An illustrator finds her distinctive style – the line weight, the muted palette, the odd perspective – reproduced in hundreds of images generated by strangers typing her name into a prompt box. Meanwhile, a marketing team generates a campaign visual in thirty seconds and then discovers, during trademark filing, that nobody can say for certain whether they own it. Generative AI and copyright collide from both ends: what goes into the models, and who owns what comes out.
The honest answer to most questions here is "it depends on where you are, and some of it is still being decided in court". This article lays out what is reasonably settled, what is contested, and what creators and companies can do in the meantime. It is a general overview, not legal advice.
The two copyright questions generative AI raises
It helps to separate the debate into two distinct questions:
- Input: Is it lawful to train an AI model on copyrighted works without permission?
- Output: Is AI-generated content protected by copyright, who owns it, and can it infringe someone else's work?
They are often blurred together in headlines, but the law treats them very differently.
Training data and copyright: the input question
Training a model involves copying works – at least temporarily – into a dataset. Whether that copying is permitted depends heavily on jurisdiction.
United States. The central question is fair use, a flexible defence that weighs four factors: the purpose and character of the use (including whether it's transformative), the nature of the work, the amount used, and the effect on the market for the original. AI developers argue training is highly transformative; rights holders argue it copies entire works and competes with them. Numerous lawsuits by authors, news publishers, music labels and visual artists are working through the courts, and early decisions have been narrow and fact-specific rather than establishing a general rule.
European Union. The 2019 Directive on Copyright in the Digital Single Market created text and data mining (TDM) exceptions. Research organisations get a broad exception; for everyone else, TDM is allowed on lawfully accessed works unless the rights holder has reserved their rights in an appropriate manner – for online content, in a machine-readable way. The EU AI Act adds that providers of general-purpose AI models must have a policy to respect those reservations and publish a summary of training content.
United Kingdom. UK law has a TDM exception for non-commercial research only. The government has consulted on broadening it, and the issue remains politically contested.
Elsewhere. Some countries, such as Japan, have relatively permissive data-analysis exceptions, though with limits. If you train or fine-tune models, local law matters.
Who owns AI-generated content: the output question
In the US, the Copyright Office's position is that copyright protects only human authorship. Content generated entirely by an AI system from a prompt isn't registrable; a work that combines AI-generated material with sufficient human creative contribution – selection, arrangement, substantial modification – can be protected for the human parts. Courts have so far upheld the human-authorship requirement.
In the EU, copyright protects a work that is the "author's own intellectual creation", a standard developed by the Court of Justice that is widely read as requiring human creative choices. Purely machine-generated output is unlikely to qualify.
The UK is an outlier on paper: its copyright law has, since 1988, included protection for "computer-generated works" with no human author, assigning authorship to the person who made the arrangements necessary for the work's creation. How that applies to modern generative AI is untested, and the provision is under review.
The practical consequence is significant. If your company's logo, character design or product copy is purely AI-generated, you may have weak or no copyright protection, meaning competitors could copy it. Trademark law, contracts and trade secrets may fill some gaps, but not all.
Can AI output infringe copyright?
Yes. Even if the training question is resolved in developers' favour, a specific output that reproduces a substantial part of a protected work – a recognisable character, a passage of text, a melody – can infringe, just as it would if a human had copied it. Style itself is generally not protected by copyright, but the line between "in the style of" and "a copy of" is a factual question.
Trademarks and personality rights add another layer: generating images featuring a famous brand's logo or a real person's likeness can create liability that has nothing to do with copyright.
What creators can do
- Use machine-readable opt-outs on your website (robots.txt rules for AI crawlers, metadata reservations). In the EU this has legal weight for commercial TDM.
- Check platform terms. Some portfolio, stock and social platforms license uploaded content for AI training; read what you've agreed to.
- Register your work where registration helps enforcement, as in the US.
- Document your process. If you use AI in your own work, records of your human contribution support a copyright claim.
- Watch collective action. Licensing schemes and class actions are where much of the bargaining power currently sits.
What companies should do
| Risk | Practical step |
|---|---|
| Weak ownership of AI output | Keep meaningful human creative input in key assets; don't rely on pure AI output for brand identity |
| Output that copies protected work | Review outputs; avoid prompts naming living artists, brands or characters |
| Vendor risk | Prefer tools that offer contractual indemnity for output claims and read its conditions |
| Confidential data leakage | Use business plans that don't train on your inputs |
| Disclosure duties | Label AI-generated content where law or platform rules require it |
| Policy gaps | Write an internal policy on which tools may be used for what |
The bottom line
Generative AI and copyright law are still settling, and they are settling differently in the US, the EU and the UK. What's reasonably clear today: purely AI-generated output has weak protection in most places, outputs can still infringe, and training-data rules depend on where the training happens. Keep humans meaningfully involved in work you want to own, read your vendors' terms, respect opt-outs, and revisit your policy as court decisions arrive.
