym for generative AI itself — a system can be generative without hallucinating in a given instance, and non-generative systems don’t “hallucinate” in the same sense (they misclassify). Prompt engineering and system prompt describe how a user or developer steers a generative system’s output, not a property of the system itself. Synthetic data and synthetic image describe a generative system’s output when that output is used to stand in for real-world data or imagery, a specific downstream use case rather than the technology category.

Why the generative/discriminative distinction matters for governance

The disclosure-relevant distinction is that generative AI produces artefacts that can be mistaken for human-authored work, which is what creates the authorship, plagiarism, and citation-integrity questions covered elsewhere in this dictionary (see AI tool disclosure and generative-AI disclosure statement). Discriminative AI (for example, a model classifying tumours from imaging features) raises different governance concerns — bias, validation, clinical accuracy — but not authorship or fabricated-content questions, because its output is a decision over real input rather than newly synthesised content.

Worked examples

ChatGPT producing a paragraph of prose, DALL·E producing an image from a text prompt, and GitHub Copilot producing a code suggestion are all generative AI in the operational sense used here. A logistic-regression model classifying tumours from imaging features is discriminative, not generative — it selects among existing categories rather than synthesising new content.

Sources

NIST AI Risk Management Framework Generative AI Profile (2024); Regulation (EU) 2024/1689 (EU AI Act), which does not use “generative AI” as a formal risk-tier category but addresses it through Article 50 transparency duties and the GPAI regime.