AI in the creative sector: six practical legal tips

DALL-E makes images, Sora makes videos, Suno composes complete songs and ChatGPT writes your copy. For creators that means a toolbox full of possibilities, but also a stack of questions. May an AI company use your work to train a model? And is what you make with AI actually yours?

In this article I set out six practical legal tips for creators and for those who publish their work. Two concern your own work as training material, the other four what you make with AI.

First, briefly, the starting point. A work is protected by copyright if it has its own original character and bears the personal stamp of its author. It may then in principle only be copied or made public with the author’s consent. The Dutch Copyright Act contains exceptions to that rule, one of which matters most to creators in the age of AI.

What the Dutch Copyright Act says about AI training on your work

Since June 2021, article 15o of the Dutch Copyright Act has contained an exception for text and data mining. Anyone with lawful access to a work may analyse it automatically and make copies for that purpose. AI companies rely on this rule when they train models on what is available online.

The exception does not apply where the author or rightholder has expressly reserved that use in an appropriate manner. For work made available online, the Act gives a reservation by machine-readable means as its example. That is your lever.

Excluding your work from AI training with an opt-out

Tip 1. Do not give AI systems lawful access, or make a reservation.

Work behind a login or a paywall is not lawfully accessible to anyone who circumvents those conditions, so a crawler that gets into your members’ area without an account cannot rely on article 15o.

If your work is freely available online, the reservation is the route. A line such as “All rights reserved” or “copyright reserved” at the foot of your site is probably not enough, because courts look at whether a machine can read the reservation. In the case brought by photographer Robert Kneschke against LAION, the Oberlandesgericht Hamburg held on 10 December 2025 that a reservation in plain language was not machine-readable in 2021. The court expressly left open whether that has changed since.

Against one group an opt-out never works. Research organisations and cultural heritage institutions may carry out text and data mining for scientific research even if you have made a reservation (article 15n of the Dutch Copyright Act). The same Hamburg case showed how broadly that can apply: the court treated the compilation of the LAION dataset as scientific research, even though commercial models were later trained on it. The court gave leave to appeal to the Bundesgerichtshof.

Why a robots.txt file only works if it names the right crawlers

The usual tool is a robots.txt file, a small text file on your server with instructions for crawlers, the bots that scour websites. For each bot you indicate which parts of your site are off limits.

How precise that has to be became clear in the Netherlands in the case of DPG Media and Mediahuis against HowardsHome, a service that searched their news articles automatically. The publishers had excluded a number of AI bots in their robots.txt, such as GPTBot, ChatGPT-User and CCBot. HowardsHome’s bot was not among them. On 30 October 2024 the Amsterdam District Court therefore held that copyright had not been reserved in an appropriate manner and that HowardsHome was entitled to rely on article 15o (ECLI:NL:RBAMS:2024:6563).

Three things follow from that:

• Name every crawler you want to keep out in your robots.txt, and keep that list up to date. New bots appear regularly.
• You can block all bots at once, but then you also disappear from Google. Google has a separate token, Google-Extended, which lets you exclude training of Google’s AI models without dropping out of the search results.
• Record your reservation in a second way as well, for instance with the TDM Reservation Protocol, a technical standard that lets a website state in a single line that text and data mining is reserved, and with a clause in your website’s terms of use.

There is no legal obligation to follow robots.txt. But anyone who ignores your reservation can no longer rely on the text and data mining exception, so copying your work is then an infringement. In addition, providers that signed the European code of practice for general-purpose AI models (July 2025) have committed to using crawlers that follow robots.txt and to not circumventing paywalls.

Evidence that an AI model was trained on your work

Tip 2. Gather evidence if you suspect your work has been used.

An AI model is a black box: you cannot tell from the outside what went into it. Because many providers are also reticent about their training data, evidence is the hardest part of any case.

That is changing. Since 2 August 2025, providers of general-purpose AI models have had to publish a public summary of the content they trained on (article 53(1)(d) of the AI Act). The European Commission has drawn up a mandatory template for it. For models already on the market before that date, the obligation applies by 2 August 2027 at the latest. Such a summary does not list individual works, but it does name, among other things, the main websites from which data was collected. It is a start.

In the meantime, record what you see yourself. If your style, character, melody or text recognisably reappears in what a tool generates, keep the prompt, the output, the date and the name of the tool. In the case of GEMA against Suno, that was exactly the evidence: recognisable melodies in generated songs, even though the prompts did not ask for them. Read more in our discussion of the GEMA judgment.

Proceedings are under way outside Europe too. The New York Times has been suing OpenAI and Microsoft since December 2023 over the use of its articles to train ChatGPT.

When a prompt is protected by copyright

A question I hear regularly is whether a prompt can be protected by copyright. A prompt is the instruction you give an AI model to obtain particular output. There is no single answer. If a court sees a prompt as an idea, it is not protected, because ideas are free. But if the idea is worked out in sufficient detail, contains enough human creative choices and is not too banal or trivial, a prompt can be protected.

“A rabbit in a top hat” is too banal. But if you spend months writing long, creative prompts for a comic book, those prompts may well be protected. Someone else may then not publish or copy them without your consent, as prompts or otherwise.

What you make with AI is not automatically yours

After the input of AI systems in tips 1 and 2, the next tips concern the output. Here too I am often asked whether the output of an AI system can be protected by copyright. The answer follows from the ordinary rules. Output is protected where it is sufficiently creative and original, and those creative choices must have been made by a human.

Enter the prompt “a rabbit in a top hat” into a generative AI system, and DALL-E produces this image (at least, it did for me):

That image is not protected by copyright, because no human creative choices were made in generating it, only artificial ones. It makes no difference that a human gave the prompt, or whether that prompt is itself protected.

A court in Prague reached the same conclusion in October 2023 about an image made with DALL-E. The claimant argued they were the author because the image was based on their own instructions, but could not substantiate that sufficiently. The court did not close the door entirely: with better evidence of creative prompts, the outcome might have been different.

How to work with AI and still obtain copyright

Tip 3. Collaborate with the machine and do not let AI do all the work.

Copyright in your work means you decide who may copy or publish it, and that you can ask a fee for doing so. If someone makes your work public without consent, you can take action. Output generated entirely by AI does not give you that position, however creative your prompt.

If you do want that protection, add something yourself. Draw a field of flowers around the rabbit, give it a pair of glasses or design a pattern on the hat. The more of your own editing, the greater the chance that copyright subsists in the result.

The comic book Zarya of the Dawn showed how that works in practice. In February 2023 the US Copyright Office refused protection for its images, which were made with Midjourney. The text of the book and the selection and arrangement of the images remained protected, because that is exactly where the human contribution lay.

Source: Creator gets no copyright in comic made with artificial intelligence | de Volkskrant (in Dutch)

There are tools, such as Copyleaks, that estimate whether a text or image was made with AI. They give a probability rather than certainty, but a client can still use them to ask questions. Using AI for an assignment is not a problem in itself, but if nothing was agreed or it was not the intention, it can lead to a dispute (see tip 6).

Proving your own input when someone disputes your copyright

Tip 4. Gather evidence of your collaboration with AI.

If someone calls your copyright into question, you need to be able to show that you made creative choices yourself. Record the steps in your creative process and keep sketches, notes, intermediate versions and the prompts you used, so that you can show the edits are yours.

What the Munich court asked of creators →

Can you simply use AI-generated images

Tip 5. Check your output before you publish it.

Not without more. AI output can contain recognisable parts of existing work. In the GEMA case, melodies from well-known songs reappeared in generated music. If you publish such a result, you run the risk of a claim, not the tool.

So check what you want to use. Search Google for terms relating to the generated work, or upload an image to Google Images or TinEye. That gives an indication of whether someone else’s work is in it, although it is not watertight.

Read the tool’s terms of use as well, because not every tool gives you the rights in the output, and some do so only on a paid plan.

The disclosure obligation for deepfakes and AI text under the AI Act

Since 2 August 2026, article 50 of the AI Act has applied. If you publish a deepfake, you must disclose that the image or sound has been artificially generated or manipulated. The same applies to AI-generated text that informs the public on matters of public interest, unless it has undergone genuine human editorial review. For evidently artistic, creative, satirical or fictional work, the disclosure may be made in a way that does not hamper the display of the work, for instance in the credits.

How broadly the European Commission interprets the concept of a deepfake, and which exceptions apply, is set out in the transparency obligation under article 50 of the AI Act.

Arrangements on AI in your contract with a client or creator

Tip 6. Think about the use of AI and record it in your contract.

To my mind this is the most important tip. Some questions to ask yourself:

1. May AI be used? Decide, as client or contractor, whether AI may or even must be used. If a client does not want it, the contract can say so. A contract between a film producer and an actor can, for instance, state that the actor may not be replaced by AI-generated imagery. As a creator you may also have an interest in AI not being made compulsory, because that can be reflected in your fee. That is what the screenwriters in the US feared, which is why they had it written into their collective agreement that a writer can never be required to use AI.
2. On what conditions? If AI may be used, the client can attach conditions, for instance that AI may only serve as inspiration, or that the end result must always contain human creative choices (see tip 3). Copyright protection then remains possible. You can also agree that AI may only be used to repair poor recording quality.
3. Who is liable? Liability for copyright infringement often rests with the creator. But who bears the risk if the client makes AI compulsory? If the fee is lower because AI use is compulsory, it makes sense to share liability or place it with the client. The loss if it turns out that no copyright subsists in the result, because AI did all the work, belongs in that allocation too.
4. Does the creator want an opt-out? A creator may, for whatever reason, not want their work ending up in AI systems. Have the contract provide that the exploiting party makes a machine-readable reservation on publication (see tip 1).
5. Who discloses the AI use? Since 2 August 2026 the disclosure obligation under article 50 of the AI Act has applied. As a client you will want to know whether your contractors used AI, and to what extent. Agree a disclosure duty for the creator.

Concrete model clauses for these questions can be found in our article on updating the author contract for AI.

Do you have questions or need help drawing up or amending contracts in which AI plays a role? Get in touch or call 020 675 88 21.

This article was written in collaboration with Harm Boon. It was first published in April 2024 and revised in October 2026: the explanation of the opt-out, the case law and the AI Act has been updated.