What is generative AI?

Maybe you've noticed a new button in your email app, with a label like "Help me write." Type a few words about what you want to say, and a whole draft appears. Or someone in the family group chat shares a picture of a dog in a spacesuit, and no camera ever took it. Both came from the same kind of software. It's called generative artificial intelligence (AI), and it's the kind most people now mean when they talk about AI at all.

A lot of the AI you've used for years picks an answer from a set of choices. Is this purchase fraud? Which song might you like next? Generative AI does something else. It produces something that wasn't there before: a paragraph, a picture, a melody, a few lines of computer code. That's all the word means. It generates.¹²

So how does it make something? During training, it works through enormous amounts of existing material, such as text, images, or sound, and learns the patterns in it: which words tend to follow which, what a dog usually looks like, how a sentence tends to end. When you ask for something, it builds the answer by producing what's likely to come next, piece after piece, based on those patterns. Put plainly, generative AI is AI that makes new things, such as text, pictures, sound, or code, from patterns it learned in a great many examples.¹³

Text is the best known, from chatbots that draft emails to tools that summarize a long document. But the same idea now makes images from a description, voices and music, short videos, and computer code. Some tools mix them, so you can describe a picture in words, or ask a question about a photo you've taken.¹

Generative AI isn't brand new. Researchers have built models that generate things for decades. What changed in the last several years was scale and design. Models got much bigger, they were trained on far more data, and a design called the transformer, described by a team of Google researchers in 2017, made it practical to learn from huge amounts of text at once. Together, those made the results good enough to put in everyday products.¹⁴

Something newly made can look completely right and still be wrong. A generated paragraph can include a fact that's slightly off, or a source that doesn't exist, written in the same confident tone as everything around it. The tool is producing what's likely, and likely is a different thing from true. So the parts that matter, such as names, numbers, dates, and quotes, are worth a quick check before you use them.⁵

Where it helps most is the first draft. A starting point for the email you've been putting off. Ten ideas for a birthday party, so you can pick the two you like. A summary of a long report, so you know which pages to read closely. In each case a person stays in charge: you decide what's good, check what matters, and make it yours. Next time that button appears, try it on something small, and read what comes back the way you'd read a helpful friend's first draft.