What is generative AI? Definition, examples and limits

Generative AI is a form of artificial intelligence that creates new content, whether text, images, sound, video or code, from a simple written or spoken request. ChatGPT, Claude, Gemini and image generators are the best-known examples: they are very handy for producing a quick first draft, but they still make mistakes you need to learn to spot.

Definition: AI that makes things instead of sorting them

For a long time, artificial intelligence was mainly used to recognize or predict: flag an email as spam, read a ZIP code, recommend a movie. Generative AI does something else: it produces a result that didn’t exist before you asked. You type « write a letter canceling my broadband contract », and a complete text appears, slightly different every time.

That content is generally not copied word for word from a database. It is built, piece by piece, from what the model learned during training. That is what makes it flexible and fast, and also what explains its mistakes: it produces plausible output, not verified output.

An everyday example shows the difference. Your bank has used AI systems for years to analyze your card payments and spot likely fraud: they answer yes or no and create nothing. A generative AI assistant, on the other hand, can write the letter disputing that suspicious payment for you. Same broad technical family, very different results.

To place generative AI within the wider field, our guide to what artificial intelligence is covers the main families and their history.

What generative AI can create

The term covers very different tools. Here are the main categories, with the technique that dominates each one.

Type of content	Example uses	Main technique
Text	Email, summary, translation, meeting notes	Large language models
Code	Spreadsheet formula, small script, bug fix	Large language models
Images	Illustration, poster, photo editing	Diffusion models
Voice and music	Reading a text aloud, background music	Specialized audio models
Video	Short clip from a description	Variants of diffusion models

The boundaries keep shifting: more and more assistants combine several of these abilities in a single conversation. You show a photo, ask a question out loud, and get back text and an image. This is called multimodal AI, which we cover in multimodal AI: when AI sees, hears and speaks.

How it works, by type of content

Text: predicting what comes next, word by word

Assistants such as ChatGPT, Claude or Mistral AI’s Le Chat (renamed Mistral Vibe in 2026) rely on a large language model. It was trained on vast amounts of text, its training data, to predict the most likely word fragment after what came before. By repeating that prediction hundreds of times, it writes a whole answer. The mechanism is explained step by step in how ChatGPT works.

Images: starting from noise and cleaning it up

For images, the most widespread technique is the diffusion model. The principle is surprising: the model is taught to gradually remove « noise » (a kind of random static) from an image. Once trained, it starts from pure noise and cleans it up step by step, guided by your description, until a coherent image emerges.

Voice, music and video

The same ideas show up elsewhere. For voice, models learn to turn text into natural speech, with its intonation and pauses. Video means producing a sequence of images that stay consistent with one another, which is much harder: an object can change shape from one shot to the next, or a hand can grow an extra finger. These glitches are becoming rarer with each generation of tools, but they remain a useful clue for spotting generated content.

The general public discovered text-based generative AI with the launch of ChatGPT on November 30, 2022. Since then, these tools have made their way into messaging apps, office suites and phones.

A very fast rise, in numbers

Public data show both a surge in investment and already broad adoption, at least in Europe. According to Our World in Data (figures collected on October 22, 2026), global investment in generative AI has evolved as follows:

Year	Global investment in generative AI
2021	$5.3 billion
2022	$3.7 billion
2023	$25.4 billion
2024	$43.7 billion
2025	$143.8 billion

On the usage side, the share of people aged 16 to 74 who had used a generative AI tool in 2025 reached, according to Our World in Data based on Eurostat: 47.02% in Switzerland, 44.7% in the Netherlands, 42.01% in Belgium, 37.88% in Spain, 37.46% in France, 32.25% in Germany and 19.86% in Italy. In other words, more than one in three people had already tried these tools in several large European countries. More data series are gathered on our AI in numbers page.

Everyday examples

• At home: plan the week’s meals around what’s left in the fridge, write a polite complaint to a delivery company, invent a bedtime story featuring your kids’ names.
• At work: turn meeting notes into minutes, suggest three versions of a tricky email, translate a product sheet, write a spreadsheet formula.
• For a club or charity: illustrate a poster, draft an event announcement, summarize long bylaws.
• For paperwork: rephrase a confusing official letter in plain English, list the documents to gather before an appointment.
• For learning: ask for an explanation with an example, then a quiz to check you’ve understood.

In every case, the AI gives you a first draft. It is up to you to reread it, adjust it and check its substance.

The limits to know about

It sometimes makes things up. A generative model can produce false information with total confidence: a hallucination. A date, a quote, a legal reference or a source can be fabricated. Our guide to AI hallucinations explains how to protect yourself.

It reflects its data. What the model learned comes from existing texts and images, with their blind spots and stereotypes. Its output can reproduce them.

It makes fakes easier. The same technology can create deepfakes: doctored photos, voices or videos, used among other things in scams. A very realistic image is not proof.

It raises privacy questions. Depending on each service’s settings, your conversations may be used to improve the models. Don’t paste personal or confidential data without checking those settings first.

Free or paid: what actually changes

Most major generative AI assistants come in a limited free version. According to the site’s survey of the providers’ official pages as of October 22, 2026, ChatGPT, Claude, Gemini and Le Chat all offer capped free access, plus paid subscriptions. The table below sums up what each one offers.

Assistant	Free	Paid from	Web search	Agent mode	Voice	Images
ChatGPT	Limited	$8/mo	Yes	Yes	Yes	Yes
Claude	Limited	$20/mo	Yes	Yes	—	—
Gemini	Limited	€4.99/mo	Yes	—	Yes	Yes
Le Chat (Mistral Vibe)	Limited	$5.99/mo	Yes	Yes	Yes	Yes

Free-tier caps usually apply to the number of messages, generated images or uploaded files, and to access to the most powerful models. For occasional use, an email here and a summary there, the free version is often enough. If you use it every day for work, compare the plans with your real usage rather than with the marketing pitch.

Four habits for using it well

1. Be specific. Give the context, goal, audience and expected format. The site’s prompt generator helps you structure your request.
2. Let it handle the form, keep control of the substance. Structure, tone, rewording: yes. Facts, figures, decisions: check them yourself.
3. Ask for sources when the information matters, and open them.
4. Reread before sending any text that carries your name, your company or your money.

Check yourself Does an image generator create a picture by searching the internet for it?

No. A diffusion model starts from random noise and gradually turns it into an image, guided by your description. It draws on what it learned during training, but it does not fetch an existing photo at the moment you make your request.