Definition
Generative AI is AI that creates new content, such as text, images, audio, video or computer programs, by learning patterns from existing examples. Predictive AI assigns a label or number to an input, while generative AI produces something new that resembles what it learned from.
In simple words
A predictive model looks at a photo and says "cat". A generative model takes the word "cat" and draws a new one. It's like a student who has read so many stories that they can write a new one in the same style. It doesn't paste from a store of stories, but it can sometimes repeat what it saw.
Why it matters
It has changed how people write, build software, design and search. It's useful, and it's risky: it can be wrong in a confident tone, raises copyright questions, and makes deepfakes easy. Anyone using or building it should understand both sides.
Key terms
- Generative AI
- AI that creates new content by learning patterns from existing examples.
- Predictive AI
- AI that gives a label, number or ranking for an input.
- Foundation model
- A very large model trained on broad data that can be adapted to many different tasks.
- Prompt
- The instruction a person gives to a generative tool.
- Training data
- The examples a model learned from.
- Synthetic content
- Content made by a machine instead of being recorded from the real world.
Explained in detail
From judging to making
For most of its history, AI was used to judge things that already existed. Is this email spam? Is there a tumour in this scan? Which route is fastest? These are tasks with a right answer. Generative AI does something different. Given a request, it produces something that did not exist before: a paragraph, a picture, a tune. There is no single right answer, only better and worse ones. This shift, from choosing among answers to composing new material, is why generative AI feels so unlike the AI that came before.
How it arrived
| Year | Step |
|---|---|
| 2014 | Researchers train two networks against each other, one making fake images and one trying to detect them, and realistic generated faces appear |
| 2017 | The Transformer design makes it possible to train much larger language models |
| 2020 to 2021 | Large language models write convincing paragraphs; a new method produces high-quality images from noise |
| 2022 | Image generators and a chat assistant are released to the public and reach many millions of users within weeks |
| 2023 onwards | Tools become multimodal, handling text, images, audio and video together |
One model, many jobs
Older AI systems were built one task at a time: a separate model for translation, another for summarising, another for answering questions. The large models behind generative AI are different. A single model, trained once on a very broad collection of data, can translate, summarise, answer, write programs and more, simply by being asked. Such models are called foundation models, because many products are built on top of them. This is efficient, and it also means that a weakness or a bias in one foundation model is carried into every product that depends on it.
What is new, and what is not
What is new is who can create. A person with no training in drawing, music or programming can now produce a picture, a tune or a working tool by describing it. What is not new is the foundation. Generative AI is still machine learning. It learned from examples made by people, it reproduces the patterns in those examples, and it inherits their gaps and prejudices. It does not understand its output as a person would, and it has no way of checking whether what it produces is true.
Questions it raises
- Truth: if anything can be generated, how do we know what is real?
- Ownership: who owns generated work, and what is owed to the people whose work trained the model?
- Learning: what happens to skills when a tool can do the task for you?
- Work: how will jobs in writing, design, translation and programming change?
- Fairness: whose language, culture and point of view do these models represent best?
How it works
| Makes | Main technique | Example tools |
|---|---|---|
| Text, computer programs | Large language models | ChatGPT, Claude, Gemini, Llama |
| Images | Diffusion models | Stable Diffusion, Midjourney |
| Audio, music | Diffusion and Transformer models | ElevenLabs, Suno |
| Video | Video diffusion models | Sora, Veo |
Predictive vs generative
A predictive model learns the chance of a label given the input. A generative model learns the patterns of the data itself, so it can sample new examples from what it learned.
Try it yourself
How a new picture is made
It doesn't look up an existing picture. Watch where the new one comes from.
Predictive or generative?
A predictive model gives a label or a number. A generative model produces something new. Which is each one?
Making a song from one line of description
Writing a poem about the monsoon
Looking at a photo and saying "cat"
Estimating the price of a house
Marking an email as spam or not
Drawing a new cat from the word "cat"
Common mistakes
- Thinking it looks up answers. It generates them, unless connected to a search or document tool.
- Sharing private company or personal data with public tools without checking their data policy.
Check yourself
Summary
- Generative AI creates new content, by learning from examples.
- LLMs make text, and diffusion models make images.
- Output can be wrong, so check what matters.
Finished reading?