Lawsuits Put Generative AI Training to the Test

Generative AI has run into one of its first major legal battles. Visual artists and image companies argue that several models were trained on copyrighted works without authorization, while tech companies face a question that courts have yet to answer clearly: does using millions of works to teach a machine amount to copyright infringement?

The issue extends well beyond the creative sector. The same training methods are used to build systems that generate text, music, code and video. What courts decide about images could help set the economic rules for a large part of artificial intelligence.

Artists take on the creators of Stable Diffusion

On January 13, artists Sarah Andersen, Kelly McKernan and Karla Ortiz filed a proposed class action in a federal court in California against Stability AI, Midjourney and DeviantArt. The case focuses on Stable Diffusion, a model capable of creating images from written prompts, as well as products built on similar technologies.

The lawsuit claims that Stable Diffusion was trained on a vast collection of images gathered from the internet, including copyrighted works. The model learns statistical patterns from those materials to produce new images: it does not work like a search engine retrieving a specific file, but rather as a system that generates pixels based on what it learned during training.

That technical distinction will be central to the litigation. The fact that a generated image is not an exact copy does not, by itself, settle the question of where it came from. The plaintiffs argue that making copies during training and the subsequent commercial use of the models infringe their rights. They also challenge tools that allow users to request results in the style of identifiable artists.

Stability AI has publicly argued that Stable Diffusion does not store images in a database or reproduce a catalog of works. That explanation is relevant, but it does not automatically determine whether the training was lawful. Judges will have to distinguish between how the model works technically, the copies needed to process the data and the kinds of results it can deliver to users.

Getty Images takes the dispute to the UK

The dispute is not limited to a lawsuit brought by individual creators. Getty Images announced on January 17 that it had launched legal proceedings against Stability AI in the High Court of Justice in London. The agency accuses the company of copying and processing millions of images from its catalog without a license.

Getty occupies a different position from many artists. It manages a commercial archive, licenses photographs to media organizations and businesses, and has visible branding on some of its images. The case therefore adds two elements to the debate: the possible reproduction of copyrighted images and the use of material associated with a licensing platform.

The conflict exposes a business-model tension. AI companies need huge volumes of data to build competitive systems. Rights holders, meanwhile, are watching works that once had licensing value feed tools capable of generating alternatives in seconds and at low cost.

There is no single legal answer

Tech companies typically argue that training may qualify as transformative use: the work is analyzed to extract relationships and patterns, not redistributed. In the United States, that debate is tied to the fair use doctrine, an exception that permits certain unauthorized uses and is assessed on a case-by-case basis.

But a technological purpose does not make every use lawful. Courts generally consider, among other factors, the nature of the use, the amount of the work used and its effect on the original market. Training a model on millions of images raises the problem of scale: each individual work may have only a minimal influence on the output, but the collection as a whole is essential to building the product.

In Europe, exceptions for text and data mining — the automated analysis of large collections of information — provide a different framework. The European copyright directive establishes a broad exception for such analysis, while allowing rights holders to expressly reserve certain commercial uses of their works. How that framework applies in practice to generative models has yet to be determined.

Licensing, transparency and a long battle ahead

The lawsuits filed so far will not resolve the issue immediately. Before any trial, courts will have to decide which claims can proceed and what facts need to be proven. The outcome could take years.

In the meantime, the pressure is already pushing the industry toward two changes. The first is the search for licensed, public-domain or purpose-built training data. The second is greater transparency around datasets: knowing what materials go into a model will be crucial for authors, customers and companies that want to use AI without taking on legal risks that are difficult to measure.

The question is not whether generative AI will keep advancing, but under what arrangements it will do so. If large-scale training requires licenses, compensation or effective mechanisms for excluding works, the cost and availability of data will become as strategic as chips and computing power.