AI in the creative sector

The possibilities with AI are endless. Think of DALL-E for generating images, Sora for making videos, Suno for composing complete songs, and of course the well-known ChatGPT. All of this is impressive, but what are the risks and pitfalls for content creators, and for those who publish their work, when using AI? To help them avoid some of those legal risks and pitfalls, I set out some practical legal tips on AI.

In this article I refer a number of times to copyright concepts. So, very briefly, a recap of when a work is protected by copyright. A work may be protected by copyright if it has its own original character and bears the personal stamp of its author. Where a work is protected by copyright, it may in principle only be published and copied with the author’s consent. The Dutch Copyright Act contains a number of exceptions allowing a work to be made public or reproduced without the author’s consent. In this piece I focus on the exception that is specifically relevant to content creators and publishers in the context of AI.

The Copyright Act already takes AI into account. Since 2021 it has given the developers of AI systems the possibility of text and data mining. This is laid down in article 15n of the Copyright Act and means that, in principle, everything that can be found publicly (and lawfully) may be ‘used’ by an AI system to train itself.

The exception is not unlimited, however. So you can make sure your work stays out of AI’s clutches.

Tip 1: set conditions and/or add an opt-out

If you want to prevent AI systems from using your content, do not give them lawful access to it. You can do that by using a login system with conditions, for instance, where certain material only becomes visible after logging in, or by using a paywall. If something or someone then accesses the website without meeting those conditions, that access is not lawful.

Another tip for keeping your content out of the clutches of commercial AI, should you wish to, is to add an opt-out, also known as making an express reservation.

Note that you cannot rely on an opt-out against non-profit scientific research organisations or cultural heritage institutions. Against them you cannot prevent AI from ‘using’ your content.

There are various ways of opting out.

Examples of a simple opt-out are: “All rights reserved”, “no datamining” or “copyright reserved”. Online, that reservation must be machine-readable. Precisely what counts as machine-readable also depends on how clever the reading machine is. Exactly how this should be done is still unclear. Hopefully clearer guidance will come from Europe before long.

To increase the chance that the system can read your opt-out, you could also add a “next-level” opt-out. Here are a few tips on how to implement that on your website. First, you could include terms such as “noAi” in your website’s metadata.

You can also set up a robots.txt file. This is a small text file containing guidelines and instructions for what are known as search engine ‘crawlers’, also called bots. Those bots scour websites and store what they find in a database. A robots.txt file tells search engines which of a website’s pages should be included in, or skipped from, their search results. You can use it to try to prevent content on the website from being used by AI systems for training purposes.

Be aware that search engines are not legally required to observe those robots.txt files, so they can simply ignore them. It is worth knowing that OpenAI, the company behind ChatGPT and DALL-E, has already indicated that it will respect an opt-out in robots.txt files.

In the Netherlands, Mediahuis and DPG Media use such a robots.txt opt-out file, and internationally CNN and the New York Times do so.

Another opt-out option is to implement specific protocols and tags in your website’s HTML. These tags make clear to AI systems that the information may not be harvested.

Tip 2: gather evidence that your content has been used by an AI tool

If, despite following tip 1, you suspect that your content is being used by an AI tool and you want to do something about it, gather evidence.

An AI system is something of a black box, in which the input and the activities of that system are not visible. Many AI tools are anything but transparent about disclosing their training data. That makes it difficult to find out whether your content is inside the black box and, if so, how it is being used.

Fortunately there are now a number of websites that check whether your content sits in an AI database and is being used to train AI. One example is https://haveibeentrained.com/. Unfortunately that website does not yet cover the databases of every AI system. So if your content cannot be found there, that does not rule out that some AI system has used it.

The New York Times also discovered that ChatGPT was training on its content, and was not happy about it. At the end of December 2023 the paper therefore brought proceedings against OpenAI, the company behind ChatGPT, and Microsoft. The New York Times argues that the tech companies infringe its copyright by using its articles without permission to train chatbots such as ChatGPT.

In the US, more and more copyright holders are bringing proceedings against companies such as OpenAI for training on their copyright-protected content. In the Netherlands no such case has come before the courts to date. The journalists’ union NVJ expects, however, that the first Dutch case will not be long in coming.

Prompts

A question I hear regularly is whether a prompt (an instruction given to an AI model to generate particular output) can be protected by copyright. There is no single answer. A prompt could be seen as an “idea”, and ideas are not protected by copyright. But if an idea (a prompt) is worked out in sufficient detail, contains sufficient human creative choices and is not too “banal and trivial”, then a prompt could well be protected by copyright. The prompt “a rabbit in a top hat”, for instance, is too banal and trivial and therefore not protected. But suppose you make a comic book with the help of AI, and spend months devising creative and lengthy prompts: those prompts might well be protected by copyright. That would mean someone else may not publish or reproduce them, as prompts or otherwise.

Output

Where tips 1 and 2 concerned the input of AI systems, the tips that follow concern the output. Another question I hear regularly is whether the output of an AI system can be protected by copyright. The Copyright Act applies to AI output too. That means the output is protected by copyright where it is sufficiently creative and original. One requirement for those creative choices is that they must have been made by a human mind.

Suppose you enter the following prompt into a generative AI system: “a rabbit in a top hat”. DALL-E then produces the following image (at least, it did for me):

The image that ‘rolls out’ of the AI system as output is not protected by copyright, because no human creative choices were made in generating it. The AI system makes only artificial choices. Note that this concerns only the question whether the output is protected. It therefore makes no difference whether the instruction (prompt) was given by a human and/or whether that prompt is itself protected by copyright.

Recently, for the first time in Europe (in the Czech Republic), a court ruled that an image generated by an AI tool (specifically DALL-E) could not be protected by copyright because no creative choices had been made by a flesh-and-blood person. The claimants’ argument that they were the authors of the image because it had been created on the basis of their specific instructions (the prompt) did not help them either, because there was insufficient evidence for it. What is interesting about that reasoning is that the Prague judge did not close the door entirely. So perhaps, had the claimants adduced sufficient evidence, they could have claimed copyright in the generated image on the basis of their creative prompts.

If you generate content with AI and you want that content to be protected by copyright, then:

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

It can matter a great deal that your content is protected by copyright. You then have the exclusive right to exploit it. Any form of copying or publishing your content requires your consent as its author (or that of the other authors as well, if you worked together). And you can ask a fee for that consent. If someone makes your copyright-protected content (or content that closely resembles it) public without consent, you can take action.

Earlier in this article I wrote that where output is generated entirely by an AI system, even if a human entered a creative prompt, that output cannot be protected by copyright. If you nevertheless want the output protected, here is a tip: collaborate with the AI. So, taking our rabbit as the example, add a field of flowers you have drawn yourself, add accessories such as a pair of glasses, or draw a pattern on the hat. The more editing you do, the greater the chance that copyright will subsist in the content.

Another legal debate about whether content made with AI can be protected by copyright, admittedly in the US, has already taken place in the creative world of comics. An American comic-book creator had created the character Zendaya entirely with AI. The US Copyright Office held that this was not protected by copyright, because the character had been created entirely with AI. The creator argued that she had worked for years on the right prompts, but that did not help her.

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

Since more and more (legal) disputes appear to be arising about copyright in content created by AI, there are tools such as Copyleaks that can show that content was ‘made’ by an AI system and not by a human.

So be aware that tools of that kind can expose your use of AI. That is no problem where you were permitted to use AI for an assignment, but where that is not the intention, or where nothing has been agreed about it, I can imagine it leading to difficulties. More on the contractual arrangements that can be made about this below.

Tip 4: gather evidence of your collaboration with AI

Should someone subsequently call your copyright into question, gather evidence that you worked with the AI tool and that you made sufficient edits to the output.

You can do that, for instance, by recording the various steps in the creative process. Sketches and notes can help to show that human creative choices were made.

Tip 5: check your output

If you use AI, you will of course not want a claim on your hands, so tip 5: check your output. If you have ambitious plans for your AI output, be aware that it may infringe someone else’s (copy)rights.

To reduce the risk of a copyright claim when you use AI-generated content, it is sensible to carry out a check. One way of doing so is to search Google for terms relating to the AI-generated work, or to upload that work. One way of doing that with an image is to upload it to Google Images or TinEye. That could give an indication whether the work contains copyright elements belonging to a third party and whether there is therefore a risk of infringement. I should point out that these methods are of course not watertight.

The AI Act

In a legal article about AI, the AI Act cannot of course go undiscussed. The AI Act is European legislation which mainly lays down obligations for providers of AI systems. The text of the AI Act will be finally published at the end of April or in May 2024, but a version has already been leaked. For deployers, developers and users of AI systems, all that follows from that leaked version is a transparency obligation for deepfakes and informative texts. That obligation will apply one year after the AI Act enters into force. The other provisions of the AI Act mainly concern obligations for providers of AI systems. I will address those in a subsequent article.

Once the period for the transparency obligation begins, take care when using deepfakes and when using informative texts. It follows from the AI Act that a deployer must state clearly that a deepfake or an informative text has been made with AI. The AI Act defines a deployer as: a natural or legal person, public authority, agency or other body using an AI system under its own responsibility. And, for clarity, the AI Act’s definition of a deepfake: where audio, image, video or text content has been artificially generated or manipulated (with the help of an AI system).

For (evidently) artistic, creative, satirical or fictional use of deepfakes, the transparency obligation is somewhat less strict. The disclosure may be made in an appropriate manner “that does not hamper the display or enjoyment of the work”. Exactly how that will have to be given shape is not yet clear. I can imagine that a mention in the description of a YouTube video or in the credits of a film will be enough.

Incidentally, for personal, non-professional use of a deepfake there is no obligation to disclose that it is a deepfake.

Nor does this disclosure obligation apply where informative texts generated with AI (specifically: text that informs the public about matters of public interest) have undergone a process of human review or editorial control and where a natural or legal person bears editorial responsibility for publishing the content. So if you do not want to disclose that you wrote a news article with the help of AI, tip 3 comes in handy again: collaborate with the machine.

N.B. The disclosure obligation for deepfakes and informative texts also does not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences.

Update: the AI Act is now in force

This article dates from April 2024 and describes the AI Act as it then stood, on the basis of a leaked version. The AI Act has since been adopted and entered into force on 1 August 2024, with its rules applying in stages. The transparency obligation discussed above is set out in article 50 of the AI Act and has applied since 2 August 2026. In July 2026 the European Commission published guidelines and a Code of Practice to go with it. What that means in concrete terms for deepfakes, AI-generated texts and chatbots, and which exceptions apply to creative work and to content with human editorial oversight, is set out in AI content must be recognisable. The remainder of this article, on copyright, opt-outs and contractual arrangements, remains fully usable.

Tip 6: think about the use of AI and record it in a contract

A final tip, and to my mind a very important one, tip 6: think about the use of AI and record it in a contract.

Some questions you can ask yourself:

1. May AI be used? Consider whether, as client or as contractor, you want AI to be permitted, or even required, at all. If as a client you do not want a contractor to use AI, you can record that. For instance in a contract between the producer of a film and an actor, recording that the actor may not be replaced by AI-generated imagery. As an author, too, you may have an interest in the use of AI not being made compulsory, because that might be reflected in your fee. That, at any rate, was what the screenwriters in the US feared. They headed off that concern by having it recorded in their collective agreement that an author can never be required to use AI.
2. If AI may be used, on what conditions? If you agree that a contractor may use AI, the client could attach certain conditions. Think of a provision permitting the use of AI, but only for inspiration. Here I refer back to tip 3: collaborate with the machine. Agree, for instance, that where AI is used, human creative choices must still be made. That way you keep open the possibility of copyright protection in the content. You could also agree that AI may only be deployed to repair poor recording quality.
3. If AI may be used, what about liability? An important contractual provision to think about when using AI is the one governing liability. It is fairly common in contracts for liability for copyright infringement to rest with the author. But if the client makes the use of AI compulsory, who is then liable for any infringements? I can imagine that where the author’s fee is lower because of the compulsory use of AI, liability for an infringement is shared or even rests with the client. Consider too any loss arising if it turns out that no copyright at all subsists in the generated content, because the author let AI do all the work and made no creative choices.
4. Does an author have an interest in an opt-out? An author may, for whatever reason, not want their content ending up in AI systems. As an author you could then have it recorded in the contract that opt-outs are to be added to the published content.
5. Prepare for the transparency obligation under the AI Act. With the AI Act coming, it can be useful as a deployer or client to know whether your contractors, for instance the authors of informative texts, have used AI and to what extent. As I wrote earlier in this article, the transparency obligation for informative texts generated with AI does not apply where human review has been involved. You could agree a form of disclosure duty for the author.

Do you have questions, or do you need help amending or drawing up contracts in which AI plays a role? Please get in touch.

This article was written in collaboration with Harm Boon.