What is generative AI?

Generative AI versus other AI

Most AI before the 2020s was predictive. It looked at an input and gave back a label or a number: is this email spam, what is the price of this house, which face is in this photo. Generative AI turns the direction around. It takes a short input and gives back a whole new piece of content.

Both kinds are trained on examples, and both are statistical. The difference is the shape of the output. A spam filter outputs one word. A language model outputs a paragraph that did not exist before.

Many products mix the two. A speech-to-text tool is generative in the narrow sense, because it writes a transcript, but it is close to prediction, because there is one correct transcript.

How generative models work

Every generative model follows the same two-phase idea. In training, the model sees millions or billions of examples and adjusts its internal numbers, called parameters or weights, so that it gets better at reproducing the patterns in that data. In use, which is called inference, the model starts from your prompt and builds new output step by step, guided by what it learned.

The model does not store a copy of each training example and paste it back. It stores a compressed statistical picture of what text, images or sounds tend to look like. That is why it can produce new combinations, and also why it can produce confident mistakes.

Three families of models do most of the work today:

The transformer architecture, introduced in 2017, is behind almost all large language models. Diffusion models became the main approach for images after 2020. GANs came first, in 2014, and are still used for some fast image tasks. The lines are blurring: some image models are autoregressive, and some video models combine diffusion with transformers.

What generative AI can make

The same idea applies to every kind of media. What changes is the training data and how the output is represented.

Not every AI media tool is generative in the full sense. cv.cm's photo to video tool, for example, estimates a depth map from one photo and moves a virtual camera through it. The camera moves, but the subject does not, because no new frames of motion are invented. A text-to-video model, by contrast, generates every frame.

A worked example: one idea, several models

Suppose a small shop wants a 15-second clip to announce a new ceramic mug. A typical generative workflow looks like this:

1. Text: ask a chat model for three short captions and a 30-word voiceover script. Edit the one you like by hand.
2. Image: describe the shot to an image model, for example a product on a wooden table in soft morning light. Generate several and pick one.
3. Video: give that image to an image-to-video model with a short motion prompt, such as a slow push-in while steam rises.
4. Voice: turn the script into speech with a text-to-speech model, or record it yourself.
5. Subtitles: run the audio through a speech-to-text model to get an SRT file, then fix any errors.
6. Review: check the product looks like the real one, the text is spelled correctly, and you label the video as AI-generated where required.

Each step uses a different model, but all of them do the same basic thing: they turn a short instruction into new content that matches patterns from training. Each step also needs a human check, because each model has its own typical failures.

Weak prompt:
mug ad

Stronger prompt:
A white ceramic mug with a thin blue rim on a light oak table,
soft morning window light from the left, steam rising,
shallow depth of field, product photo, no text, no logo.

The stronger prompt names the subject, the material, the setting, the light and what to leave out. Generative models fill every gap with the most typical choice from their training, so the more you specify, the less they guess.

What generative AI does well and what it does badly

Generative AI is strong at producing a first draft quickly, rewording text, summarizing long material you provide, translating, brainstorming many options, and turning a rough idea into a picture or a sound.

It is weak in predictable places:

• Facts: language models can state false information with full confidence. This is called hallucination or confabulation.
• Counting and exact details: five fingers, three apples, a 12-word sentence. Models approximate.
• Text inside images: signs and labels often come out misspelled, though this has improved at the time of writing.
• Physics and continuity in video: objects can merge, vanish or move in impossible ways.
• Fresh information: a model only knows what was in its training data, up to a cutoff date, unless it is connected to search.
• Rare languages, dialects and accents: quality drops where training data was thin.

The practical rule is to use generative AI where a mistake is easy to spot and cheap to fix, and to verify everything where a mistake is costly.

Using generative AI responsibly

A short checklist before you publish or rely on AI output:

1. Verify facts, numbers, quotes and links against a primary source.
2. Do not paste private data, passwords, client files or ID documents into tools you do not control. Browser-based tools that run the model on your device avoid uploading the file.
3. Check the terms of the tool for commercial use, and do not prompt for real people, logos or characters you have no right to use.
4. Label AI-generated or AI-edited media where the law or the platform requires it. China, the EU and many platforms now have labeling rules.
5. Keep a human responsible for the final result, especially in health, law, finance and news.

Generative AI is a tool that predicts plausible output. It is very useful when you treat it as a fast assistant that drafts, and not as an authority that knows.