Yes, artists can sue over training data, the images and text copied to teach an AI model. They must prove the company copied their protected work and defeat a fair-use defense.
The answer is not automatic payment for every artist whose work appears online. It turns on proof of ownership, copying, and how the company obtained the work. Sourcing and evidence decide more than anger about scraping.
Official resources:
- Read the official guidance from Berkeley — Use this primary source to verify the official guidance.
- Read the original research from Nyu — Use this page to review the original research.
Table of Contents
- What the image-generator lawsuit claims
- Can a training-data claim move forward?
- When is training fair use?
- What should artists do now?
What the image-generator lawsuit claims
Visual artists Sarah Andersen, Kelly McKernan and Karla Ortiz sued Stability AI, Midjourney and DeviantArt in January 2023. They alleged image generators were trained on billions of internet-scraped images. The NYU Journal of Intellectual Property and Entertainment Law describes the claim in its Andersen case summary as use without consent, credit or compensation.
That theory targets direct copying during training, not every output that looks similar. An artist still needs rights in the work and proof the training set included it. Style alone, without copied protected expression, is a weak basis for payment.
Can a training-data claim move forward?
Judge William Orrick let core direct-copyright claims in Andersen v. Stability AI survive dismissal. The case moved toward discovery and a jury trial then reported for 2026-2027. The NYU Journal of Intellectual Property and Entertainment Law explains that path in its dismissal analysis, while stressing artists must still prove copying and rights.
Surviving dismissal is not a win on liability. It means the court saw enough to allow evidence gathering and testing by a jury. Weak documentation, licensed transfers, or independent creation can still defeat a claim. Any future payment would depend on court-approved notice and proof you fall in a covered set. Save evidence now instead of responding to ads promising guaranteed checks.
When is training fair use?
On June 23, 2025 Judge William Alsup ruled in Bartz v. Anthropic that training language models on books was highly transformative fair use. The same ruling held that downloading millions of pirated Books3 and LibGen copies for a permanent library was not fair use.
The Guide to Artificial Intelligence details that split in its 2025 rulings guide. The lesson is narrow and practical. A transformative purpose can protect training, while pirated sourcing can undo that protection. Courts weigh how the defendant got the works, not only what the model learned.
What should artists do now?
Anthropic then agreed to pay $1.5 billion for about 465,000 to 500,000 works. The deal was preliminarily approved on Sept. 25, 2025. The Society of Authors describes it in its settlement report as the largest U.S.
copyright settlement, tied to piracy-based sourcing. That result does not by itself pay artists whose images were scraped. It shows payment risk grows when a permanent library came from piracy. Artists who want to preserve options should: Do not pay a service to file for you based on a press headline. Keep dated source files together; without them, you cannot tie your work to any future covered set.
- Keep dated originals, project files, and publication records.
- Save copyright registrations, licenses, and takedown history.
- Record where work appeared and whether scraping was allowed.
- Follow court dockets for approved notice before sharing bank details.