What is generative AI?

Generative AI covers models that produce new content — text, images, audio, video, code — rather than classifying existing content. The technical difference is that a classifier learns the boundary between categories while a generative model learns what the data itself looks like, which is a substantially harder problem and the reason generation arrived later than recognition.

How does image generation actually work?

Diffusion models are trained by adding noise to real images step by step until nothing remains, then learning to reverse each step. To generate, the model starts from pure random noise and removes a little of it repeatedly, guided by your text description, until an image emerges. There is no database of images being searched or assembled.

Does a generative model copy from its training data?

Not by design and occasionally in effect. The model stores statistical patterns rather than images, so most output is genuinely new. Images duplicated many times across the training set can be reproduced closely, which is a real phenomenon and the basis of several legal disputes, and it is the exception rather than the mechanism.

What does temperature do in text generation?

The model produces a probability for every possible next token, and temperature reshapes that distribution before one is chosen. Low temperature concentrates probability on the most likely tokens, giving repetitive but predictable output; high temperature flattens it, giving varied and sometimes incoherent output. It controls variability, not correctness.

Can I use generated images commercially?

It depends on the tool's terms and your jurisdiction, and the position is genuinely unsettled. Several jurisdictions have held that purely machine-generated work cannot be copyrighted, meaning you may be unable to prevent others from using your output. Separately, whether training on copyrighted material was lawful is still being litigated, so an indemnity clause in your provider's terms is worth reading carefully.

Why does generated text state false facts so confidently?

The model is trained to produce plausible continuations, not true ones, and it has no mechanism for distinguishing the two. A fabricated citation is generated by exactly the same process as a correct one, which is why confidence carries no information about accuracy and why anything factual must be verified independently.

Generative AI: how machines produce images and text

For most of the field's history, models answered questions about existing things: is this a cat, is this review negative, will this customer churn. Producing something new was a different and much harder problem, and it was solved recently enough that the consequences are still being worked out.

What this course sets out to do: explain what actually happens when you type a prompt and something appears, so you can predict when it will work and when it will not.

What this course does not do: teach you to train generative models. The premium catalogue covers Stable Diffusion, fine-tuning and production pipelines.

What you are about to discover

Course contents

#	Lesson	Main goal	Time
1	What generating means	Discriminative versus generative, and why it is harder	8 min
2	How images are generated	Diffusion, from noise to picture, and how text steers it	9 min
3	How text is generated	Next-token prediction, sampling, temperature, context	9 min
4	Using it well	Where it pays off, how to prompt, and what to automate	8 min
5	The limits that matter	Fabrication, copyright, deepfakes, environmental cost	9 min
6	Recap and FAQ	Synthesis, a usage guide, and 12 common questions	6 min
7	Quiz and attestation	Validate what you learned with 5 corrected questions	3 min

Is this course for you?

You use these tools daily and want to know what is happening underneath.
You are deciding whether to put generative AI in a product and need to judge the risks.
You keep hearing "diffusion", "temperature" and "hallucination" and want them precise.
You need to explain to a colleague or a client what these tools can and cannot be trusted with.

No prerequisites beyond Introduction to AI.

What you will be able to do at the end

Explain the difference between recognising and generating, and why the second is harder.
Describe diffusion without hand-waving, and say why prompts steer it.
Explain next-token prediction and what temperature actually changes.
Judge which tasks are worth automating with generation and which are not.
State the copyright position accurately, including what remains unsettled.
Recognise fabrication as a structural property rather than a bug awaiting a fix.

Estimated time

Around 45 to 55 minutes of reading.

Prerequisites and next steps

Prerequisites: Introduction to AI. Deep Learning helps for lesson 2.

Natural continuation:

Large Language Models — the text side in depth
Computer Vision — reading images rather than producing them
Ethics of AI — consent, attribution and misuse

Frequent questions, answered in one line

Is generative AI creative?

It recombines patterns from an enormous body of existing work, which produces genuinely novel combinations and involves no intention, taste or judgement about what is worth making. Whether that counts as creativity is a question about the word rather than the technology, and the practically important part is that the model cannot tell whether its output is good.

Will it replace designers and writers?

It has already changed what they spend time on, which is different from replacement. Producing a first draft or twenty variations is now nearly free, and deciding which one is right, whether it serves the brief and whether it is accurate remains entirely human — and is where most of the value was.

Can I tell whether something was generated?

Increasingly not, reliably. Detection tools exist and their false positive rates are high enough to make accusations dangerous, particularly against non-native writers. Provenance approaches that sign content at creation are more promising than detection after the fact.

Do I need a powerful computer?

To use hosted tools, no. To run image generation locally, a consumer graphics card with enough memory is sufficient and increasingly common, which matters when your inputs cannot leave your infrastructure.

Want to build with generative models rather than read about them?

The premium catalogue covers Stable Diffusion, fine-tuning, LoRA and production pipelines, with a verifiable certificate after a 40-question examination. Included in every paid plan.

Ready? Start with lesson 1 →

What you are about to discover​

Course contents​

Is this course for you?​

What you will be able to do at the end​

Estimated time​

Prerequisites and next steps​

Frequent questions, answered in one line​