Explained / Copyright and training data
Did they have the right to train on all that? And who owns what it makes?
The short answer
Nobody can give you a clean answer, because the law is genuinely unsettled and differs by country. Two separate questions are usually muddled: whether training on copyrighted work was lawful, and who owns what a model produces. Courts and governments are still working through both, and anyone sounding certain is guessing.
Two questions, often confused
The first is about the input. Models were trained on enormous amounts of text and images, much of it copyrighted, mostly without asking. Whether that was lawful is being argued in courts in several countries right now.
The second is about the output. If you generate an image or a paragraph, who owns it, and can you use it commercially? That depends on the tool's terms and on national law, and the two do not always agree.
People argue past each other constantly because they are answering different questions. It is worth being clear which one you mean.
BEEP / THE MYTH-BUSTER
I was trained on a very large amount of work by people who were not asked. I cannot tell you whether that was legal. I can tell you that finding out is not my job and is currently several courts' problem.
What is actually known
Some things are reasonably settled. Copying a work to train a model is generally an act that copyright law has something to say about, which is why the arguments centre on whether an exception applies.
Several providers now offer indemnities to business customers — a promise to defend them if output leads to a claim. Offering that is a commercial decision, not evidence that the law is settled.
Some models are trained on licensed or public-domain material and say so. That is a genuine difference, and it is checkable.
In the United Kingdom, the position has been under active government consultation, which is itself a fair signal that the law was not clear enough.
What this means for you
For private use — drafting, thinking, explaining — none of this is likely to trouble you.
For commercial work, it matters, and the practical risks are not really about landmark court cases. They are mundane: output that closely resembles an existing work, a client contract that forbids generated material, a stock library whose terms you did not read.
Questions worth answering before commercial use
- What do the tool's terms say?
- Whether you own the output and may sell it. This varies more than people expect.
- What does your client's contract say?
- Many now require disclosure, and some prohibit generated work outright.
- Does it resemble something existing?
- Close resemblance is a risk whatever the training argument turns out to be.
- Was it trained on licensed material?
- Some tools say so plainly. Worth preferring when the work is commercial.
The shape of the diligence rather than legal advice. For anything significant, ask somebody qualified in your jurisdiction.
What to actually do
- Read the output-rights section of any tool you use commercially.
- Ask clients what they expect to be disclosed, before the work rather than after.
- Treat close resemblance to an existing work as a problem regardless of the law.
- For anything valuable, take proper legal advice. This page is not it.
Why we are not going to pretend otherwise
It would be easy to write a confident paragraph here. Plenty of sites have. The honest position is that this is live litigation and active policy in several countries at once.
We will update this page when something actually changes, and say what changed. Until then, treating it as unsettled is not fence-sitting; it is the accurate description.
See it for yourself
Reading about it only gets you so far. This one is written to show you the thing described above, rather than describe it again.
If I publish something you helped me write, what should I check before I do? Tell me where those rules are actually written down rather than summarising them.
Then paste it into whichever assistant you use — ChatGPT, Claude, Gemini or Copilot. Not sure which? Meet them all.
If you remember nothing else
- Two separate questions: was training lawful, and who owns the output.
- Both are genuinely unsettled, and the answers differ by country.
- For private use this rarely matters. For commercial work it does.
- The practical risks are resemblance and contracts, not landmark rulings.
- Anyone giving you a confident answer today is guessing.
Where this came from
- UK Government, Copyright and Artificial Intelligence consultationchecked 26 September 2026
Anything here that turns out to be wrong is worth telling us about. Figures that move — prices, sizes, energy use — are kept in the registry with their dates, rather than written into the prose where they would quietly go stale.
Questions about “Copyright and training data”Asked by readers, answered by us. Open to read them.
Open to load them.
Ask about this page and we will answer it here, where the next person can read it too.
Nobody else can answer you, and you cannot answer anybody else. This is not a forum — we would rather say that than run a room we cannot watch at three in the morning.
Your question goes up with the answer and without your name. Nothing identifies you, ever.
If we cannot answer it we will tell you why, and that one stays between us.