What is generative AI?

Generative versus discriminative models

Machine-learning models fall into two broad kinds. A discriminative model learns to tell categories apart — for instance, whether an image shows a cat or a dog. A generative model instead learns the underlying distribution of the data well enough to produce new examples — a plausible new cat image, or a new sentence. Generative AI is the applied use of generative models to create content, sampling from the learned distribution to produce outputs that resemble, but are not copies of, the training data.

How generative models work

Different media use different generative architectures. Large language models generate text by predicting tokens in sequence; generative adversarial networks and, more recently, diffusion models generate images by learning to reverse a noising process.

Many systems are foundation models: large models pre-trained on broad data and then adapted to many downstream tasks. This reuse is a defining feature of the current generation of generative AI.

Capabilities and limitations

Generative AI can draft text, synthesise images, write code, and produce audio, often from a natural-language prompt. Its limitations mirror those of the underlying models: outputs are probabilistic and may be inaccurate or fabricated, can reflect biases in training data, and raise questions about provenance and the rights to training content. Outputs are not authoritative and, in research contexts, must be verified before use.

Generative AI in research

In research, generative models are studied for their capabilities and risks and used cautiously as tools — for example to draft text or generate synthetic data. Synthetic data can help where real data is scarce or sensitive, but it must be validated so that it does not introduce artefacts. Reproducibility is challenging because outputs are stochastic and models change; documenting the model, version, prompt, and settings is essential, as is disclosing where generative AI was used.