---
title: 'Generative AI'
description: 'AI whose central capability is producing new text, code, images, audio, video, or structured data from a learned generative model.'
canonical_url: 'https://darkfactory.dev/glossary/generative-ai'
markdown_url: 'https://darkfactory.dev/glossary/generative-ai.md'
collection: glossary
date_published: '2026-08-03T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Generative AI


## Definition

Generative AI is AI whose central capability is producing new text, code, images, audio, video, or structured data from a learned generative model. The system generates an output conditioned on a prompt, example, control signal, or other input, or samples without an external condition.

"New" means the system constructs an output for the request. It does not prove originality, novelty, ownership, or factual accuracy. A generated passage can repeat memorized material, combine familiar patterns, or state a plausible falsehood. Those questions require separate provenance, similarity, and correctness checks.

The current label covers several technical families, including autoregressive language models, diffusion models, variational autoencoders, and generative adversarial networks. Generative modeling predates the recent public use of "generative AI," and the umbrella phrase does not identify one architecture or inventor.

## Distinguish it from nearby terms

- **Discriminative AI** estimates labels, scores, or boundaries from inputs rather than modeling how to generate samples. One system can contain both kinds of model.
- **Predictive AI** forecasts an outcome. A generative model also predicts during generation, but the product distinction concerns whether the system returns content or a bounded prediction.
- An **agent** selects actions and uses tools toward a goal. A generative model can power an agent without being agentic by itself.
- A **foundation model** is defined by broad reuse. It may be generative, but generative output alone does not make a model foundational.

## Operational significance

Evaluation must inspect the produced artifact, not only whether the model returned one. Define factuality, safety, style, rights, privacy, and format requirements for the use case. When output can trigger tools or publication, add validation and authorization before the generated content becomes an action.

## Check your understanding

A system drafts an email but cannot send it. It is generative AI. Adding permission to choose recipients and call a mail tool changes the surrounding system into an agentic workflow, not the underlying definition of generation.

## Also called

GenAI, generative artificial intelligence

## Related terms

- [Causal language model](https://darkfactory.dev/glossary/causal-language-model)
- [Diffusion model](https://darkfactory.dev/glossary/diffusion-model)
- [Generative adversarial network (GAN)](https://darkfactory.dev/glossary/generative-adversarial-network)

## Evidence and further reading

- [Google Machine Learning Glossary](https://developers.google.com/machine-learning/glossary/)
- [MIT Sloan Generative AI Basics Glossary](https://mitsloanedtech.mit.edu/ai/basics/glossary/)
- [Deep Learning](https://www.deeplearningbook.org/)