What is generative AI?

The AI you already used sorted things. Generative AI makes things

For twenty years, the AI in daily life was invisible and it classified. The spam filter decided spam or not. The photo app decided which face was whose. The bank decided fraud or not. The recommendation engine ranked what to show next. Every one of those takes something that exists and puts a label or a score on it. What is AI covers that broader family.

Generative AI reverses the direction. Instead of taking an image and outputting "cat", it takes "a cat on a windowsill at dusk" and outputs an image. Instead of scoring an email, it writes one. The model has learned the patterns of millions of examples well enough to produce a new one that fits.

That reversal is the entire reason the technology reached the public. Classification is a feature inside someone else's product. Generation is a product on its own: type a request, receive a thing. ChatGPT launched on 30 November 2022 and reached an estimated 100 million monthly active users in January 2023, two months after launch, per a UBS analysis reported by Reuters on 1 February 2023. It was the fastest consumer application to that mark at the time. For the first time, the model's output was the point.

What followed wasn't a steady trickle. The model release timeline tracks 41 major launches since that day, and 31 of them landed in 2025 and 2026. Orange marks a reasoning model or a family that ships with a thinking mode; the shaded stretch starts with o1-preview.

Every generative model learns what a plausible example looks like, then makes one

Every generative model is trained the same way at heart: show it an enormous number of examples and make it learn what a plausible example looks like. Then, at use time, ask it to produce a new sample that fits the pattern and the prompt.

For text, the model learns to predict the next token, and generates an answer one token at a time. What is an LLM is the full account. For images and video, the dominant approach is diffusion. The model learns to remove noise from a picture in small steps. At generation time it starts from pure noise and removes it step by step while a text description steers what emerges. The recipe most image models use is the 2020 Denoising Diffusion Probabilistic Models paper by Ho, Jain and Abbeel; the approach it displaced, generative adversarial networks, dates from Goodfellow et al. in 2014. Music generators mostly borrow one of the two.

All of them produce what fits, not what's true. A text model will produce a citation that fits the shape of a citation. An image model will produce a hand that fits the pattern of hands, sometimes with six fingers. The fluency and the errors come from the same place, and what is an AI hallucination covers the text version in detail.

Four media, four kinds of model, one window

In 2022 generative AI meant text and, separately, images. By 2026 a single chat product routes to four kinds of generator, and the model behind each is different.

Whizi meters these differently because they cost differently to run. A text answer is a fraction of a cent. An image is a separate counter. Video and music are the expensive rows and sit on the top plan.

Image generation isn't included on Starter, is 100 a month on Pro and 500 a month on Powerhouse, and video and music generation are Powerhouse features, per how Whizi credits work. The gap between a text answer and a video clip is the gap between generating 500 tokens and generating several seconds of frames, and the pricing follows the compute.

What generative AI is good for, and the honest edge of it

It's good at first drafts of anything with a known shape: an email, a summary, a function, a product photo, a background track. It is good at transformation, taking a thing that exists and producing the same thing in another form: translated, shortened, rewritten for a different reader, restyled. And it is good at volume, producing twenty variants where a person would produce two.

It's weak, by construction, at three things. Facts it hasn't been given, because it generates what fits rather than what is. Precision in the long tail: the sixth finger, the code that compiles and does the wrong thing. And judgement about whether the output is good, which stays with the person who asked. The how to use AI guide is the practical version of working inside those limits.

Generative AI is the right tool when you can recognise a good result faster than you can produce one. If you can't tell a good output from a bad one, it isn't the tool for that task yet.

Generative AI is a branch of machine learning, which is a branch of AI

Artificial intelligence is the umbrella: any software that does something we would call intelligent if a person did it. Machine learning is the method most modern AI uses, learning rules from data instead of being programmed with them. Generative AI is the branch of machine learning whose output is new content. A large language model is the generative model for text. A chatbot is the app wrapped around one. Each is a subset of the one before it, and the AI glossary has the full set of terms in one place.

When a product says it "uses AI", ask which kind. If it sorts, ranks or predicts, it's the old kind and probably works well. If it generates, it's the new kind and needs a person checking the output.