What’s Really at Stake: How Getty v Stability AI will shape provenance for AI-generated art
No Robots Allowed. “We call it study when they do it. When you do it, we call the principal.” Same assignment, different rules.
Consent, Signals, and What Comes Next
Every era of art has a threshold moment when something new enters the room and asks whether the old rules still apply. The printing press did it to scribes, photography did it to painters, and now generative models are doing it to everyone at once. They learn by consuming what already exists, and the open question is whether that consumption is creative, exploitative, or something we have not fully named yet. That question is no longer hypothetical. It is now in court.
The case drawing the most attention is Getty Images v Stability AI. Getty is one of the largest stock photography libraries in the world, an archive that quietly powers magazines, websites, campaigns, and visual economies. Stability AI develops Stable Diffusion, a widely used text-to-image model that synthesizes new images from learned statistical relationships rather than storing originals. Getty alleges that Stability trained on more than twelve million Getty photographs, along with captions and metadata, without permission. (BakerHostetler summarizes the allegation at this scale.)
“Provenance is part of how art earns trust.”
One detail makes the dispute feel concrete for non-lawyers. If you query the model in certain ways, you can sometimes see a warped, ghost-like version of the Getty watermark appear in the output. Getty treats this as evidence that the system digested its branded material and that artifacts now risk confusing viewers about origin. Associated Press reporting from the London courtroom captured that focus during the trial. (AP News) Stability’s response is that training learns patterns, not copies, and that stray watermark artifacts are rare, unintended technical noise rather than trademark use. Reuters’ coverage of opening arguments underlined how both sides framed these stakes for the court and for the broader industry. (Reuters)
Why this matters to Mindset Art Collective is simple. Provenance is part of how art earns trust. When we curate digital collections, arrange stories on screens, and present work to people at home or in public spaces, audiences rely on signals about origin and license to understand what they are seeing. That is also why this series sits alongside our ongoing conversations about what AI adds to human creativity and how we talk about consent and authorship.
What the Courts Are Weighing
Two questions sit at the heart of the litigation. First, is training itself unlawful when the training set contains copyrighted or trademarked material. Second, even if training is not directly unlawful, can specific harms flow from it, such as watermark artifacts or look-alike images that erode licensing markets and confuse viewers.
Key Takeaways
Identify the core questions: Is training on copyrighted or trademarked material unlawful, and do AI-generated outputs cause real harm by leaking watermarks or producing look-alike images that confuse viewers and erode licensing markets?
Explain the UK fight: The focus is secondary infringement (liability for dealing with infringing copies), whether model weights qualify as an “article” (an item that can be imported or distributed), and trademark and passing off (misrepresenting source) based on watermark artifacts.
Outline the U.S. battle: The analysis turns on fair use, substantial similarity, and de minimis overlap (so minimal the law treats it as negligible), which will determine how courts treat model outputs.
In the United Kingdom, the High Court held a multi-week trial in June 2025. As closings began, Getty narrowed its claims after the record did not support that model training occurred in the UK. The case refocused on secondary infringement and on marks. In UK law, secondary infringement covers dealing with infringing copies, including importing or distributing an “article” that embodies the infringement. The open question is whether a pre-trained model’s weights can be treated as such an article when imported or used domestically. Practitioner analysis has flagged sections 22 and 23 of the Copyright, Designs and Patents Act 1988 as central to that inquiry. Alongside that, the court is evaluating trademark and passing off. Passing off in the UK addresses misrepresentation that leads consumers to believe a work comes from a particular source, causing damage to the source’s goodwill; here the focus is whether occasional watermark artifacts function as a badge of origin in viewers’ minds. (Osborne Clarke) The court had already streamlined the case in January 2025 by limiting an attempt to run broad representative claims, a posture spelled out in the approved judgment. (Courts and Tribunals Judiciary) And Associated Press reporting confirms the late-stage narrowing as the trial drew to a close. (AP News)
“If a generated image occasionally carries a distorted watermark, the risk is not only legal exposure. It is audience confusion and erosion of trust”
In the United States, Getty first filed in Delaware in 2023. That action bogged down in disputes about jurisdiction and transfer. On August 14, 2025, Getty voluntarily dismissed in Delaware and refiled in the Northern District of California the same day, a venue that already hosts several generative-AI suits. Bloomberg Law’s report on the refiling repeats the “more than 12 million photos” allegation and frames the new complaint’s scope. (Bloomberg Law) Early docket entries in the California case show routine assignment steps and scheduling toward an initial case management conference this fall. (CourtListener) The U.S. merits fight will likely hinge on fair use, substantial similarity, and de minimis overlap. Substantial similarity asks whether protected expression from an original appears recognizably in the accused material. De minimis describes an overlap so trivial that the law does not recognize it as infringement. How courts measure those thresholds in the context of training and generation will influence how model developers design filters and how platforms like ours evaluate outputs.
“We support paying artists and legitimate licenses, but we do not support royalties merely because a model was trained on other images. Study and influence are part of art.”
Watermark artifacts deserve special attention from curators. A watermark is a signal about licensing and origin. If a generated image occasionally carries a distorted watermark, the risk is not only legal exposure. It is audience confusion and erosion of trust. The factual questions for courts mirror the practical questions for curators. How often does it happen. How visible is it. Does the artifact function like a brand in the viewer’s mind or merely like a glitch. Those questions are at the center of the UK trademark and passing off analysis, and they will echo in any U.S. claims that turn on source confusion. (Osborne Clarke)
We will keep this discussion grounded in our larger editorial approach to AI, human authorship, and collaboration. When you want a companion lens on how we think about the creative upside, see our Journal reflections on why AI can expand, not replace, human intent, and how we hold provenance and consent together in practice.
Status and How We Curate
Timeline
2023
Jan 16: Getty files UK High Court claim against Stability AI
Feb 3: Getty files U.S. case in the District of Delaware
2025
Jan 14: UK court trims representative claims
Jun 9: UK trial opens in London
Jun 25: Getty narrows claims to secondary infringement and marks
Jun 30: UK trial concludes; judgment reserved
Aug 14: Getty dismisses Delaware case and refiles in Northern District of California
What’s next
UK: Written judgment expected late 2025; if delayed, early 2026
US: ND California case management conference Nov 13, 2025; scheduling order to follow
Watchpoints: secondary infringement as “article”; trademark and passing off; fair use; substantial similarity; de minimis overlap
Series: Part Two publishes after UK judgment or the first substantive U.S. ruling
Key takeaways for October 2025: the UK trial is complete and judgment is pending. The live UK issues are secondary infringement and marks. Getty has restarted the U.S. action in the Northern District of California. The scale allegation remains large, with Getty citing more than 12 million photographs plus related text metadata. (Bloomberg Law, Reuters, AP News)
For readers following the docket, here is the short timeline. In the UK, a January 14, 2025 judgment trimmed the case’s representative posture, trial opened in June, and multiple outlets reported that Getty narrowed its claims at closings to focus on secondary infringement and marks. A written decision is expected this term. (Courts and Tribunals Judiciary, Reuters) In the U.S., Getty’s August 14, 2025 refiling in the Northern District of California is now the live case, with standard scheduling underway. (Bloomberg Law) None of that tells us how the merits will land. It does tell us what to watch next: whether UK law treats a pre-trained model as an article for importation and distribution claims, how the court measures confusion around watermark artifacts, and how California judges frame the relationship between training copies and expressive use.
Here is how we approach this at Mindset Art Collective. If you create an image, whether you paint, photograph, run Stable Diffusion locally, or use a hosted service like Midjourney, we recognize you as the rights holder, provided you obey the tool’s terms. We support paying artists and legitimate licenses, but we do not support royalties merely because a model was trained on other images. Study and influence are part of art.
To prevent confusion, we remove any image that contains a residual watermark, logo, or signature suggesting a license we do not hold. Our review is conscientious, though not exhaustive. We may miss a rare artifact, but we act quickly on any valid takedown request. Contributors license their work to us and confirm they have the necessary rights. We keep comprehensive provenance records internally so we can protect artists and answer hard questions when they arise. Publicly, we show the essential trust signals: the usage license and the creator’s chosen display name. Any deeper disclosure is left to the creator’s discretion.
Part Two will follow when the UK judgment lands or the California case clears its first substantive hurdle. We will translate the rulings into concrete steps for artists, curators, and collectors, and we will keep choosing consent and clarity while the law finds its footing.
Sources
Reuters, UK trial opens, June 9, 2025. Reuters
Associated Press, narrowing at closings in London, June 25, 2025. AP News
Osborne Clarke, analysis of remaining UK issues and secondary-infringement theory. Osborne Clarke
UK Judiciary, approved judgment on representative-action posture, Jan 14, 2025. Courts and Tribunals Judiciary
BakerHostetler, case tracker summary of allegations including “more than 12 million” photos and related text. BakerHostetler
Bloomberg Law, Delaware dismissal and refiling in ND Cal and scale allegation. Bloomberg Law
CourtListener, ND Cal docket and early case-management activity. CourtListener
