AI and copyright: training, output and what you can publish

AI and copyright is really three questions, and mixing them up is why the arguments go nowhere. One: was it lawful to train a model on copyrighted work? Two: does the output infringe anyone's copyright? Three: can you own what the model produced for you? The short answers today are: unsettled and being fought out in court, usually no but sometimes yes, and only to the extent a human contributed. Here is where each line sits and what it means if you want to publish AI assisted work without a nasty surprise.

This is general information, not legal advice. Copyright is national law, the cases are moving, and the answer for a novel is not the answer for a logo.

Was the training lawful?

Training copies. To learn from a book a model has to ingest it, and ingesting means making a copy, which is the act copyright controls. The defense is that the copying serves a transformative purpose and the model is not a substitute for the works. Whether that holds depends on where you are.

In the United States the question is fair use, decided case by case on four factors, and there is no blanket ruling either way. Courts have been more sympathetic to the argument that training itself is transformative than to how the training material was obtained. The distinction that keeps surfacing is between works a company bought or licensed and works taken from pirate libraries: the first has a serious fair use argument, the second is a straightforward reproduction problem no matter how the model behaves afterwards. Litigation is ongoing across text, images, music and code, and settlements and licensing deals are being signed in parallel, which tells you the industry does not regard the answer as settled either.

In the EU the route is different. There are exceptions for text and data mining: a broad one for research organizations, and a general one that allows mining lawfully accessible content unless the rightsholder has expressly reserved their rights in an appropriate machine readable way. That reservation is the opt out. It shifts the burden onto creators to say no, and it only works if the reservation is expressed in a form a crawler can read, which in practice means site level signals rather than a line in your bio. The AI Act adds a compliance layer on top by requiring model providers to have a copyright policy and publish a summary of training content, explained in what the AI Act asks of general purpose model providers.

The UK has only a narrow exception for non commercial research and has consulted repeatedly on a broader one without landing. Practical effect for you: how a model was trained is a fact about the provider, not something you can fix, so your protection comes from indemnities and from what you publish.

Does the output infringe?

Usually not, and the reason is mechanical. A model stores statistical relationships, not files, so typical output is a recombination rather than a copy. Infringement needs substantial similarity to a specific protected work, and most generated text or images do not reach it.

The exceptions are real, and they cluster:

• Memorization. Works that appear many times in the training data can be reproduced closely, especially famous song lyrics, poems, well known code snippets and iconic images. Prompting toward a specific work makes this dramatically more likely.
• Characters and franchises. A generated picture of a famous cartoon character is an infringing copy of that character, whatever produced it. The same applies to logos and trade dress, which bring trademark law in as well.
• Style versus substance. Copyright does not protect style, so a picture in the manner of an artist is not automatically infringing. But naming a living artist in a prompt tends to pull the model toward their actual compositions, and in some places it raises separate claims around passing off, unfair competition or moral rights.
• Voice and likeness. Cloned voices and faces are governed by publicity and personality rights rather than copyright, and several jurisdictions have tightened those rules specifically because of synthetic media.
• Code. Suggested code can reproduce licensed source, and copyleft licenses carry conditions that travel with the code regardless of how it arrived in your repository.

Can you own the output?

In the United States the rule is human authorship. Purely machine generated material is not protected by copyright, and the Copyright Office will not register it. What can be protected is the human contribution: text or art you wrote yourself, meaningful edits and modifications, and the creative selection, arrangement and coordination of AI generated pieces into a larger work. A comic where the pictures are generated but the story, sequence and layout are yours can be registered for those human elements, with the generated images disclaimed. Prompting alone, however long and detailed the prompt, has not been accepted as authorship, on the reasoning that you are describing a result rather than controlling its expression.

Registration is where this bites. In the US you generally need a registration to sue for infringement, and the application asks you to disclose and disclaim AI generated content. Getting that wrong can undermine the registration later.

Elsewhere the picture varies. The UK has a long standing provision for computer generated works with no human author, giving a shorter term of protection to the person who made the arrangements necessary for creation, and its future has been under review for years. Most EU countries require a human author reflecting a personal intellectual creation, so purely generated output is similarly unprotected. In practice this means your AI generated marketing image may be something nobody can stop a competitor from reusing.

The rules that keep you out of trouble

1. Do not name living artists, current franchises or specific copyrighted works in prompts. This single habit removes most output risk, and it is the one most commonly ignored.
2. Check the platform terms. They decide whether you get commercial rights, whether outputs are exclusive to you, and whether an indemnity exists. Those are contract terms, and they belong with the other clauses to check in an AI supplier agreement.
3. Check the indemnity conditions. Where a vendor promises to defend copyright claims on outputs, the promise typically requires a paid tier, filters left on and no prompting toward a specific work. Violate the conditions and the cover evaporates.
4. Verify before publishing. Reverse image search a generated image you plan to use commercially. Check quoted text and any citation, because models invent sources as readily as they invent sentences, which is why made up sources are so common.
5. Keep records. Save prompts, dates, model used and your own drafts. If ownership or originality is ever questioned, contemporaneous records of the human contribution are the evidence that matters.
6. Label where it counts. Some platforms, ad networks and jurisdictions now require disclosure of synthetic media. Embedded provenance data, described in how content credentials work, is becoming the standard way to carry that claim with the file.

If you are the creator being trained on

Your options are practical rather than legal in most countries. You can express a reservation of rights on your site in a machine readable form, which carries real weight in the EU and is a documented signal elsewhere. You can block known AI crawlers at the network edge or in your robots file, list works in a creator opt out registry, and put valuable material behind a login. The mechanics of all of these are in how to keep your work out of training data, and what the training pool actually consists of is covered in where AI training data comes from.

What none of these do is affect models already trained. There is no practical mechanism today to remove a specific work from a finished model, and any service promising it is selling something else.

What to do before you publish

Decide two things for each piece of work. First, how much of it is yours: if the answer is nothing but the prompt, do not rely on owning it, and do not build a brand asset like a logo out of it. Second, how close the prompt steered toward someone else's work: if you named a person, a character or a title, redo it. Then save the prompt, the model and your drafts in the same folder as the final file.

If you publish images at scale, read the safe path for using AI images legally too, since platform rules and disclosure duties change faster than copyright law. Treat the area as live: the direction of travel is toward more licensing deals and more disclosure, not one clean ruling that settles everything.