Creative works have always been a medium of human expression, creativity, and innovation. In recent years, the creative industries have been reshaped by the rise of generative artificial intelligence (AI). AI-generated content is the product of machine learning models trained on vast quantities of existing works, which can then produce new images, music, and video in response to simple prompts. Programs like Midjourney, Stable Diffusion, and image generators built into ChatGPT and Google Gemini have put this capability into the hands of millions of users. Although generative AI has gained widespread attention for its ability to create images, musical compositions, and even animations that rival human-created work, this innovative technology has raised many complex and novel issues surrounding copyright law.
Recent cases have begun to address these issues, but the law remains unsettled, especially where AI-generated outputs compete directly with human-created works or established licensing markets.
Machine Learning and Copyright Infringement
Machine learning, a subset of AI, enables computer programs to learn from specific datasets and improve automatically, without being explicitly programmed. Machine learning algorithms require large amounts of data to learn from, and this data often includes copyrighted material such as images.
Stability AI, a leading AI company, created Stable Diffusion, a text-to-image model capable of generating images from written prompts. Stable Diffusion was allegedly trained on billions of image-text pairs drawn from large datasets that included material scraped from websites.
In 2023, Getty Images sued Stability AI, claiming that Stability AI used datasets that contained thousands of copyrighted images and text that were copied from its website. In November 2025, the UK High Court issued the first ruling in the dispute. Getty largely lost on the claims that remained before the court, but the ruling did not decide whether training an AI model on copyrighted images is lawful. Getty dropped its main copyright claim after admitting it couldn’t prove the training happened in the UK, so the court never actually decided whether training an AI on copyrighted images is legal. On the claim that was left, the court found that Stable Diffusion’s underlying model is not an “infringing copy” of Getty’s images. In simple terms, a trained AI model picks up patterns from the images it learns from, but it doesn’t keep copies of the images themselves. Getty’s only win was a small one: early versions of Stable Diffusion sometimes produced images with distorted Getty watermarks on them, and the court said that amounted to trademark infringement in certain limited situations. Getty is appealing the decision, and its separate U.S. lawsuit, which will directly address the fair use question the UK court never answered, is scheduled for trial in 2028.
At the heart of that case is a question courts have yet to fully resolve: can the use of copyrighted works to train AI models be excused as fair use? To understand why the answer is so contested, it helps to start with what fair use actually requires.
What is Fair Use?
Fair use is one possible defense against copyright infringement claims involving AI-generated content and AI training. This legal doctrine allows limited use of copyrighted material without permission from the owner, under certain circumstances. The purpose of this doctrine is to balance the protections that copyright grants to its owners with the public interest in promoting creativity and allowing the free exchange of ideas.
Courts are just beginning to answer whether AI companies can legally train their models on copyrighted works, and so far the answer is: it depends on the facts. In one case against Anthropic, a court found that using books to train an AI model was allowed because the AI transforms those works into something new rather than simply copying them. However, the court treated the use of lawfully acquired books differently from the alleged use of pirated copies, which raised separate copyright concerns.
A court reached a similar result in a case against Meta, but mostly because the authors suing didn’t show that AI training actually hurt their book sales — the judge hinted that with better evidence, authors could win.
What Factors Are Used to Determine Fair Use?
The fair use doctrine, codified at 17 U.S.C. § 107, requires courts to weigh four non-exclusive factors to determine whether a particular use of copyrighted work is a fair use and therefore not infringement:
(1) the purpose and character of the use, including whether it is commercial or nonprofit;
(2) the nature of the copyrighted work;
(3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
(4) the effect of the use upon the potential market for or value of the copyrighted work
Can AI Training Be Considered Transformative?
Although machine learning algorithms can generate images that are similar to the original works, they also have the ability to produce new creative works. This raises questions about the extent to which the resulting AI training and AI-generated outputs can be considered transformative or whether they are simply derivative works that infringes on the original copyright. To benefit from the protections of the fair use doctrine, machine learning algorithms, like Stable Diffusion, need to demonstrate that there is a transformative use. Transformative use involves using the copyrighted work in a way that creates something new and adds value, rather than simply copying the original work.
Getty Images claims that Stable Diffusion can produce images that are highly similar to and derivative of its copyrighted work. Stability AI’s position, however, is that the purpose of the model is not to replicate any original work, but to give users a tool to create new, original works — an infinite variety of images and styles — without the need for expensive equipment or specialized training.
For years, the closest analogy was Google Books, where a court found fair use because Google copied books to build a searchable database — a purpose entirely different from the authors’ expressive one. The recent AI training decisions embrace this logic: the Anthropic court called AI training “spectacularly transformative” because the model learns statistical patterns from works rather than reproducing them, and the Meta court agreed.
But transformative purpose is no longer the end of the analysis. The Anthropic court held that training on lawfully purchased books was fair use, while downloading pirated copies was separate infringement that transformativeness could not cure — a distinction that led to a $1.5 billion settlement. The Meta court, meanwhile, called market harm the most important factor, suggesting that AI outputs flooding the market with competing works could defeat fair use on stronger evidence. That may matter greatly for Getty: unlike the book authors, it runs a licensing business that AI-generated imagery competes with directly. The fair use question will likely turn on how Stability AI obtained its training data and whether Getty can prove the market harm the authors could not.
In conclusion, a finding of fair use could help create new possibilities for creativity and expression. As AI-generated content continues to advance and transform the field of art, AI companies are seeking clarity that their methods of training models will not be subject to copyright infringement claims. This would be a more practical solution, rather than having to get approval to use every single copyrighted work. At the same time, courts will have to balance that interest against the rights of artists, photographers, and other copyright owners whose works are used to train these models.
Contributions to this blog by Hannah Warther.