Telescope Logo
Telescope Logo

Illegal content in the AI supply chain

Read time: 6 minutes


The commercial logic of generative AI is simple: collect as much data as possible, as cheaply as possible, and let scale do the rest. There is now an open question: is that logic colliding with the law?

When Getty Images sued Stability AI, it alleged that Stable Diffusion had been trained on LAION-5B, a dataset of 5.85 billion image-text pairs built by scraping links from across the web, including Getty’s own sites. In the United States, authors alleged that Anthropic had downloaded millions of books from the pirate libraries LibGen and PiLiMi.

The commercial logic of generative AI is simple: collect as much data as possible, as cheaply as possible, and let scale do the rest. There is now an open question: is that logic colliding with the law?

When Getty Images sued Stability AI, it alleged that Stable Diffusion had been trained on LAION-5B, a dataset of 5.85 billion image-text pairs built by scraping links from across the web, including Getty’s own sites. In the United States, authors alleged that Anthropic had downloaded millions of books from the pirate libraries LibGen and PiLiMi.

Photography on a beach

Getty Images dropped its claims regarding primary copyright infringement and lost its claims regarding secondary copyright infringement (Getty Images is appealing the latter). The claims against Anthropic were settled without Anthropic admitting liability. But are these claims indicative of a structural problem? Is it the case that pirated books, unlicensed photographs and personal data scraped in breach of terms are moving through training pipelines and into commercial models? If that is right then this raises serious questions for off-takers or customers of generative AI, for if the rule-breaking happens upstream, the end user may never see it or be able properly to quantify their own risk. In 2023, the Ada Lovelace Institute compared this risk in AI supply chains to illegally mined gold entering technology manufacturing despite, as with AI, the existence of supplier codes of conduct and audit processes. However, the key technical point that matters for commercial enterprises is this: model weights (numerical parameters that specify how the model represents connections to leverage between pieces of data to achieve the desired output) are a byproduct of the training data. Unlawful inputs therefore cannot simply be recalled once a model is on the market producing output, because the model (and therefore the output) is inherently formed by its (potentially unlawful) training data. The question for businesses is whose balance sheet absorbs the consequences.

Global legal landscape

The threshold question is whether a model can itself be an unlawful artifact.

In the United States, the Copyright Office’s May 2025 report considered that every stage of the pipeline, from data collection and curation through training to output generation, can be a prima facie infringement of the right to reproduce works. The USCO considered that, where AI generated outputs are substantially similar to the training data inputs, there is a “strong argument” that the models’ weights themselves infringe the reproduction and derivative work rights of the original works. However, this point has been sharply debated in the US, with different courts taking opposing views (e.g., with the US District Court for the Northern District of California (Kadrey v. Meta Platforms, Inc) describing allegations that a model can infringe copyrighted works as “nonsensical”).

There are potentially serious consequences for getting it wrong. In Bartz v Anthropic (a class action brought in the US, that settled with final approval in July 2026), the court held in June 2025, that the use of lawfully acquired copyrighted books to train Claude was fair use, because the training of AI models is a “transformative” process. However, the court considered that Anthropic had sourced some of its training material unlawfully i.e., it had sourced pirated copies of books and used those to train Claude. The sourcing of the pirated copies was not fair use. The message seems to be clear: how the data was acquired (and whether it was acquired lawfully), not how the data is used, is key in considering liability.

AI Studio screen

In a recent case in England and Wales, Getty Images v Stability AI, Getty initially brought a claim for:

  1. primary infringement of its copyrighted material (i.e., that Stability had infringed Getty’s copyright through unlicensed training of its AI model Stable Diffusion using Getty’s image library); and
  2. secondary copyright infringement of its copyrighted material (i.e., Stable Diffusion contained copies of Getty’s copyrighted images, and making Stable Diffusion available for download in the UK was importation of infringing copies of Getty’s images).

Getty dropped its primary claim because it could not prove that the training of Stable Diffusion took place in the UK. It therefore remains to be seen whether, as a matter of English law, training AI models in England on images protected by copyright can constitute infringement (as above, the US court has considered it does not).

In respect of Getty’s secondary claim, the Court held that AI model weights are not a “copy” of the images in the sense required by the Copyright, Designs and Patents Act 1988 (“CDPA”). The CDPA requires a recognizable reproduction of a work or a derivative in which it is embodied. An AI model, by contrast, contains statistically trained parameters, not stored copies or reconstructions of photographs. Additionally, the model weights of the various iterations of Stable Diffusion did not store the visual information in the copyright works and therefore could not infringe Getty’s copyright. Getty has appealed this part of the decision with arguments to be heard at about the time this article will be published.

In the EU, all providers of general-purpose AI models must, under Article 53 of the AI Act, adopt a copyright policy and publish a training-content summary using the Commission’s mandatory template. Crucially, the template requires:

  1. disclosure of information about data sources (e.g., whether data is sourced from public or private datasets); and
  2. providers to detail the measures taken to remove “illegal” content from data. One example of “illegal” data given within the template is “the non-authorized use of material protected by intellectual property rights.”

Failure to complete the template can lead to fines of up to 3% of worldwide turnover or EUR 15 million (whichever is higher, unless the infringer is an SME).

Operational impact

The above instances all focus on the AI models themselves. How does this affect other corporate entities in the generative AI supply chain?

The current academic thinking is that, as generative AI becomes more embedded, claims concerning the lawfulness of the data used to train the model will resemble claims in product liability. That is, liability will be fragmented or shared across different actors in the supply chain.

The “taint” problem of unlawful datasets is real. For any parties downstream (e.g., a supplier that is deploying AI to an end user), the risk is inheriting any “‘negligence”‘ or “‘unlawfulness”‘ from the upstream supply chain. To assess, mitigate and avoid these risks, it is essential for deployers and off-takers to have information on how the upstream model was trained, what data it used and how it was tested.

Opportunities and practice points

Provenance is becoming a product feature. Developers of AI models that keep auditable records of where data came from, what filtering was applied and when, together with a record of where training, storage and hosting take place will win procurement processes and regulatory inquiries alike. Indeed, in Getty Images, it was the fact that Stability was able to show how it trained its model that allowed it to defeat Getty’s primary infringement case. Providers and deployers alike must allocate liability within their contracts for non-compliance and third-party claims. It is very likely that we will see an increase in deployers looking for warranties on data provenance or indemnities for training data claims and audit rights. The other side of the increased risk or burden this can bring is that curators and trainers that can offer such warranties credibly will command a premium. Every organization should also have a tainted-data playbook covering takedown, model rollback or retraining and notice to customers and regulators.

Key takeaways

  1. Acquisition is the new liability trigger. Courts are separating lawful training from unlawful sourcing, so audit how your data and your vendors’ data were obtained, not just how they are used.
  2. Deployers cannot outsource the risk. Visibility and regulatory attention sit downstream, so build provenance warranties, indemnities and audit rights into every model procurement now.
  3. Treat tainted data like a product defect. Screen upstream, document throughout and test a remediation plan before a regulator or court sets the standard for you.

Key takeaways

  1. Acquisition is the new liability trigger. Courts are separating lawful training from unlawful sourcing, so audit how your data and your vendors’ data were obtained, not just how they are used.
  2. Deployers cannot outsource the risk. Visibility and regulatory attention sit downstream, so build provenance warranties, indemnities and audit rights into every model procurement now.
  3. Treat tainted data like a product defect. Screen upstream, document throughout and test a remediation plan before a regulator or court sets the standard for you.

Authors


Phillip Richardson

Phillip Richardson

Partner, UK

View profile
Emily Quinn

Emily Quinn

Senior Associate, UK

View profile

In this edition


In this edition


Children on phones

Age assurance as digital infrastructure

Read time: 7 minutes

Read more

The content licensing reset

Read time: 5 minutes

Read more

Platform governance across multiple rulebooks

Read time: 5 minutes

Read more

AI agents and consumer liability: new interfaces, new risks

Read time: 8 minutes

Read more

Creators, advertising, and monetization

Read time: 6 minutes

Read more

Protecting premium content at internet speed

Read time: 6 minutes

Read more

Disputes and enforcement in the digital content sector

Read time: 4 minutes

Read more

Previous page

Next page

Arrow

© Eversheds Sutherland. All rights reserved. Eversheds Sutherland is a global provider of legal and other services operating through various separate and distinct legal entities. Eversheds Sutherland is the name and brand under which the members of Eversheds Sutherland Limited (Eversheds Sutherland (International) LLP and Eversheds Sutherland (US) LLP) and their respective controlled, managed and affiliated firms and the members of Eversheds Sutherland (Europe) Limited (each an "Eversheds Sutherland Entity" and together the "Eversheds Sutherland Entities") provide legal or other services to clients around the world. Eversheds Sutherland Entities are constituted and regulated in accordance with relevant local regulatory and legal requirements and operate in accordance with their locally registered names. The use of the name Eversheds Sutherland, is for description purposes only and does not imply that the Eversheds Sutherland Entities are in a partnership or are part of a global LLP. The responsibility for the provision of services to the client is defined in the terms of engagement between the instructed firm and the client.

Connect with us

linkedin logo
facebook icon
youtube icon