AI and copyright: should authors be paid?

Should the authors whose texts or songs were used to train an AI be paid? On AI and copyright, the debate pits two images against each other: plunder, and the reader who learns. The law, for its part, moves forward through a finer distinction. Learning from an accessible work is largely allowed in Europe, unless the rights holder has opted out. Memorising and reproducing a work, on the other hand, is starting to be judged as reproduction.

For a company that publishes content, the practical question is not a philosophical one: should you opt out of training, and what do you lose in visibility in answer engines if you do?

Position 1: training without paying is plundering

Authors and the societies that manage their rights make a simple argument. Language models were built from a huge body of protected works, without permission or payment, and they then compete with those same works: a generated summary can replace the article, and a generated image can replace the work commissioned from an illustrator.

This position has won an important first round in Germany. On 11 November 2025, the Munich I Regional Court ruled largely in favour of GEMA, the German society for the collective management of music rights, against two companies of the OpenAI group (case 42 O 14139/24). The dispute concerned song lyrics written by nine well-known German authors. The court found that these lyrics were memorised in the models, by comparing the texts in the training data with what the models reproduced. Its reasoning distinguishes three stages:

The court states that the ruling is not final.

Position 2: learning is not copying

On the other side, model providers and some legal experts reply that a model is not a library. It learns statistical regularities from billions of texts, just as a person who has read a lot later writes in a style shaped by that reading, without paying every author they came across. That is the defence OpenAI presented in Munich: according to the court’s press release, its models do not store or copy specific training data but reflect in their parameters what they have learned from the whole. Requiring a licence for every public text would make training impossible, except for the few players rich enough to negotiate with every publisher, which would reinforce their dominance.

This side also has legal backing in Europe. The 2019 copyright directive provides, in its Article 4, an exception for text and data mining on works that are “lawfully accessible”. Learning is therefore not forbidden in principle.

What the law already says about training

The rights holders’ opt-out. The same Article 4 of the directive sets the limit, in paragraph 3: the exception applies on condition that the use of the works “has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online”. A publisher can therefore say no.

The obligation on model providers. Article 53 of the EU AI Act requires providers of general-purpose AI models to put in place a copyright compliance policy, in particular to identify and respect that reservation of rights, and to publish a sufficiently detailed summary of the content used for training. These obligations have applied since 2 August 2025; models already on the market before that date have until 2 August 2027 to comply (Article 111).

What we think

We think the distinction taking shape is the right one. Banning all learning from public texts would freeze the technology for the benefit of those who can afford to buy everything. But a model that spits out a song or an article word for word is not “taking inspiration”: it reproduces, and there is no reason for it to escape the rules that apply to everyone.

On payment, we are more cautious.

This mechanism really benefits only players able to negotiate a licence in exchange for withdrawing their opt-out. A small publisher or an SME that publishes a blog does not have that clout: for them, the only real question is choosing between opting out and visibility.

What would change our minds is a final decision, on appeal or before the Court of Justice of the European Union, that held training itself to be a reproduction, regardless of any memorisation.

What it changes in practice for a company

1. Decide consciously, not by default. Opting out of training is done mainly through machine-readable rules, including the robots.txt file, which tells crawlers what they may read. Before blocking, check in each AI vendor’s documentation which bot is used for training and which for search: they are not always the same. Also check the file your site actually serves: it is one of the first checks of a serious audit, as we detail in our guide to what an AI SEO agency should really deliver.

2. Measure what you lose. Content that no crawler can read will not be cited by an answer engine. We explain what gets a site picked up in our guide to how to get cited by ChatGPT and answer engines. For a site that lives off its visibility, blocking everything is rarely the right choice; for a site that sells its content, the opposite is often true.

3. Distinguish what you protect. Nothing forces you to treat the whole site the same way. Commercial pages are better off being read; premium content or a proprietary database can be reserved.

4. Do not confuse the files. The llms.txt file is not for opting out of anything: it is a table of contents for agents. We detailed it in our article on what the llms.txt file really does.

5. Check your own use of AI. If you publish generated texts or images, keep a record of the human intervention that transformed them. That is what counts when asking whether a work made with AI can be protected, as we explain in our analysis of AI creativity and the protection of generated works.

Where to start

In October 2026, the line between learning and copying is not fixed, and the Munich ruling is not final. In the meantime, the decision that is yours to make is an editorial one: which pages to open to training and search bots, and which to reserve. Start by rereading the robots.txt file your site actually serves, then decide bot by bot. If you choose visibility, our offer to get found on Google and cited by AI fixes whatever stops their bots from reading the site before writing anything at all.

Frequently asked questions

Sources

1. Munich I Regional Court (new tab)justiz.bayern.de
2. 2019 copyright directive (new tab)eur-lex.europa.eu
3. EU AI Act (new tab)eur-lex.europa.eu