Key takeaways from the Getty v. Stability AI UK ruling

The recent UK High Court ruling in the case of Getty Images vs. Stability AI, the company behind the Stable Diffusion model, has provided some much-anticipated clarity on how copyright law applies to training AI models.

The judgment, a 205-page read, offers important insights for anyone involved in data gathering and AI development.

Getty Images had alleged that Stability AI unlawfully used its extensive library of stock images to train its AI model.

However, a key aspect of Getty's claim, centered on the act of copying during the training and model development stage, was actually abandoned early on.

Getty had brought the case in the UK, but Stability successfully argued, during the case, that the training did not happen in the UK - the images never touched a server in the UK, as they were on AWS services in the US.

This jurisdictional issue meant the court never ruled on whether the initial copying of images for training constituted primary copyright infringement (copying the images to train the model) under UK law.

What the court did examine were two other points: secondary copyright infringement and trademark infringement.

Imported model did not hold ‘copies’

The most compelling part of the ruling dealt with secondary copyright infringement, which includes “importing” the already-trained model into the UK jurisdiction.

The judge stated that an intangible electronic copy of an image could constitute secondary copyright infringement by such “importation”.

But the judge further stated that this would only be an “infringing copy” if the AI model itself had at some point contained a copy of the copyrighted image used to train it.

In this case, the judge found that Stability AI is not itself a library of images or text. Rather, the model is a network of statistical weights and parameters that describe relationships, mathematical representations of images, not a copy of the data itself.

The judge found the model never "saw" or "retained" an image in the way a human does, so it couldn't be considered a copy.

While the court affirmed that an AI model could, in principle, be something capable of infringing copyright, in this specific instance, it was not.

Getty’s trademark infringed

This left only the issue of trademark violation.

Getty alleged that the Stability AI software could reproduce images from its library, with a version of the Getty watermark.

While the generated watermark was frequently distorted (for instance, "Getty" sometimes had three 'T's), the judge found that Stability AI had indeed violated trademark laws by reproducing Getty’s logo.

But this concerned Getty’s trademark and not its photos themselves.

So, what does this mean for those of us in the web scraping and data gathering space?

1. Jurisdiction is paramount

Stability AI averted the primary infringement claim because the training occurred entirely outside the UK.

This underscores a critical point: where the copying for training purposes takes place determines which laws apply. Developers using training processes that don't involve the tool storing or reproducing the data itself may not be subject to UK copyright law if the training happens elsewhere.

2. The ‘mathematical representation" distinction is key (for now)

The UK court's acceptance that the model contains learning patterns, not storing content, is a significant, if temporary, victory for AI developers. The ruling that a model containing mathematical weights and parameters is not an "infringing copy" of the data it was trained on provides a potential defense against claims of secondary copyright infringement in the UK.

3. Beware of trademarks

The one area where Stability lost was on the reproduction of Getty's watermarks. Data gatherers and model trainers must be vigilant about the potential for their models to reproduce trademarked elements in their outputs. This could necessitate filtering training data or fine-tuning models to avoid this specific type of infringement.

4. Don't ignore international implications

While this ruling offers some respite in the UK, it's not a global get-out-of-jail-free card. EU law subjects the marketing of an AI system in the European Union to compliance with EU copyright rules. This means that even if you train your model in the US, you will still have to comply with EU rules if you want to market it there.

The legal landscape is still evolving, and a US-based training operation doesn't grant a free pass to enter the EU market without careful legal consideration.

On the matters actually dealt with, the case creates some precedent in the UK.

But whether the act of scraping and training an AI model on copyrighted data within the UK constitutes infringement was not answered by the court.

While the ruling is a key decision and provides valuable guidance, it has been described by some as a “damp squib” because it didn't rule on the core issue of training.