Willemijn Kornelius is a graduate of Civil Law (Intellectual Property) (LL.M.) Leiden University and a current Legal Research Master student at Utrecht University, the Netherlands. Her research focus lies within the field of law & technology from a European perspective. During her research internship at CEPRI (University of Copenhagen), she is looking into digital platforms and liability for illegal content.
1. Introduction
This year AI Research and Deployment company OpenAI introduced DALL-E 2, an even more sophisticated version of their previous system DALL-E. DALL-E 2 is a system that uses artificial intelligence technologies to generate images from plain texts (combinations of words, so-called “prompts”) and to ‘outpaint’ existing images. Do you ever wonder what the room where the ‘Girl With a Pearl Earring’ of Johannes Vermeer was standing could have looked like? Or what a 1990s cartoon with ‘teddy bears mixing sparkling chemicals as mad scientists’ would look like? Wait no longer! , DALL-E 2 generates these images within a second.
DALL-E 2 is an artificial intelligence system (“AI”). This means that it uses technology to automate tasks normally associated with intelligent human behaviour. DALL-E 2 combines Contrastive-Language-Image Pre-Training with diffusion models which take prompts as input and produce images as output. It is trained on a set of pairs of images and the corresponding texts to describe these images. DALL-E 2 is so advanced, that the only human input necessary to create an image is the prompt. Still, users have pointed out that coming up with a good combination of words still allows for quite some trial-and-error if one aims for a good picture. There is even a guide made with tips and tricks for the perfect prompt. A more technical description of DALL-E 2 is given in this paper of OpenAI.
AI increasingly gets more intelligent. DALL-E’s interpretations of prompts show that it also increasingly gets creative. Creativity is the ability to create something tangible of one’s imagination or original ideas, such as making art and writing books. It is traditionally seen as a unique human feature. The development of AI, such as DALL-E 2, disrupts the fundamentals of the area of law dealing with creativity: copyright. Are these images protected by copyright? Who enjoys the copyright? Should we offer copyright protection? Interestingly enough, OpenAI has these questions as well. On 4 November 2022, they updated their terms of use, giving the user of Dall-E “full ownership rights to the output”. However, it is not clear whether this concerns copyrights as well. Moreover, it is not as simple as that.
2. Copyright and AI-generated outputs
Copyright law offers protection to creators, authors, of creative works, such as art paintings, pictures, books, music and news articles. It does so by giving the authors an exclusive right to exploit their creations. In that way they can prevent others to copy their works without permission and make money by selling them. Art. 2(1) of the Berne Convention, signed by 181 countries, provides that works are “every production in the literary, scientific and artistic domain, whatever may be the mode or form of its expression”. For this blogpost, I focus on artificially generated productions that fall under this definition of “works”, which I define as “artificially generated output”. This form of output could potentially activate copyright rules. The images generated by DALL-E fall under this definition. The question whether the prompt itself is copyright protected is left outside the scope of this blogpost. However, I will show that it could play a role in assessing whether the image is protected.
Under the Berne Convention, copyright protection is initially granted to the author. The Convention does not define “author”, but its text suggests that this is the natural person that put his intellectual labour and creativity in the work the work. However, with systems as DALL-E, human intervention and especially human creativity is not so evident. This year, the US Copyright Review Board refused to grant copyright protection to a machine. In that case (link here), the painting was ‘autonomously created by a computer algorithm’ running on that machine. Legal academics and legislators at international, European, and national level attempt to provide clarity on the question who enjoys copyright in the case of artificially generated output. Their consensus seems to be that, when a human contributes in a sufficiently creative way to the work, he enjoys the copyright. However, they do not want to burn their fingers on the question of protection for ‘purely artificially generated output’, when human intervention is less prominent or even absent. See for example the resolution of the AIPPI in 2019 or the vision in this Commission’s report in 2020.
In this blogpost I approach the unresolved issue of protection of ‘purely artificially generated output’ by comparing two different legal systems. First, I look into the system of the UK. The UK is sometimes described as the “odd one out”, because it is one of the few countries with a specific provision on computer generated output. Second, I look at the Dutch system, as a representative of the EU tradition of copyright protection. Dutch copyright law currently lacks specific provisions dealing with this type of output. Both the EU (in the Commission’s intellectual property action plan in 2020) and the UK (in the government’s response to a call for views on artificial intelligence and intellectual property) have made it their ambition to make their intellectual property systems fit for the future. To tell if they are fit to deal with artificially generated output, it is necessary to understand what protection is currently offered to purely artificially generated output and why we (would) offer copyright protection. I compare the systems to provide insights on this matter, using the example of DALL-E 2.
3. The United Kingdom: producing an incentive for creativity
3.1 Current legal framework
The UK government expresses that copyright is first and foremost an economic tool to reward creativity and stimulate it. This has two sides. On the one hand, the creative investment of the author needs to be protected against unfair competition of people that can copy or use his work freely. On the other hand, copyright protection helps to ensure that the works are publicly available, because authors have an economic incentive to exploit their works.
The UK’s copyright tradition is centred around this economic rationale. How much weight is given to its counterpart, that copyright is a natural right belonging to the creator of a work, is less evident. Generally said, the provisions of the applicable 1988 Act provisions protect ‘original works’, viewed to reflect the personality of the creators.
A copyright is a property right that exists, according to Article 1(1)(a) 1988 Act in literary, dramatic, musical and artistic works (“LDMA-works”) which are original. What is considered to be original is subject to discussion. Traditionally, the originality test was quite low: it merely meant that a work must originate from its creator and must not be copied from another work. According to Section 9(1) the author is the person who creates the work and is, following from Section 11(1), primarily the first owner of the copyright. Through the influence of EU law, when the UK was still a Member State, and the famous Infopaq-ruling of the CJEU, the originality-test seem to have become a bit stricter. Thus, the test is related to the author: the work must show intellectual labour of the author.
3.2 The rightholder of purely artificially generated output
In light of this originality-test, it is not immediately clear who is the author of artificially generated output. The author is generally assumed to be a natural person and thus not a machine. The UK government clarified that to the extent that a person creates a work with the assistance of AI (human intervention is thus present), the work is handled as a normal LDMA-work under the 1988 Act. The originality-test applies: when the human input is not creative enough, no normal copyright protection exists. The artificially generated DALL-E images are the result of a difficult automated process. Although the images are based on the prompt, human intellectual labour is (almost) absent.
However, the 1988 Act additionally provides protection under Section 9(3) to LDMA-works that are computer-generated. Section 178 clarifies that this means that a work is generated by a computer in circumstances such that no human author of the work is identifiable following the “normal” rules. The formal author and thus rightsholder in this case is the ‘person by whom the arrangements necessary for the creation of the work are undertaken’.
One of the only court cases in which this provision occurred is Nova Production v. Mazooma Games (2006). The court found that the frames generated and displayed on a screen by users when playing a videogame were “computer generated works”. The court ruled that the programmer, and not the player, “had devised (…) the rules and logic by which each frame was generated and he had written the relevant computer program” and was therefore “the person who made the arrangements”.
It is generally assumed that artificially generated output falls under this definition and is thus protected by this provision. Applied to the case of DALL-E 2, there are two potential authors: the programmer of DALL-E 2 and the user that came up with the prompt. “Necessary arrangements” of the programmer could be designing the DALL-E software (and thus rules and logic) and deciding on the images and descriptions that formed the training data set. The arrangements of the users are deciding on the combination of words that lead to the actual image and the trial-and-error-process described earlier. In line with Nova Production v. Mazooma, the programmer of DALL-E made more of these relevant arrangements and has thus the best chances to be the rightful author. Arguably, however, there could be co-authorship.
3.3 The doctrinal debate: UK government reviews justification of protection
Section 9(3) offers a solution for authorship on artificially generated output, although questions about its application arise. Not only practical application difficulties exist, but the UK government launched a call for views on the question how the law should react to artificially generated output. It clarified that Section 9(3) was motivated by the hope for more investments in AI. In a follow-up Impact Assessment, the government explained that copyright protection for artificially generated output is only justified when a balance is achieved between the benefits of incentivising the creation of artificially generated output and the costs of a wide protection (users’ access to information and competition). Recently, the government declared that it has not been able to successfully collect evidence for this evaluation and has decided not to change Section 9(3) for now.
4. The Netherlands: exclusivity needs justification
4.1 Current legal framework
The Dutch copyright framework is governed by the Dutch Copyright Law (“DCL”). This law incorporates the provisions harmonised by EU law. Because EU copyright law is harmonised to a great extent, the Dutch system can be seen as a representative of the European copyright tradition. Two main justifications for copyright protection underlie the Dutch framework. The first one is a consideration of fairness. It is considered to be fair that an author enjoys the rights to the work he created and to enjoy potential economic benefits. The second is the utilitarian consideration: copyright protection benefits society as a whole. Exclusive protection promotes the creation of and investments in arts, literature and works. However, the Dutch and European legislator have emphasised that protection is not unlimited: it should not disproportionately hinder the accessibility of information, in light of the freedom of expression. This makes copyright protection a balancing act with the rights of users of the works.
The DCL uses open and flexible concepts, clarified in case law of the Dutch Supreme Court. This Court established that the essence of a “work” fit for copyright protection is its originality. In the landmark case Endstra, the Supreme Court explained that a work must 1) “have an own, original character” and 2) “bear the personal stamp of the maker”. This originality-test is said to align with the originality-test developed by the CJEU in its Infopaq and Painer–judgments: the work must result from the “creative choices of the maker”. These judgments demonstrate the close link between the work and its author in the European copyright tradition.
European law does not harmonise the concept of “author”. However, that the author has to be a human can be derived from the directives and the CJEU’s case law. The anthropocentric focus (human-centred) is most clearly found in the opinion of the AG to the Painer-case, approved by the CJEU: ‘‘only human creations are therefore protected, which can also include those for which the person employs a technical aid, such as a camera’’ (para. 92).
4.2 The rightholder of purely artificially generated output
Since Dutch law does not contain specific provisions for artificially generated output, the normal originality test is applied. As a general rule, an author can use a technical tool as AI to create his work and still enjoy copyright protection, as long as his personal stamp can be identified in the result of the application of AI (“human intervened artificially generated output”). The human intervention with the artificially generated images of DALL-E exists in the words that form the input. This requires a person to make some choices, but the form and the actual presentation of the image lack his personal stamp, because this is decided by DALL-E. In line with the previously mentioned anthropocentric focus, as seen in the AG’s conclusion to the Painer-case, the Dutch (and EU) framework does not seem to accept a non-human author. Presumably, artificially generated output of DALL-E 2 is not protected in the Netherlands.
4.3 The doctrinal debate: justification is disputed
In the Netherlands, the legislator has not meddled in this debate yet. The issue whether copyright protection should be granted to artificially generated outputs is primarily discussed by legal scholars. The discussion centres around the question whether protection would incite the creation of such works. This requires solid empirical economic evidence that the absence of protection damages the European economic welfare, as Quintais and Hugenholtz explain. De Bruin argues that such protection is hard to defend, because autonomous AI technologies will be insensible to financial incentives. Simultaneously, the European Parliament has called upon the Commission to, at least, create uniform copyright provisions applicable to artificially generated output.
5. Protection of DALL-E’s images? Two sides of the sea, two different views
The availability of a provision that applies to artificially generated output in the UK system and the lack thereof in the Dutch system makes it interesting to compare these two systems. The importance given to the identified rationale(s) for copyright protection differ for both systems. In the UK, copyright protection is arguably mostly seen as an economic tool: it produces an incentive for creativity by rewarding creativity. In this way, it serves a public good: the created works become publicly available through exploitation. In the Netherlands, and generally in the EU, the exclusivity of copyright needs justification: two important justifications are the fairness and utilitarian considerations. Protection in the Dutch system should not disproportionately hinder the accessibility of information.
When we turn to the object of this enquiry, protection images generated by DALL-E 2, we see that in the UK, artificially generated output is assumably protected by Section 9(3) when there is no human author. For DALL-E 2, there are two potential authors: the programmer and the user. It is argued that the programmer made the most necessary arrangements, but this is still open for discussion. In the Netherlands, copyright protection is only offered when the work resembles choices of a (human) person during its establishment. A work made with the use of AI is protected as long as the human person made a sufficient and creative contribution to the work and the work bears his personal stamp. DALL-E 2’s images, purely artificially generated output, are therefore presumably not protected.
Ultimately, the question if DALL-E’s creations are protected by copyright raises interesting questions about the reason why we offer copyright protection. The underlying theme in the discussed doctrinal debates is whether copyright protection for purely artificially generated output is desirable and how it should be regulated. I have shown in this brief overview that these rationales lead to different solutions in two countries that are just divided by one sea.