In recent weeks, generative AI platforms, such as Chat GPT-4, DALL.E 2, and Stable Diffusion, have made headlines for a number of reasons. Many have questioned the impact this technology will have on business, education, and the creative sectors, and the potential benefits it may bring. Others have considered the potential risks of the technology and whether the regulatory landscape is suitable to tackle these concerns. One such risk is whether these generative tools are at risk of infringing the work of human creators and their intellectual property rights.
Under copyright law, copyright holders are granted several rights to make use and dispose of creative works as the holder sees fit. In most countries, this includes: the exclusive right to make copies of a work; the ability to prepare derivative works; the right to distribute copies of the work to the public; and the right to perform or display the work publicly.
An AI’s learning/training and generative processes present two key infringement risks to these rights:
- The right to make copies of a work (reproduction right): if the AI learning/training process or the model allowing the AI to function require copies to be made from a data set containing copyright works, the rights of the holders of the data used in these processes may be infringed.
- The right to prepare derivative works (transformation right): if the AI system has been trained using a dataset containing copyright works, there is also the potential risk that the output produced by that AI system will materially resemble, or completely reproduce, parts of those copyright works, therefore potentially infringing the rightsholder’s right to make derivative works.
A right to behold: Risks and exceptions to the use of protected media and data
When seeking to reproduce copyright works, creators and/or trainers of generative AI platforms may find comfort in Article 4 of the EU Digital Single Market Copyright Directive, which has been transposed into domestic legislation across all EU Member States. The provision creates an exception to the restriction of a non-rightsholder’s ability to reproduce copyright works for the purposes of text and data mining (i.e., any automated analytical technique aimed at analysing text and data in digital form to generate information). This means that, unless the rightsholder has reserved its rights through the appropriate means (namely, using machine readable-means, like metadata or terms and conditions), the activity of reproducing copyright-protected materials for the purposes of training the AI would typically be covered by the exception. The effect of these measures allows the rightsholder to opt out of this exception and therefore prohibit the reproduction of their protected works. It should be noted, however, that this exception only extends to the right of reproduction, and not any other rights, such as rights of transformation or communication to the public, which still might be infringed by the output of the AI system.
Creators and trainers of generative AI in the EU and the UK must also be aware that the training phase of the programme could also infringe sui generis database rights, regardless of whether such activity infringes copyright. This issue will typically arise when an entire database (which satisfies the requirements for protection), or a substantial part of one, is copied or re-utilised for the purposes of training an AI system, and/or replicated in the AI system’s ultimate output. In the UK and the EU, the contents of a database may be protected by database rights if there has been a substantial (qualitative or quantitative) investment in obtaining, verifying, and/or presenting the contents of the database. In such cases, extracting, or re-using, the whole or a substantial part of a protected database, without permission, may amount to an infringement of these rights.
While there is some harmonisation of intellectual property rights (e.g. in the UK and EU) as noted above, creators/trainers of generative AI (which generally consists of supranational technology) must also consider the rights of parties at a national level, that do not necessarily reflect the position of other jurisdictions. For example, certain EU Member States, such as Spain, Germany, and Austria, grant additional limited protections to the ‘creators’ of photographs that are not considered photographic works for the purposes of copyright protection. This would mean that if these works were to be used for the purposes of training a generative system, developers/trainers may find themselves to have breached the domestic rights of those from which the photograph originated. However, it should be noted that this right is limited, and does not extend in the same way it would if the holder was protected via the copyright regime. For example, the limited right will not typically protect the creator against acts of transformation or adaptation. The use of such photographs to create derivative works would not therefore generally amount to an infringement, though each instance would require assessment on a case-by-case basis.
Similar rights considerations may also include, where generative AI draws from sound or audio-visual data, the exclusive neighbouring rights of audio-visual producers, producers of phonograms, and/or performers. The extent of these rights varies from country to country. In general, however, musical performers and phonographic producers are often protected across the EU, the UK, and the US.
In some territories, such as Spain, Germany, and the State of California, creators and trainers must also consider the protections that can attach to images of real people, whether alive or deceased. As with many of the rights noted above, the exact degree and scope of the protection of these images varies. However, in each case, failure to adequately ensure that use of such images does not infringe the rights of the rightsholder may result in parties being found to be in breach of the domestic protections.
Rage against the machine: A developing opposition to the proliferation of generative AI
Unsurprisingly, anxious rightsholders are already pushing back against generative AI platforms using their works.
On 7 February 2023, the UK Government withdrew a proposal to extend copyright exceptions for text and data mining for training AI systems for commercial purposes. The existing UK exception is limited to AI training for non-commercial research purposes. In 2022, the UK Intellectual Property Office proposed to introduce a wide exception to copyright protection in the case of text and data mining (whether commercial or non-commercial), with no ability for rightsholders to opt out. The argument for such an exception was that it would promote the use and development of AI. The main reason driving the recent decision to withdraw this proposal for a wide exception was the substantial backlash from the music industry and other creative sectors who were of the view that any form of open access text and data mining should be on an opt-in basis for right holders (i.e. that this is use of their copyright works that they should be entitled to license).
However, it does appear that some legislation aimed at further facilitating the use of AI tools will follow in the UK. When the UK Government noted that the originally proposed wide text and data mining exception would not progress in its original form, it indicated that alternative proposals may emerge for consultation later. This aligns with the statement by the UK Chancellor in his budget speech on 15 March 2023 that the Government would work with the Intellectual Property Office to provide clarity on IP legislation, so that companies working on generative AI can access the material they require.
In recent weeks, Getty Images (“Getty”) has indicated its objection to the use of its copyright works by generative AI and has initiated court proceedings in the US and the UK against Stability AI, a generative AI platform which creates artwork based on text instructions input by users. Getty has claimed in a statement on its website that Stability AI has:
“unlawfully copied and processed millions of images protected by copyright and the associated metadata owned or represented by Getty Images absent a license to benefit Stability AI’s commercial interests and to the detriment of the content creators.”
The cases will raise interesting issues as to whether and when the use of copyright works to train AI systems will infringe copyright, and as to whether any defences may apply.
“Can a robot write a symphony?”
Can a robot write a symphony? With a rapidly growing armoury of AI tools, the answer is ‘highly likely’. A more problematic question is whether it (and any other AI-created works), can attract copyright protection.
The situation in the EU is relatively clear. Although there is no formal definition of “author” in EU legal texts, case law of the CJEU establishes that only a human can be considered an author of copyrightable works. This is because for a work to satisfy the requirement of originality, it is necessary that the work reflects the personality of its author, as an expression of his/her free and creative choices (CJEU Judgment in Case C-683/17 Cofemel). EU law therefore clearly opposes the idea of an AI-exclusive authorship, which is reflected throughout the national legislation of most Member States, which in some cases even establishes that the author must be a natural person (with certain exceptions for legal entities, e.g., collaborative works).
The position under UK law is more complicated. An “author” of a “computer-generated work” is defined under UK legislation as “the person by whom the arrangements necessary for the creation of the work are undertaken”. Arguably, the person making the arrangements could be seen as the owner or user of any AI platform used to make an artwork. But the situation becomes less clear when seeking to establish whether the work is sufficiently original to benefit from copyright protection when using generative software. Originality under the traditional UK test required that the author of the work created the work through their own skill, labour and judgment, and that it was not simply slavishly copied from another work. However, the CJEU case law on originality referred to above, which introduced the concept of the author’s own intellectual creation based on free and creative choices, also continues to apply in the UK, post-Brexit. The UK courts have never fully reconciled the two contrasting tests for originality, and so it is difficult to predict how courts will seek to apply them to AI-generated works. It is uncertain, for example, whether the use of training data and image prompts to develop a new image would involve sufficient originality. Furthermore, it is equally uncertain to what extent a user would need to interact with a generative AI programme in order to have made demonstrated sufficient originality (whether assessed in the traditional way by reference to skill/judgment labour, or based on the EU standard of free and creative choices, although if the latter approach is followed the likelihood of protection seems to us to be lower).
The situation in the US is also unclear and open to interpretation. In February 2023, the US Copyright Office (“USCO”) issued a decision on artwork created by the generative AI ‘Midjourney’, in which images are created by feeding the programme a series of text prompts. In their letter to the author of the book for which the images were produced, the USCO stated that the images were not eligible to be protected by a copyright registration on the basis that the artworks which were generated by AI “are not the product of human authorship”. However, in a recent policy statement from March 2023, the USCO stated that works created by generative AI may in fact be eligible for copyright protections. For a more detailed discussion on the current US position on copyright protection for AI see this article by our US colleagues.
The question of copyright protection for AI generated works is likely to remain an obscure one due to the varying degree of human involvement required for an AI tool to produce creative works, and the nuanced interpretations that could be applied to the concepts of “originality” and “creativity”.
As with the emerging legal framework surrounding the Metaverse, the law relating to generative AI is likely to be one that develops with the technology itself. It is likely that we will begin to see more claims of infringement relating to generative AI systems as their use becomes more sophisticated and widespread.