Law: Exploring the conflicts between AI and Copyright

Note: The following article was written by Ronitth Vasani L6B (20VasaniR@students.watfordboys.org)

AI, and its implications upon the realm of intellectual property law, and thereby copyright law, are  bound to be complex due to the nature of how AI models are developed; training on large datasets provided to then be able to respond adequately when prompted. However, when discussing the impacts of the use of AI Image Generators by both private citizens and commercial entities, the issue of copyright infringement becomes significantly more convoluted, as training data stems largely from works of art/other copyrighted content. Inevitably, tracing AI Image outputs to a single source material is nearly impossible given the large scope of AI training data sets (naturally excluding the recreation of famous characters, such as Darth Vader (1)), and thus identifying the victim of AI copyright infringement a tedious task. Nevertheless, despite the potential economic and technological benefits associated with generative AI (potentially increasing productivity by 1.5% (2)), the unrestricted use of copyrighted material for AI training raises substantial legal and ethical concerns that existing copyright frameworks may be ill-equipped to address.

When approaching the question from a legal perspective, the issue of what constitutes copyright infringement is relatively clear, as presented in the Copyright, Designs and Patents Act 1988. However, when considering the use of copyrighted works within AI training data, the application of legislation becomes slightly more unclear. To argue in favour of AI training being copyright infringement, one must consider firstly what a “copy” is defined as legally, which in the CPDA, is also constituted as “storing the work in any medium by electronic means” (CPDA 1988 - s17c2). Therefore, when considering the process of downloading, indexing and caching data in anticipation of its use in training an AI model, this may certainly be considered a form of copyright infringement due to the creation of additional duplicates of the data. Furthermore, under the CDPA s9, when considering the copyright implications of computer-generated material, statute requires “the author [to] be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken”, strengthening the case for the use of copyright material being infringement.

When applied to the case of AI Image Generators, training data act functionally as the “arrangements necessary for creation of the work”, particularly where generated outputs are heavily derivative of stylistic features extracted from copyrighted works, and thus arguably shifts the authorship of a given AI’s output back to the authors of the AI training data (when used without a licence/otherwise given permission), instead of either an entity prompting an AI models output/the company running the AI model. The strongest copyright claim against generative AI may not concern the outputs themselves, but rather the initial reproduction and storage of copyrighted works during training. 

Additionally, in consideration of the exclusive rights granted to the copyright holder, who hold the right solely to issue copies/adaptations of the work to the public (CDPA s16). Arguably, despite the partially transformative nature of an AI model, given both the free access to the public of these AI image generators (e.g. Midjourney, whose services were available completely freely/anonymously through the social media platform, Discord etc.), as well as the incentive for companies to create more apt models in order to begin to charge consumers for use (Midjourney garnering over $300mn in revenue in 2021), already seen through services provided by companies such as Google, OpenAI and Anthropic. In the context of the AI’s output being a “substantial part” (typically constituting a part containing elements that express the author’s intellectual creation, skill, labour, or originality)  of an author’s work, although this may be hard to assess, services like those provided by Vermillio provide an opportunity to quantitatively analyse how much of an AI’s output derives from copyrighted material. Although artistic “style” is not protected by UK Statute, given that in the case of many copyright infringement cases, not just the style but significant parts of the work such as characters/scenes are reproduced, meaning when quantitatively assessed, a high percentage of the work is determined to have derived from copyrighted material.

In these scenarios, copyright infringement is explicit, thereby furthering the validity of the argument that an AI’s output is not always sufficiently transformative. 

Conversely, according to the CDPA s29, permissible use of copyrighted material includes use for  a “person who has lawful access to the work [to carry out a computational analysis”, which certainly would include the training of AI models, as ingesting of large volumes of data is impossible to avoid, as well as the exception to the provision for crediting the authors use being met, once again due to the volume of the work being processed by AI models.  Further, proponents of the expansion of AI use argue that the AI training process focuses primarily upon the AI analysing the patterns within the work, rather than the expressions in the work itself. Nevertheless, despite there being some arguments in favour of AI training not violating IP laws, the argument is contingent upon the “analysis” being for “non-commercial” purposes, which in the majority of cases is simply untrue, as even non-profit AI companies such as OpenAI eventually plan to capitalise on the significant volume of AI research through monetisation of their services. 

However, while the legal implications of AI image generators are extensive, the ethical considerations are certainly more intricate. Firstly, the implications of allowing AI models to scrape data from the internet freely clearly violate an ethical principle of consent from the creative sector, and more specifically the artists that create the work. Artists must consent explicitly to their work being used, as argued by those in the creative industry, and as such one must argue that there needs to be better protections for creatives to prevent use of their work in AI training models. Furthermore, given the controversial nature of AI in regards to many of those in the creative industry seeing AI as a threat to their livelihood (4) , use of their work in training models presents a further ethical challenge, as continued production and distribution of work online by artists furthers the likelihood that AI may be able to flawlessly replicate artists work, thereby diminishing demand for illustrators, animators and graphic designers.

Further, given use of the use of creative’s work in an AI training data set, there ideally should be some form of a contract giving explicit permission for use, most importantly requiring some form of consideration from both parties. Currently, artists have very little control over what AI companies use their data, how it is used and receive no compensation/recognition, thus not meeting the legal requirement for consideration. While obviously, it is unrealistic to expect agreements with the multitude of artists whose work is scraped by these AI tools, some form of recognition/notice is the minimum that artists should expect when their work is used, although implementation of a said system is once again hard to implement.

The economic interests of allowing of more widespread data scraping of AI also present another consideration, as according to estimates presented by the IMF, AI has the potential to boost productivity by nearly 1.5%, potentially leading to economic benefits for the wider public (i.e. the collective), adding a dimension of protection of individual rights in situations which may benefit society as a whole.

Although a proposed solution to this issue is presented in the Government’s discussion of balancing the interests of AI companies and artists through implementation of an opt-out system, by including machine-readable text into published work explicitly reserving rights against use of said work in the training of AI models, this solution is inherently flawed by the fact that although this addition is possible for artists/entities with the means and knowledge to implement this into their distributed work, such as companies like Disney, many smaller/independent artists may either be unaware of these changes, how to implement them, and may not have the means to issue copyright claims in court due to financial restrictions, creating effectively further exploitation of artists widespread AI scraping had already greatly impacted, arguably requiring updated legislation to expand rights of these groups, meaning that although they have equal protections in law, in reality they lose exclusive rights granted to them in the CDPA 1988.

Overall, uncertainty arises at both stages of AI development: the training phase, where issues of copying, storage, and the text and data mining exception under the CDPA 1988 s29 are highly contested, and the output phase, where questions of substantial similarity, authorship and originality remain difficult to resolve. Recent consideration of these issues in cases such as Getty Images v Stability AI (2025) has led to judgements in favour of generative AI companies by confirming use of data for training purposes does not constitute as secondary copyright infringement, ambiguity over how this may apply to primary copyright infringement given the jurisdictional issues forcing Getty Images to drop claims of primary infringement, as well as the issues of the act of “scraping” itself. The result is a legal framework that provides partial answers but lacks clarity in addressing the scale, automation, and statistical nature of generative AI systems, and the challenges faced by both artists and AI companies warrant either the revision or expansion of relevant statute law to ensure adequate protections for economic interests and artistic rights.