Artists sue Stability AI, Midjourney and DeviantArt over training data

Background

Text-to-image generators produce a picture from a written prompt. To do that, they are trained on very large collections of images: the model repeatedly adjusts its internal settings so that its output comes closer to the patterns found in that collection. Whether such a collection may be assembled and used in this way is a question of copyright law, and it is the question at the centre of this case.

The suit

On 13 January 2023 the artists Sarah Andersen, Kelly McKernan and Karla Ortiz filed a class-action copyright suit against Stability AI, Midjourney and DeviantArt.^[1][4] A class action is a lawsuit brought by a small number of people on behalf of a much larger group in the same situation; here the proposed group was artists whose work, the plaintiffs argued, had been used to build the companies' image generators without permission.^[1]

The complaint was filed by Matthew Butterick, a lawyer and typographer, together with the Joseph Saveri Law Firm, which specializes in antitrust and class-action cases.^[5] The plaintiffs alleged that the companies infringed the rights of millions of artists by training their tools on five billion images scraped from the web without the consent of the original artists.^[5] "Scraping" in this context means gathering images automatically from websites at scale, rather than licensing them one by one from the people who made them.

The July 2023 ruling

In July 2023 Judge William Orrick dismissed most of the claims but allowed the artists to file a new complaint.^[2] A ruling at this stage is about whether the complaint states its claims clearly enough to go forward, not a full examination of the evidence on both sides. Because the artists were permitted to refile, the dismissal did not end the dispute; it moved the case into a further round of written pleadings.

The amended complaint

In November 2023 the artists filed an amended complaint, which added Runway as a defendant.^[3][4] The amended filing kept the case alive against a somewhat different set of parties and gave the plaintiffs another chance to put their allegations in the form the court had asked for.

Why it matters

Generative image tools were built on training collections far larger than any single studio or publisher could assemble by negotiating licences. That scale is exactly what makes the legal question hard: the individual contribution of any one image to a trained model is diffuse, while the collection as a whole is what gives the model its capability. The suit put that tension into a courtroom, asking whether the way these tools were built is compatible with the rights of the people whose images were used.^[1][5]

The case did not by itself settle how copyright law treats training data. What it did do was make the question concrete and public: the same dispute over scraping, consent and the rights of millions of artists has since shaped how developers describe their data sources and how artists organise to defend their work. Following the July ruling and the November amendment, the argument continued over what exactly the plaintiffs had to show — a procedural question with large consequences for anyone building or licensing generative models.^[2][3]