What is generative AI?

Generative AI is AI that makes things. Give it a prompt and it can write an email, draw an image, compose music, generate video or write computer code. It does this by learning the patterns in enormous amounts of existing data and then producing new material that follows those patterns.

How it works

• Learn from huge datasets. Models are trained by repeatedly predicting the next piece of a sequence, like the next word or part of an image. The result is called a foundation model; for text, it's usually a large language model.
• Transformers. Introduced in 2017, this neural network design tracks context with a mechanism called attention, and it powers tools like ChatGPT.
• Diffusion models. Most image generators learn to turn random noise into a picture, step by step.
• Refine. Models are then fine-tuned, including with human feedback, to be more helpful and safer.

Older machine learning mostly made predictions about existing data, like spotting a tumor in an X-ray. Generative models create something new.

What it can make

• Text and code: chatbots such as ChatGPT, Claude and Gemini
• Images: tools such as DALL-E, Midjourney and Stable Diffusion
• Audio, music and video
• Science: research systems that design new proteins and materials

The risks

• Made-up facts. Models can state false things confidently, known as hallucinations.
• Copyright and misinformation. NIST warns it makes copying protected work and producing misinformation easier, and deepfakes can fool people about what's real.
• Energy and water. MIT News reports that training GPT-3 used about 1,287 megawatt-hours of electricity, and data centers use water for cooling.

A quick history

IBM traces the roots to the ELIZA chatbot in 1964. The key building blocks arrived in 2014 (image generators) and 2017 (transformers), and ChatGPT's launch in 2022 made generative AI a household term.