When a model trains on millions of works made by real people, is it learning the way we do — or copying? The sharpest question in the subject.
August 2026 · Case 4 of 6
As you read — hold this question
If a model reads everything humans have made in order to make something new, is that fair-use reading — or copying at a scale no human could manage?
2025
2025 is when Australia’s Productivity Commission proposed letting AI train freely on copyrighted work — the creative industries revolted, and the government ruled it out and turned to licensing instead.
Copyright exists the instant a work is made, and almost everything online is copyright. So training a model on millions of songs, books and images is, at its heart, a copyright fight. The question the traditional account poses exactly right: is the model learning, the way we do, or is it copying?
The core question
The examinable core
Learning, or copying?
When a model trains on copyrighted work, is that learning or copying? Australia tried to answer: in 2025 its Productivity Commission proposed letting AI train freely on copyrighted material. The creative industries revolted, and the government ruled it out and turned to licensing — if you want to train on someone’s work, you pay for it.
There’s a bind: Australia also wants its own AI industry, and making training harder at home just pushes it offshore.
Carry this
The training fight is a copyright fight. Australia chose licensing over free training — pay to train — but that same choice risks pushing AI development overseas.
Two serious people, opposite reads
The other side
Marcus versus Sag
Gary Marcus: generative AI has a plagiarism problem — these models can reproduce near-copies of real work even when no one asked. Training on it without paying isn’t learning; it’s copying. Matthew Sag: training is reading in order to learn, which is fair use — unless the model actually reproduces the original. Clamp down too hard and you kill the research.
The risk that cuts deepest
One harm is worse than knock-offs and hollowed-out brands: the loss of control over Indigenous art and culture — sixty thousand years of story and symbol scraped into a dataset and sold back as a pattern on a t-shirt. That isn’t a market problem; it is a kind of theft the law is still struggling to name.
Quick recall — without looking back
Test yourself on this case
Question 1 of 3
What did Australia decide about training AI on copyrighted work?
The 2025 Productivity Commission proposed free training; after the creative industries revolted, the government ruled it out and chose licensing — pay to train. The bind: it risks pushing AI development offshore.
Question 2 of 3
Summarise the Marcus vs Sag disagreement.
Marcus: models reproduce near-copies, so training without paying is copying (a plagiarism problem). Sag: training is reading to learn = fair use, unless the output reproduces the original; over-clamping kills research.
Question 3 of 3
Why is copyright the natural home of the training debate?
Because copyright is automatic the instant a work is made, and almost everything trained on is copyright. So training is, at its heart, a copyright question.