Welcome back. Let's look under the bonnet
In Level 1 we looked at what AI is actually changing in creative work. Now we'll take on a question that comes up in almost every conversation between creatives: how do these tools learn to make things, and why are so many artists, writers and musicians angry about it?
First, a quick note. This course gives you general information, not legal advice. When something specific comes up in your own work, a qualified professional is the right person to ask.
Ready? Let's go.
Training: learning from a mountain of examples
Here's a question for you. How did you learn what a cat looks like? Nobody gave you a rulebook. You saw hundreds of cats, and over time your brain picked up the pattern.
Generative AI works in a loosely similar way, but at a scale that's hard to picture. A model is "trained" by showing it huge numbers of examples, like text, images or audio, and having it adjust millions or billions of internal settings until it gets good at predicting what tends to go with what.
Take an image generator. It's usually trained on pictures paired with text descriptions, so it learns that the words "golden retriever on a beach at sunset" go with certain shapes, colours and textures. One well-known public dataset, LAION-5B, released in 2022, contains about 5.85 billion image-and-text pairs gathered from across the web. It was used in building models such as Stable Diffusion.
Text models learn by predicting the next word, again and again, across vast amounts of writing. Music models do much the same thing with sound.
In simple terms, the process goes like this:
- Collect a huge amount of material, often by automatically "scraping" the public web.
- Pair it with labels or descriptions where needed.
- Train the model to spot and reproduce patterns, nudging its settings over countless rounds.
- Generate new output from those learned patterns when someone types a prompt.
Why training data matters so much
This is the key idea of the whole lesson: a model can only produce what its training data has taught it.
Think of it like a cook who has only ever tasted dishes from one cookbook. Their "new" recipes will still taste like that book. If a model has seen thousands of paintings by a living illustrator, it can produce images "in the style of" that illustrator. If it's never seen a particular style, it can't really imitate it.
That's why creatives care so much about what went in. Their portfolios, novels, photos and songs are often part of that mountain of examples. In most cases nobody asked them first.
The three Cs: consent, credit, compensation
So what exactly are creatives objecting to? A handy way to remember it is the "three Cs".
Consent. Was the creator asked? Most of the time, no. Work posted online to win clients or share with fans got swept into training sets automatically.
Credit. Is the creator acknowledged? An illustrator whose style shapes thousands of AI images usually gets no mention at all.
Compensation. Is the creator paid? Their work helps build commercial products, and in many cases they receive nothing, while those same products may compete with them for jobs.
You can see how this lands across different fields. A concept artist finds people prompting with her name. A session musician hears AI tracks that sound uncannily close to his. A copywriter is told a chatbot can "do it for free", and that chatbot learned from writing like his.
This pushback is organised and large. The "Statement on AI Training" says that unlicensed use of creative works for training generative AI is "a major, unjust threat" to creators' livelihoods, and tens of thousands of people have signed it, including Björn Ulvaeus, Kazuo Ishiguro and Max Richter. In February 2025, about 1,000 UK musicians released a silent album called Is This What We Want? as a protest. In the same month, UK newspapers ran front-page "Make It Fair" wraps.
The other side of the argument
Now, to understand this debate properly, you need to hear both sides. So what do AI developers and their supporters say?
Their first argument is that the model learns patterns rather than storing copies, much as a human art student studies the masters. Their second is about scale and practicality. In evidence to a UK House of Lords committee in December 2023, OpenAI said it would be "impossible" to train today's leading models without copyrighted material, because copyright covers nearly all modern human expression. They also point to wider benefits, like new tools, research and creative possibilities, and some offer opt-outs for creators.
Creatives answer back. A human student doesn't learn from billions of works in weeks and then produce on demand for millions of users. Opt-outs put the burden on the creator, and often only after training has already happened. And "it's impossible without your work" sounds to many like a strong reason to pay for that work, not to take it free.
So who's right? Honestly, that's still being argued in courts and parliaments, and we'll get to the law in later levels. What matters for you now is being able to explain both sides clearly and fairly. That makes you far more convincing when you talk to clients, peers or policymakers.
Your practical action
This week, pick one AI tool that's relevant to your field. Spend ten minutes finding out what it says publicly about its training data. Then jot down where it stands on each of the three Cs: consent, credit and compensation. You'll start to see the landscape much more clearly.
Key takeaways
- Generative models learn patterns from huge collections of text, images and audio, and they can only produce what that data taught them.
- Much training data was scraped from the web without asking creators.
- Creative objections centre on the three Cs: consent, credit and compensation.
- Developers argue that models learn rather than copy, that training at scale needs copyrighted works, and that the tools bring wider benefits.
- Creatives reply that the scale is different, that opt-outs shift the burden onto them, and that essential work should be licensed and paid for.