Getty v. Stability AI After the UK Ruling: Copyright, Watermarks, and the Duty to Keep the Record Clean

A solitary figure faces a monolithic form emerging from a quiet, textured seascape.

When we published our first piece on Getty v. Stability AI, the central question was unresolved. No one yet knew how far copyright law would reach into the training and output of generative models.

The High Court in England and Wales has now ruled. It is not the sweeping answer many people expected. The court rejected Getty’s surviving secondary copyright claim. It also found limited trademark infringement tied to certain watermark-bearing outputs, especially from earlier model versions. The decision closed off one of Getty’s main copyright theories, but it did not resolve the larger fight over AI training. On the copyright side, Getty did not get the broad ruling it wanted.

The case had narrowed well before judgment. Getty dropped its primary copyright and database-right arguments after conceding there was no evidence that relevant training occurred in the UK. What remained was a secondary infringement theory and trademark claims tied to outputs. In that context, the court held that the final Stable Diffusion model does not store or contain reproductions of Getty’s works and therefore is not an infringing copy under UK law. The holding turns on that specific theory and that specific record.

“Generative systems break trust when they push ownership signals back into circulation without context or permission.”

Most coverage treats this as a broad ruling on training. It is not. The court ruled on a narrowed case, not on AI training as a whole.

What the decision does settle is one recurring claim about trained models. Getty’s remaining theory did not persuade the court that a trained model should be treated as a hidden archive of the works it learned from. Under the theory Getty left before the court, model weights were not treated as stored copies of the underlying images. That should end the habit of talking about trained models as though they were just image vaults in compressed form.

That shift narrows the field. If model weights are not treated as stored copies, future arguments will have to focus more heavily on what happens during training, where the relevant acts occur, what the law treats as reproduction at that stage, and what resurfaces at generation time. Those are harder questions. They are more technical, more territorial, and less friendly to slogans.

What the Court Clarified

The trademark findings move in a different direction. The court found limited infringement tied to watermark-bearing outputs, particularly in earlier versions of the model. Those outputs sometimes carried distorted Getty or iStock marks. Consumers could interpret those marks as signals of origin or affiliation. That is enough to create real legal exposure, because origin signals are not visual static.

For a court, this is a question of trademark and consumer confusion. For anyone working with images, it is a question of provenance. A watermark carries authorship, licensing, and ownership with it. When that signal appears in a generated image without authorization, the chain of custody is broken.

We do not think this is complicated.

We protect artists’ intellectual property, including artists who use AI-assisted workflows. The tool does not change the obligation. If a creator has the right to produce and distribute a work, that right holds. If a system produces an image that carries another party’s watermark, signature, or logo, that work should not pass through curation unchanged. The obligation runs both ways. Authorship has to be defended, and permission has to be visible.

Where the Real Fault Line Is

Here the case becomes more revealing. It does not settle the legality of training. It does make clear that outputs can create concrete, recognizable harm when they carry residual ownership signals. Getty has already been granted permission to appeal the dismissal of the secondary copyright claim, so even the copyright side of this case remains in motion.

Courts will continue to work through training-stage questions, including where copying occurs, which jurisdictions apply, what counts as fair use or infringement, and how licensing markets should function. That process will take years. The broader dispute is still active.

We do not need to wait for final doctrine to know how to behave.

We do not treat learning from existing culture as a royalty trigger. Human artists study, reference, and transform existing work. That has always been part of the creative process. The harder problem is provenance failure. It shows up in large-scale ingestion without transparency and in outputs that reintroduce ownership signals without authorization. Generative systems break trust when they push ownership signals back into circulation without context or permission.

“That should end the habit of talking about trained models as though they were just image vaults in compressed form.”

Provenance failure is harder than a simple copyright argument because generative systems are built to optimize outputs, not to preserve lineage. They are good at abstraction, compression, and recombination. They are poor at traceability unless someone goes out of their way to design for it. A system can emit a watermark fragment, a signature shape, or a brand marker without carrying any built-in account of where that signal came from, why it appeared, or what obligations attach to it. That is one reason the provenance problem is so hard to contain.

That is also why the provenance question cuts deeper than the usual abstraction about training data. Artists, platforms, and collectors do not work in the abstract. They work with objects, records, sales, captions, editions, takedowns, and public trust. Once an ownership signal is pushed back into circulation without context or permission, the burden shifts downstream. Someone has to detect it. Someone has to remove it. Someone has to answer for it. That is where a system stops being defensible.

For artists working with AI tools, this translates into discipline. Understand the systems you use. Review outputs for residual marks, signatures, or fragments that do not belong to you. Do not treat those artifacts as noise.

For platforms and curators, the responsibility is heavier. Detection, review, and removal of contaminated outputs should be ordinary platform hygiene. Provenance should be treated as part of the work itself, not as metadata that can be ignored when it becomes inconvenient.

That standard applies whether or not a lawsuit forces it.

The decision does not resolve AI and copyright. It cuts away one weak theory and puts a more immediate problem in plain view. Model weights are not easily reduced to stored copies. Watermark-bearing outputs are not benign. The question now is whether the systems being built around generative art can produce work without corrupting the record they depend on. That question reaches beyond Getty, beyond Stability, and beyond this one round of litigation.