What is generative AI? How machines make text and images

Generative AI is artificial intelligence that creates new content, such as text, images, audio, video or code, instead of only sorting or scoring what already exists. It learns the patterns in huge sets of human-made examples and then produces new material that fits those patterns. Chatbots that write and tools that turn a sentence into a picture are both generative AI, built on different designs.

In short

• Most AI predicts or classifies: is this email spam, will this loan be repaid. Generative AI produces something new.
• Text tools are mostly large language models that write one word-piece at a time; image tools are mostly diffusion models that turn noise into a picture.
• Both learn from vast collections of human-made work, which is why who owns that work is one of the field's open legal questions.

How is generative AI different from other AI?

Most machine learning in daily use is predictive. It takes an input and gives back a label, a score or a decision. Generative AI takes an input, often a written request called a prompt, and gives back a new piece of content. The same email shows the difference.

	Predictive AI	Generative AI
Question it answers	What is this, or what will happen?	What would a new example look like?
Email example	Is this email spam? Yes, 97% likely	Write a polite reply declining the meeting
Output	A label, number or ranking	Text, an image, audio, video or code
Typical uses	Fraud alerts, recommendations, credit scores, medical image screening	Chatbots, image and video generators, coding assistants, voice synthesis
How to judge it	Was the prediction right?	Is the output good, accurate and appropriate, which is often a matter of judgment

Both are built the same way underneath: a neural network trained on examples. What changes is what it is trained to produce. A predictive model learns to tell examples apart; a generative model learns what examples look like well enough to make new ones.

How do text generators work?

The text tools people meet, such as ChatGPT, Gemini and Claude, run on large language models. A language model is trained on a vast amount of text to predict the next small piece of text, called a token. To write, it predicts a token, adds it to the text, and repeats. Chatbots get further training on example conversations and on people's ratings of answers, which turns a raw text predictor into an assistant. The design inside nearly all of them is the transformer, the T in GPT, which lets the model weigh every word in its input against every other.

Code assistants are language models trained heavily on code, and many music and speech generators use the same next-piece approach on audio.

How do image generators work? A worked example

Most image generators are diffusion models. The method was laid out in a 2020 paper by Jonathan Ho, Ajay Jain and Pieter Abbeel, Denoising Diffusion Probabilistic Models. The trick is to teach a model to undo noise.

1. Ruin a picture, gradually. Take a training photo, say a lighthouse. Add a little random static. Then a little more. The 2020 paper used 1,000 such steps, after which the photo is pure static with nothing of the lighthouse left.
2. Learn one step back. At each step, the model is shown the noisier image and trained to predict the noise that was added, so it can be subtracted. One step back from static is still almost static; one step back from nearly clean is almost the photo. Repeated across millions of images, the model learns what real pictures look like at every level of blur.
3. Steer with words. During training, each image comes with its caption. The model learns that "lighthouse" goes with tall white towers on rocks, and "at sunset" with orange skies.
4. Generate. To make a new image, start with fresh random static, which contains no picture at all. Ask the model to remove one step of noise, guided by the prompt "a lighthouse at sunset". Then again, dozens or hundreds of times. Shapes appear, then edges, then detail. The result is a lighthouse that never existed.

Because each run starts from different random static, the same prompt gives a different picture every time. And because the model learned patterns rather than storing photos, it can combine things it never saw together, such as a lighthouse made of glass.

An earlier design, the generative adversarial network or GAN, was introduced by Ian Goodfellow and colleagues in 2014. It trains two networks against each other: a generator that makes fakes and a discriminator that tries to tell them from real examples. GANs made the first convincing fake faces, and are still used in some deepfake tools, but diffusion models replaced them for most image generation.

Where does the training data come from?

Generative models need enormous amounts of human-made material. Language models are trained on text gathered from the public web, books, code and other sources, adding up to trillions of tokens. Image models are trained on images paired with captions. LAION-5B, a public dataset released in 2022 and used to train early versions of the Stable Diffusion image model, contains 5.85 billion image-text pairs collected from the web.

Most of that material was made by people who were not asked. This is the source of the field's biggest legal fight. Authors, artists, photographers, musicians and news organizations have filed many lawsuits since late 2022, among them The New York Times's case against OpenAI and Microsoft, filed in December 2023.

Still argued

Whether training on copyrighted work without permission is legal is unsettled. AI companies argue that training is fair use, or its equivalent elsewhere: the model learns patterns from the work, as a person learns from reading, and doesn't store or republish it. Creators argue that their work is being copied at industrial scale to build products that compete with them, without consent or pay. In June 2025, two US district judges in San Francisco ruled days apart. In Bartz v. Anthropic, the judge found that training on books was fair use but that keeping millions of pirated copies in a central library was not, and Anthropic then agreed to settle with the authors for $1.5 billion. In Kadrey v. Meta, the judge ruled for Meta on fair use but stressed that the decision covered only those authors, who had not shown harm to the market for their books. Neither ruling binds other courts; a summary of both by Norton Rose Fulbright sets out the reasoning. The US Copyright Office released a pre-publication report on AI training in May 2025, and laws differ by country. The central question has no settled answer across courts and countries.

What is generative AI good and bad at?

It is strong at first drafts, summaries, rewording, translation, brainstorming, illustrations, and routine code. It lowers the cost of making content close to zero, which helps people who couldn't otherwise afford it.

The same features cause its problems. Generators produce what is plausible, not what is true, so text models invent facts and citations, a habit known as hallucination. They can repeat biases and stereotypes in their training data. They make convincing fake images, voices and video easy to produce. And because content is cheap, low-effort material made in bulk floods the web, which has acquired its own name: AI slop.

Generative AI is one branch of artificial intelligence, and the newest to reach the public. The models underneath, trained on examples and tuned by adjusting numbers, are the same kind of machine learning that has run spam filters and recommendations for years.

Check yourself

• Without looking back: what is the difference between a predictive model and a generative one, using the same email?
• Explain how a diffusion model gets from random static to a picture of a lighthouse.
• Why does the same prompt give a different image each time?

Sources

• Denoising Diffusion Probabilistic Models — Jonathan Ho, Ajay Jain and Pieter Abbeel, 2020
• Generative Adversarial Networks — Ian Goodfellow and others, 2014
• LAION-5B: An open large-scale dataset for training next generation image-text models — Christoph Schuhmann and others, 2022
• Copyright and Artificial Intelligence — U.S. Copyright Office, 2024–2025
• Two US decisions find that reproducing works to train large language models is fair use, Part 2: Kadrey v Meta — Norton Rose Fulbright, 2025
• Generative artificial intelligence — Wikipedia