A necessary update to the author's contract in the age of AI. Three tips.

The AI landscape is slowly taking on colour. Details are being filled in. Through a strike by the screenwriters in the US and the compromise that followed, through conversations with media companies in the Netherlands, and through a stream of publications.

It sometimes puts me in mind of a time-lapse shot of a landscape: an empty plain that, beneath racing clouds, quickly turns into a flowering idyll where the bees hum and bright butterflies flutter about. But the film could end differently too: with a view of a desolate plain in which the polar wind of AI chills the last life to the bone.

Which way it will go cannot be said. That authors, publishers and producers have to do something about AI, in their contracts too, is unavoidable. In this article I offer a number of tips leading to a number of model clauses. The authors I have chiefly in mind here write texts or screenplays. In principle the clauses can be applied in any contract between an author and the party wishing to exploit the author’s work.

1. use of AI

Parties must make clear how the actual use of AI is to be handled. There are several flavours. Parties can choose to allow the author to use AI. For inspiration, for instance. Such a clause might then read: “The author may use work created by AI, provided that the author does so only for inspiration.” That provision needs some further explanation, I think. In my view “for inspiration” means that individual elements from the AI output, or perhaps an underlying idea, may be used, but not that the author may take over entire parts of the AI output that would in themselves be protected by copyright. It therefore makes sense to add a further sentence: “To that end the author shall endeavour not to take over any parts of work created by AI which would be works within the meaning of the Copyright Act if they had been made by a human author.”

Parties can of course also agree that the author may not use AI. The question is what exactly the purpose of such a provision is. If the purpose is to produce a pure creation, one that comes solely from the author’s mind without being shaped in part by AI, one should bear in mind that every author draws inspiration from the work of others. Every person, and their thoughts, is the sum of the input of others. Why AI cannot be part of that too (being, ironically, itself the output of the input of many) is not clear to me. A ban on using AI is not rooted in a logic I can follow.

The mirror image of this is a possible obligation on the author to use AI. The discussion about that goes to the heart of the author’s work. One can imagine that using AI leads to faster delivery and thus, because it takes the author less time, to a price-quality ratio more favourable to the client. This is a slippery slope with an unknown destination. A more favourable price-quality ratio means, in essence, a lower fee for the author. Where the work shifts from writing to editing AI work, the total fee falls dramatically in any event. For the publisher or producer there is the risk that, given the compulsory use of AI output, the author may be tempted to lean too heavily on AI. The author then delivers a work that smells of AI and may not be protectable against being copied by third parties. After all: what the AI has written is not protected by copyright.

The screenwriters saw the risk to their fees as well, and agreed in their collective agreement that an author (screenwriter) can never be required to use AI. There are other reasons for not making it compulsory, too: the use of AI is experienced as an attack on a creative process that is so much a part of the screenwriter’s work that it amounts to an attack on the profession itself. And protecting that profession is exactly what their union was founded for.

Where parties do not wish to oblige the author to use AI, they can include: “The author shall never be obliged to make use of AI.”

2. copyright and AI

The author’s work can be used to train an AI. This too was a reason for the American writers’ union to strike: the fear that their members’ scripts would be used to train an AI that would then make them redundant.

Parties can make arrangements about the use of the author’s work for AI training. It should be borne in mind, though, that an AI which lawfully obtains access to the author’s work needs no permission to use that work to train itself. Put differently: it is not the publisher or the producer who grants a licence to the AI; it is these parties who must take action to prevent the AI from training on the material.

Where parties do not want the work to be used for data mining and AI training, they must say so unambiguously alongside the work. That means that where the author does not want their work used for AI training, they must agree with the publisher or producer that the latter will make the relevant statement each time. All in accordance with article 15o of the Dutch Copyright Act. And that, on request, they will enforce a prohibition on that use through the courts where an AI does not observe it. In my view, a reasonable cost-benefit assessment may be made before going to court, taking into account the prospects of success and the evidential position.

For example:

“The publisher/producer shall ensure that it reserves copyright in accordance with article 15o of the Copyright Act, and shall impose that obligation on each of its successors in title.”

“In deciding on enforcement in the case of a copyright infringement connected with data mining, the publisher/producer shall weigh the significance of the infringement against the costs of enforcement, taking into account the prospects of successful enforcement. Where there is the possibility of having a claim brought partly on behalf of third-party rightholders, the publisher/producer shall make as much use of that as possible.”

3. allocation of risk and indemnities with AI

Now we come to a somewhat thornier subject. Where does the risk lie if something goes wrong and the author’s work infringes the rights of others? In many of the current standard agreements the author is liable and indemnifies the publisher or producer in case their work is not original and turns out to infringe that of a third party. That is logical enough: deliberately taking over another’s work in such a way as to infringe that person’s rights should remain at the author’s expense. That does not mean, incidentally, that unconscious copying leading to a copyright infringement is impossible. It is just that in practice the risk of it is not that great, in my view. That is to say: recurring motifs are often to be found in authors’ works, a kind of fashion in subject matter, which can look like unconscious borrowing; but those shared motifs seldom if ever lead to a real infringement of another’s copyright. After all, as a reminder, the underlying ideas are not protected by copyright, and it is often those unprotected ideas and motifs that unconsciously find their place in the work of others.

In this age of AI, none of that is so simple. Certainly not where parties agree that the author may deploy AI. An AI is unpredictable and may very well reproduce large parts of someone else’s work without the author knowing. And without the AI doing so deliberately either: an AI is inscrutable and a good deal more besides, but it has no consciousness.

It seems to me, therefore, that the standard arrangement placing the risk of copying on the author needs to be qualified. It also does no harm for the author to keep the prompts and the output of the AI, so that the author can show what the AI produced and that there was therefore no deliberate infringement.

For example: “The author is liable where their work infringes the rights of others. Where the author uses AI, the output of the AI is deemed not to infringe the rights of others. Where that output nevertheless does so, the author is not liable for it.” Parties can of course also agree a limitation of liability. I would consider it reasonable (and in practice we often see this) for it to be limited to the amount of the assignment, for instance.

A further point here is that where parties agree (in line with the first model clause) that the author must always edit the AI output so as not to take over any “work” of the AI, there can in principle be no infringement by the AI of third-party rights. Exactly how the relationship between those two clauses should be shaped is a matter for negotiation. I am inclined to say that I would consider it reasonable for this clause to take precedence over the obligation to adapt the AI output so that the author takes over no “work”. Above all because determining where the line between an idea and a work lies is very complicated. That means the first model clause should be preceded by the words: “Without prejudice to the provisions of article XX (reference to this model clause), etc”.

In a further article I will go into the role of the GDPR and privacy in the relationship between contract and AI.

 

noAI

Not to be used for data mining

Copyright reserved