What is Generative AI?

How it works

Traditional machine learning mostly makes predictions about existing data: is this email spam, what will sales be next month. Generative models learn the structure of their training data well enough to produce new examples that resemble it.

The main families are:

• Large language models for text and code. They are transformers trained to predict the next token, and generate by doing that repeatedly. ChatGPT, Claude and Gemini are built on them.
• Diffusion models for images, video and audio. They learn to turn random noise into a picture step by step, guided by a text prompt.
• Multimodal models that take and produce several types, such as reading an image and answering in text.

All of them are trained on very large datasets, then usually tuned with human feedback (RLHF) and safety training so they follow instructions.

Why it matters

Generative AI made AI usable through plain language. Anyone can ask for a draft, a summary, code or an image without training a model. That is why it spread into search, office software, customer service and software development so quickly.

The risks are equally general. NIST's Generative AI Profile lists issues such as confabulation (confident false output), harmful bias, data privacy, information integrity and intellectual property. In practice, generative AI is a strong first-draft engine that still needs checking, grounding in real data (RAG) and clear limits.

Generative AI produces content when asked. Agentic AI builds on it by letting models plan and take actions with tools over several steps.

Example

A marketing team asks a language model for ten subject lines for a product launch email, then uses an image model to create a matching banner from a text description. Both outputs are new content, generated in seconds, that a person then reviews and edits.