ChatGPT drafting an email, a generator creating an image from a sentence, a tool producing a short video: that's all generative AI. The term is everywhere, but people rarely explain what actually happens inside.

Good news: you don't need math to understand the essentials. And once you get the principle, you'll know why these tools are so good at some tasks, why they get others wrong, and how to use them better.

AI that produces instead of sorting

For a long time, AI was mostly used to recognize or sort things: is this email spam, does this photo contain a cat, which movie should you watch next. It answered with a label or a score.

Generative AI does something else: it produces new content. Text, an image, a sound, a video, code. You give it an instruction (the prompt), and it creates something that didn't exist, based on everything it learned during training.

For text: predicting the next piece

Assistants like ChatGPT, Claude, Gemini or Le Chat run on a large language model. Its basic principle is surprisingly simple: predict the most likely continuation of a text.

The model doesn't read whole words but tokens, pieces of text: a short word, part of a long word, a punctuation mark. At each step, it calculates which tokens are most likely to come next, picks one, adds it to the text, and starts again. A three-paragraph answer is hundreds of predictions chained together.

That sounds limited, and yet this ability turns out to be very broad: to predict the continuation of an explanation, a line of reasoning or a program well, the model had to pick up a huge number of patterns about language and the world. Hence its ability to rephrase, summarize, translate, plan or write code.

But remember the main consequence: the model generates, it doesn't look things up. It produces what's plausible, not what's verified.

For images: starting from noise

Image generators work differently. Most use diffusion models. During training, images are gradually covered in noise (like static on an old TV) until they're unrecognizable, and the model learns to go the other way: remove the noise step by step.

When you ask for an image, the generator starts from random noise and "cleans" it little by little, guided by your text. That's why two attempts with the same prompt never give the same image: the starting noise changes.

It's also why you're better off describing visual facts (subject, setting, framing, lighting) rather than wishes ("a striking image"): the model learned to connect concrete descriptions to shapes. The 5-layer grid for better images is built on exactly this principle. To put it into practice, follow the guide to generate an image with AI, then the method to make a video with AI.

Two different moments: training and use

People often mix them up.

Training happens before you use the tool. The model is exposed to enormous amounts of text or images and adjusts its internal parameters to predict better. It's long, expensive, and done by the company that builds the model. At the end, its knowledge is frozen at a certain date.

Use (also called inference) is when you ask your question. The model learns nothing new at that moment: it applies what it learned to generate an answer. So it doesn't know about news after its training, unless the tool adds web search or documents you provide.

One nuance: some services may use your conversations to train future models, depending on your settings. That's a privacy setting worth checking in your tool.

What generative AI does well

  • Producing a first draft: an email, an outline, a description, an idea for a visual.
  • Rephrasing and adapting: changing the tone, simplifying, shortening, translating.
  • Summarizing and structuring a text you give it.
  • Suggesting variations: ten titles, five angles, three image styles.
  • Explaining a concept at your level, and answering your follow-ups.

In all these cases, you're still able to judge the result. That's what good uses have in common.

What it does poorly

  • Precise facts: figures, dates, quotes, references. It can make them up with total confidence.
  • Recent events: without web search, it doesn't know what happened after its training.
  • Exact calculation and long chains of logic, even if reasoning modes are improving.
  • In images: hands, text inside the image, logos and keeping a character consistent from one image to the next remain classic traps.

Hallucinations, a direct consequence

Since the model produces what's plausible, it sometimes produces something plausible and false: a source that doesn't exist, an invented figure, a mixed-up biography. That's called a hallucination. It's not a bug that will be fixed all at once someday: it's a side effect of how the model works. The tools are getting better, but the habit of checking remains essential. It's all explained in why ChatGPT makes things up.

A useful prompt to stay in control:

Answer my question, then add a "To check" section listing every figure, date, name or reference in your answer that you're not sure about. If you don't know, say so rather than guessing. <question> [your question] </question>

Test the principle yourself

The best way to understand is to observe. Ask the same question twice in two different conversations and compare: the answers vary, proof that the model generates rather than recites. Then ask it to explain how it works:

Explain how you generate your answers, as if to someone who has never done anything with computers. Use an everyday comparison, in under 150 words. Then give me three examples of questions where I should be wary of your answer, and explain why.

For images, run the same prompt three times in a generator. The three different results show you the role of the starting noise.

Going further

On Ovrelia, Prism island helps you understand where generated images come from: how a generator turns noise into an image, which families of tools exist, and why you should describe visual facts rather than wishes. The first chapter, "From Noise to Image," is free and plays in your browser, no account needed: play it now.

To get practical with text, start with how to write a good ChatGPT prompt. And if you want to see what happens when generative AI doesn't just answer but acts, read what is an AI agent, really?. To test your knowledge, launch the friend challenge: ten timed questions about AI, and a score to beat.

Generative AI comes in several families of tools (text assistants, image and video, code, automation). To practice telling which one fits a given task, play the free mini-game Which AI for the job?.

Frequently asked questions

What's the difference between AI and generative AI?

AI is a broad field that also includes systems that classify, sort or recommend. Generative AI is the part that produces new content: text, images, sound, video or code, from an instruction.

How does ChatGPT generate its answers?

It splits text into tokens and predicts, step by step, the most likely piece to come next, based on what it learned during training. So it produces a plausible continuation, not verified information.

Does generative AI learn while I use it?

No, not in the sense of its training: when you ask a question, the model applies what it already learned. Some tools remember information through a memory feature, and some services may use your conversations to train future models depending on your privacy settings.

Why are two images generated with the same prompt different?

Most image generators start from random noise that they gradually clean up following your text. Since that starting noise changes with every attempt, the result changes too.