What is generative AI?
In one line
Ages 8 to 10: Generative AI makes new words, pictures or sounds by following the patterns it saw in lots of examples, and it can be wrong.
Ages 11 to 13: Generative AI learns which pieces tend to follow which, then builds new content one piece at a time by picking likely next pieces.
Ages 14 to 18: Generative AI uses models that approximate the statistical patterns of their training data to sample new content, which is why fluent output can still be false.
Generative AI is artificial intelligence that makes new content, such as text, pictures, sound or video. It learns patterns from a very large number of examples, then produces new output that follows those patterns. What it makes can look convincing and still be wrong, so it needs checking.
Last reviewed: 20 October 2026 7 min read
The idea in one picture
Imagine someone who has read only two sentences:
the cat sat on the mat
the dog ran to the park
They notice which word comes after which. After "the" comes "cat", "mat", "dog" or "park", each 1 time in 4. After "dog" always comes "ran", then "to", then "the".
Now ask them to make a sentence. Start with "the", pick a word that has followed it, and repeat:
the → dog → ran → to → the → mat
"The dog ran to the mat" is in neither example. It is new, and it follows the patterns. The same chain could just as easily give "the cat sat on the park": it follows the patterns and makes little sense. Real generative AI learns far richer patterns from vastly more examples, but the core idea is the same: learn patterns, then build something new that fits them.
How it works, step by step
- Gather examples. A model is trained on a very large collection of text, images, audio or video.
- Learn the patterns. Training adjusts the model's numbers so it gets better at predicting what fits, such as the next piece of text.
- Give it a prompt. A person types or supplies an instruction or starting point.
- Build the output piece by piece. Text models add one token (a word or part of a word) at a time. Many image models start from random noise and remove it step by step until a picture appears.
- Pick with some chance. The model often chooses among likely options rather than always the single likeliest, so the same prompt can give different results.
- Check the result. People check what comes out, because it can be wrong or made up.
A little deeper (ages 11 to 13)
The two-sentence example is a probability machine. After "the", each of 4 words has a 1 in 4 chance, or 25%. To make "the dog ran to the mat", the picks after each "the" each had a 1 in 4 chance, and the rest were certain: ¼ × 1 × 1 × 1 × ¼ = 1/16. So it would make that exact sentence about once in every 16 tries starting from "the".
Real systems do the same kind of thing with far more options. A setting called temperature controls how adventurous the picks are: lower means more predictable output, higher means more random output.
This also shows why generative AI can be wrong. It picks what is likely to come next given its patterns, not what has been checked as true. Likely and true often match, but not always.
Going further (ages 14 to 18)
A US standards body’s guidance uses this definition: generative AI models "emulate the structure and characteristics of input data" to generate "derived synthetic content" such as images, video, audio and text.
Large language models generate text by predicting the next token, again and again. Many image generators are diffusion models: training gradually adds noise to images, and the model learns to reverse that process, so it can start from pure noise and denoise its way to a new image.
The same standards body describes confidently stated but false output, often called hallucination, as a natural result of the design: models produce outputs that approximate the statistical distribution of their training data.
Models can sometimes reproduce parts of their training data, including private information (memorization). And generated images, audio and video can be hard to tell from real recordings, which is why provenance records and content labels matter.
Common mix-ups
- Mix-up: "It looks the answer up."Fact: By default it generates text from learned patterns. Some systems add a search step that fetches sources first, and even then the output can mix in errors.
- Mix-up: "If it sounds sure, it’s right."Fact: Confident tone says nothing about accuracy. Check important facts against a reliable source.
- Mix-up: "Everything it makes is brand new."Fact: Most output is a new combination, but models can sometimes reproduce parts of what they were trained on.
- Mix-up: "Generative AI means chatbots."Fact: A chatbot is one way to use it. Generative AI also makes images, music, speech and video.
Using it with care
- Check facts that matter, and ask where a claim comes from.
- Keep personal details, such as full names, addresses and photos, out of prompts.
- Say when something you share was made with generative AI.
Try it (no device, 5 minutes)
Write two short sentences that share a word, such as "the". For each word, list the words that came right after it. Start with the first word, pick a word that followed it, and keep going until you reach a word with nothing after it. Make three sentences. Which are new? Which are nonsense? That is generation at its smallest, and it shows why fluent is not the same as true.
Quick check
In the two-sentence example, which word always comes after "dog"?
Answer: "ran". It is the only word that ever follows "dog", so the chain always picks it.
"ran". It is the only word that ever follows "dog", so the chain always picks it.Generative AI gives an answer that sounds very sure. Does that mean it is true?
Answer: No. It picks what is likely given its patterns, not what has been checked. Check important facts against a reliable source.
No. It picks what is likely given its patterns, not what has been checked. Check important facts against a reliable source.Why can the same prompt give two different answers?
Answer: The model often picks among several likely next pieces with some chance, rather than always the single likeliest one. A higher temperature makes the picks more random.
The model often picks among several likely next pieces with some chance, rather than always the single likeliest one. A higher temperature makes the picks more random.Questions
Generative AI is a kind of AI that makes new things, such as stories, pictures, music or videos. It learns patterns from many examples and then creates new content that follows those patterns.
Many AI systems predict a label or a number, such as whether a message is spam. Generative AI produces new content instead, such as a paragraph, an image or a sound.
No. It produces what is likely given the patterns it learned, which is not the same as what is true, so it can state false things confidently. Important facts need checking against a reliable source.
Tools that write text, make images from a description, produce speech or music, or create short videos are all generative AI. Large language models, which generate text, are one well-known kind.
Mostly it produces new combinations of learned patterns. It can sometimes reproduce parts of its training data closely, which raises privacy and copyright questions.
Mechna Academy is a city of machines for curious children aged 8 to 18, opening first to ages 8 to 10. In each mission, a child guesses how a machine decides, then opens its Peek Hatch to check, changes one thing and tests it on a case it has never seen. Children learn how AI works this way, and they never chat with an AI.
For grown-ups: join the waitlistPlaces are limited. The waitlist is for parents and guardians 18 and older.