Artificial intelligence and EU copyright law: a net of authorship claims - 19 October 2022
Blog by Sarthak Babbar and Anna Christina Stuart Aguiar, LLM (International Commercial Law) students
Human-authored art has always been at the heart of copyright, but what happens when technology, or to be more precise: artificial intelligence (AI), seeks to replace the artist? Is the final output still protected by copyright?
Dr Alina Trapova explained that when she set out to do her Ph.D. in 2017, the big question in copyright revolved around ´´Who is entitled to commercialise the art created by AI?”
It is important to note, as Dr Trapova referred, that this complicated issue has been discussed not only by many scholars but also by some of the relevant institutions (both EU and international). Even some legislatures have started to act upon it, though the vast majority are still in the preparatory phases. This evidences that this current and relevant subject has started to permeate in the different layers of academic and public debate.
The copyright discussion was fuelled by “The next Rembrandt” project. An AI machine that was designed to recreate a computerised painting that matched, as the name suggests, the work of 17th century Dutch artist Rembrandt Harmenszoon van Rijn. Dr Trapova mentioned that this project was met by a high amount of scepticism from some members of the art community and the broader public, as it led to legal and ethical questions involving art, technology and commerce. Nonetheless, the human involvement in the project was so excessive here, that the ‘autonomy’ of the machine learning process seemed to be highly exaggerated.
The issue was bifurcated by Dr Trapova into two key points. First, understanding the creation process of this “art”, by essentially seeking to identify a human author in the creative process and second, determining if this human author’s creation constitutes “free and creative choices” amounting to their “own intellectual creation” that leads to an objective and clearly identifiable work.
Essentially, Dr Trapova rightly raises the subsequent questions of whether AI generated works are protected by copyright; and if not, should they be. It is relevant to note that the answers for these were circumscribed to the final product (output) and EU copyright law (as opposed to any related or other intellectual property rights).
In an attempt to answer these questions, Dr Trapova set out to dissect the machine learning process and determine if the benchmark for copyright protectability under EU´s case law was met. However, she came to the conclusion that there is no straight jacket formula, meaning that a case-by-case analysis is needed depending on the particular characteristics of the issue at hand. Nonetheless, Dr Trapova managed to establish certain criteria that must always be met for a valid copyright claim to subsist, i.e., ‘three protectability benchmarks’:
- Designation: human authorship
- Subsistence: originality standard, which constitutes the author´s intellectual creation - this turned out to be the most difficult (and equally, the most important) part of her analysis
- Objectivity: the final work needs to be clearly identifiable
Dr Trapova researched the work done by Harvard University’s Cyberlaw Clinic at the Berkman Klein Centre, a legal clinic that provides advice and counselling to artists on a pro-bono basis. From its experience, the centre was able to identify the four stages (pillars) in the generative art systems:
- Input
- Learning algorithm
- Training algorithm
- Output.
For each stage, Dr Trapova identified the key elements used by the artist and AI systems in order to create an output: final product, where the question regarding authorship arises. Dr Trapova further explains the intricacies of each stage, focusing on the two most technical steps of the whole system: the algorithms. As she states, the learning algorithm operates all the inputs to identify their relevant characteristics. The second term, the training algorithm, extracts all the learning and information from the previous stage in order to produce an output. According to Dr Trapova, the training algorithm is the most important stage of these new machine learning techniques. Most of the times in generative art, neural networking techniques have been employed in that pillar. Such training algorithms are fundamentally based on two features: architecture and neurons connected by weights.
As Dr Trapova illustrates, the architecture is a parameter which is developed in advance and coded depending on the needs of the user, commonly referred to as the “hyper-parameter”. The weights that connect the neurons take their idea from the functioning of the brain which uses millions of neurons to pass information from one part to another. Unlike the architecture, the neurons are capable of evolving through the training process. The weights are in the form of a numeric value, while the neurons are mathematical functions. This is where the legal question of copyright comes into place. Dr Trapova explains that the three benchmark requirements needed to be mapped onto the four pillars individually in order to determine if they were protected by copyright – at each stage and in the final output.
After her analysis, Dr Trapova concluded that given the particularities of these neural network-driven systems, the art generated by them is generally not subject to copyright protection under EU law. It is evident that the detachment trend between the human author and final creative work continues to grow, as increasing layers of technology appear to stand between them. Dr Trapova states that, at the end of the day, copyright aims to protect human creations. Therefore, attempting to have artificial intelligence (based, broadly speaking, on mathematical algorithms) be the creator of art, contradicts with the roots of copyright law. Humans are not mathematical, as emotions, mistakes and imperfections are involved in their creative process and that is where the subsistence of copyright lies.