[This piece is co-authored by Siddhant Singh and Gurmehar Bedi, who are third-year students pursuing a B.A.LLB at National Law University, Jodhpur. In this piece the authors analyse the decision of Getty Images vs. Stability AI to deal with aspects of copyright in a technological context vis-à-vis safeguarding creative labour, and determining what the ruling’s global reverberations may be.]
Introduction
When Justice Joanna Smith delivered her ruling in Getty Images v. Stability AI, the UK High Court became the focus of all eyes in a contentious debate that lies at the intersection of copyright governance and artificial intelligence (“AI”). The lawsuit was a result of Getty’s allegation that Stability AI relied on countless photographs copyrighted by the former to train its image-generation model. This raised a critical concern for creative industries: whether copyright law can protect creative works when such works are being absorbed into AI training processes. While the court had the unique opportunity to protect creative works at the training stage as well, it fought shy of doing so and instead, opted for textual precision by treating learning and copying differently.
The practical ramifications of treating AI training as fundamentally different from reproduction are the weakening of copyright deterrents and concerns with respect to creativity, especially in the future. In our view, this judgement puts forth an unnaturally restrictive view of copying, one that potentially leaves authors insufficiently protected. We wish to explore the nuances and flaws of this judgement from a purposive perspective, one that aligns with the underlying objective of copyright i.e. to safeguard creative labour, and determine what the ruling’s global reverberations may be.
The Case in Brief
Getty claimed that the text-to-image model, Stable Diffusion, was developed based on Getty’s images without permission, thereby violating their copyright. The claim comprised primary infringement, alleging the usage of copyrighted material in the training process; and secondary infringement, alleging liability for commercialising works derived on the basis of copyrighted works. Additionally, Getty alleged trademark and database rights claims.
By the time of the trial, the primary claims had been dropped. Getty was unable to provide evidence of any training having been done in the UK itself, so the Court was forced to determine whether the model itself, i.e. the set of learned parameters, could be regarded as an “infringing copy.” Getty contended that the model in question was simply a derivative embodiment or result of its photographs. To this, Stability AI replied that the model was merely a set of statistical patterns and did not contain any data from the training set.
The Court’s Findings
Justice Smith arrived at two critical conclusions. Firstly, the Court held that the alleged acts of reproduction occurred abroad, and the resulting model within the UK contained no copies, thereby dismissing Getty’s copyright claim.
Secondly, she held that the model weights were not copies. She felt that Stable Diffusion “does not itself store the data on which it was trained.” It only figures out the relationships between pixels and, based on this learning, produces new images. Since the model had never stored any of Getty’s photographs, it could not be an infringing copy. We deal with each of these findings in turn.
Why Territoriality Fails in the Age of Global AI
Regarding territoriality, Getty argued that although the actual servers were located overseas, the infringement took place in the UK. Stability AI contended that since all of the training (using the Getty Images dataset) took place outside of the UK, the court lacked jurisdiction.
Favouring Stability’s arguments, the court determined the question of territoriality by looking at the physical location of the AI training. Although, territoriality is a settled principal for he exercise jurisdiction, it operates on the assumption that commercial exploitation can be fixed to a singly physical location. This assumptions breaks down in light of the distributed nature of AI development, which relies on distributed data pipelines ands globally spanning cloud infrastructure. Therefore, this formalistic application of the law is concerning in the context of AI.
Furthermore, using server location to identify infringement ignores the location of technological control and commercial exploitation. This enables businesses to structure infrastructure outside of while continuing to commercially exploit the model within in the UK. A type of copyright arbitrage is thereby produced. Instead of concentrating on the location of computation, a more purposive approach would emphasise the locus of control and profit would allow the legal regime to track control, and market impact notwithstanding the incidental location of servers.
Getty transforms Secrecy into Immunity
The structural weakness is exacerbated by the evidentiary issue that traditional rules of proof are ill-equipped to address.
Where the plaintiff traditionally bears the burden to prove infringement, in the context of AI, that burden becomes unbearable.This is because AI systems usually operate as “black boxes.” These are systems where only developers have information about their internal processes of AI training. This creates barriers that obstruct transparency, accountability and verification. All of the data in Getty was encoded in private systems and protected by claims of confidentiality and trade secrets. Hence, Stability was the only one who knew where its data was stored, how it was processed, and which datasets were used to train its models. This makes the evidentiary threshold of infringement unsurmountable in the context of AI since plaintiffs have no method to access the evidence required to prove infringement.
This unnecessary smokescreen transforms secrecy into immunity where the more confidential the process of AI training, the harder it is to prove infringement arising from it . Scholars like Pasquale concur, referring to this system as a form of “information asymmetry” eroding systems of accountability.
This issue also occurred in Anderson and Silverman where courts demanded direct proof, which was impossible to discover as a consequence of the aforementioned practice. Moreover, experts from the industry contend that black boxes in the AI industry are needless.
Why Model Weights Should Be Considered Copies
Another important question that emerges from Getty Images v. Stability AI is whether a model which has been trained with numerous mathematical “weights” can create issues of copyright infringement. Getty made the argument that by using their images, Stable Diffusion captured the essence of these images, and thus, resulted in a model that embodied the copyrighted material. Conversely, the defendant contended that model weights were solely mathematical abstractions and consequently, did not reproduce or express the material. Justice Smith, unfortunately, sided with the latter, holding that since the model “does not itself store or contain” such images, it cannot lead to the violation of copyright. Although the rationale purports to distinguish between learning and copying, it arguably misapprehends how machine-learning models internalise information.
Strictly speaking, model weights are not abstract mathematical values. Rather, they are numerical representations of visual characteristics that were abstracted from the training data. When countless images are processed, their patterns, textures, colours, and compositions all get recorded in weight matrices, which express these characteristics and relationships in a statistical or numerical manner. This process is conceptually similar to compression, wherein, while the model does not store every single image in its entirety, it does have a representation of the individual images in the weight matrices. Similarly, in this case, the weights formulate a functional copy of the original material’s creative DNA in a form that is not readable by humans.
The holding that model weights fall outside the scope of reproduction simply because their mathematical nature misrepresents what digital reproduction really means: the conversion of creative expression into mathematical code. The law should not allow abstraction to take precedence over substance because such an approach leaves creators unprotected. Their works are made commercially exploitable under the facade of abstraction, while granting developers a wide berth to monetise creative expression without any prior licensing.
The Global Reverberations of Getty Images v. Stability AI
A conundrum that courts around the world are currently facing is brought to light in the Getty ruling: can copyright law, which is rooted in the idea of tangible reproduction, regulate the intangibility of machine learning? The Court has, in a sense, stated that learning from protected material may not be deemed infringement unless there is a visible replica. The ruling serves as an example of the challenges that copyright law is facing due to the design of generative AI.
Such a message will have a significant impact beyond the UK as well. Through the fair use principle, courts across the United States are attempting parallel investigations. Federal judges have taken a practical stance in Kadrey v. Meta and Bartz v. Anthropic, concentrating on whether AI training is truly “transformative”. While this is more flexible than the UK’s reproduction-centric approach, it also significantly benefits developers. The result, then, is a lenient legal environment on either side of the Atlantic under which creators are left to grapple with evolving definitions of use and harm, while innovation is given the benefit of the doubt.
The EU has taken up a more proactive approach to clarify this uncertainty. The Digital Single Market Directive’s Copyright and the EU AI Act both mandate disclosure while limiting the exceptions for text and data mining to non-commercial purposes. However, the UK’s autonomy in the post-Brexit era has led to a more liberal and permissive environment that draws developers but runs the risk of isolating rightsholders.
Broader implications of the judgment raise alarm bells. If a model weight is not deemed to be a reproduction or a form of copying, and training AI abroad evades liability, creatives become very vulnerable. Authors, artists and photographers have no control over their own work, being completely unaware of whether or not it is being absorbed and exploited in an ocean of code. The signal of this ruling is clear: courts will continue to apply the obsolete copyright regime textually until the legislature takes charge and modernises the statutory protections.
The Way Forward
In the aftermath of Getty, courts and legislators must recognise the foundational goals of copyright: rewarding original creative effort, preventing unfair market practices, and enabling innovation that is of social value. The relevant question is not regarding whether “model weights contain expression” but rather whether training on copyrighted images enables AI to be commercially substitutive of underlying works.
In the UK, courts and the parliament ought to adopt a more pragmatic approach towards the issues of territoriality and evidence. This would be in line with the holding that intellectual property rights must respond to changes in technology and exploitation methods.
The global IP regime needs renewed and contemporary transparency norms. International soft-law standards via WIPO or OECD should aim to guide domestic regimes towards compatible evidentiary obligations which balancing auditability without exposing proprietary architecture. A black-box constraint complaint system which balances creative effort and innovation is the most sustainable and realistic path forward.