AI art lawsuits are a wave of federal cases in which visual artists, photo agencies, and other rightsholders accuse image-generating AI companies of copying their work without permission to train commercial models. The central legal question in nearly every suit is whether that mass copying qualifies as fair use under federal copyright law. Early rulings have started to arrive, and the first full trial against an image generator, Andersen v. Stability AI, is scheduled to begin September 8, 2026.
What Artists Are Actually Suing Over
The core allegation across these cases is simple: the AI company copied enormous quantities of copyrighted images, usually by scraping the open web, and fed them into a training dataset to build a model that now generates images commercially. Federal copyright law gives creators the exclusive right to reproduce their work, and each image pulled into a training set is, in the plaintiffs’ framing, an unauthorized copy.1Office of the Law Revision Counsel. 17 U.S. Code 106 – Exclusive Rights in Copyrighted Works
That framing matters because statutory damages for willful infringement can reach $150,000 per work.2Office of the Law Revision Counsel. 17 U.S. Code 504 – Remedies for Infringement: Damages and Profits Training sets contain millions or billions of images, so the theoretical exposure is enormous. The AI companies don’t generally deny the copying happened. They argue it was legal, and they lean hardest on one defense.
Fair Use Is the Fight
Fair use is the doctrine that could insulate AI companies from liability even if they copied protected images. Courts weigh four factors together: the purpose and character of the use, the nature of the work copied, how much was taken, and the effect on the market for the original.3Office of the Law Revision Counsel. 17 U.S. Code 107 – Limitations on Exclusive Rights: Fair Use No factor is decisive on its own.
The first factor asks whether the new use is “transformative,” meaning it serves a different purpose than the original rather than substituting for it. The Supreme Court narrowed this analysis in 2023 in Andy Warhol Foundation v. Goldsmith, holding that when the original work and the secondary use share “the same or highly similar purposes” and the secondary use is commercial, the first factor weighs against fair use.4Supreme Court of the United States. Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith AI companies argue that training a model is fundamentally different from displaying or selling the copied images. Artists counter that the end product competes directly with them for the same illustration and photography commissions.
The fourth factor, market impact, often carries the most weight. In Thomson Reuters v. Ross Intelligence, a federal court in Delaware ruled that using copyrighted legal headnotes to train a competing AI research tool was not fair use, precisely because the AI was built to compete in the same market as the original product. The court also found that the potential market for AI training data itself could count against fair use, even if the rightsholder had not yet licensed its data for that use.5United States District Court for the District of Delaware. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc.
Pulling the other direction is the Second Circuit’s 2015 decision in Authors Guild v. Google, which held that mass digitization of copyrighted books to build a search index was fair use because the purpose was “highly transformative” and the system did not provide a “substantial substitute” for the originals.6Justia Law. Authors Guild v. Google, Inc., No. 13-4829 (2d Cir. 2015) AI companies rely heavily on this precedent, treating training copies as an intermediate step rather than the end product.
What Early AI Rulings Have Said
The first direct rulings on AI training came from federal courts in the Northern District of California in cases involving text models rather than image generators. In Kadrey v. Meta Platforms, the court granted summary judgment to Meta, finding that copying books to train its Llama language models was “highly transformative” and that copying entire works was “reasonably necessary” for that purpose. The plaintiffs failed to show sufficient market harm.
In Bartz v. Anthropic, a separate court went further, calling AI training “quintessentially transformative” and holding it was fair use as a matter of law. But the Bartz court drew a sharp line on sourcing. While training itself was fair use, Anthropic’s decision to obtain some of its training books from pirated websites was not. Copying from a pirate site to build a central library, the court held, is “inherently, irredeemably infringing” regardless of what happens afterward. How the training data is acquired, in other words, can matter as much as what the model does with it.
These are trial-court decisions, not binding nationwide, and both involved text. Whether the same reasoning will carry over to image generators is an open question. Image models produce outputs that compete more directly with their source material than a chatbot does, which could shift the fourth-factor analysis when a court finally rules on an image case.
Other Claims Riding Alongside Copyright
Several of the complaints add a Digital Millennium Copyright Act claim for removing copyright management information. Federal law prohibits intentionally stripping identifying data like artist names, watermarks, or licensing terms from copyrighted files.7Office of the Law Revision Counsel. 17 U.S. Code 1202 – Integrity of Copyright Management Information Artists allege that AI companies strip this metadata when they process scraped images into training sets. Statutory damages run from $2,500 to $25,000 per violation, separate from any infringement damages.8Office of the Law Revision Counsel. 17 U.S. Code 1203 – Civil Remedies The hurdle is proving the removal was done knowingly and to conceal infringement; courts have not resolved how that intent standard applies to automated AI pipelines.
Another set of claims goes after style mimicry. When a user prompts an AI to produce an image “in the style of” a named artist, the output may not copy any single work but trades on the artist’s reputation. Plaintiffs invoke the Lanham Act’s bar on false designations of origin, arguing that style mimicry can confuse consumers about who created or endorsed the result.9Office of the Law Revision Counsel. 15 U.S. Code 1125 – False Designations of Origin and False Descriptions Forbidden State right-of-publicity statutes add another layer, though courts have sometimes held that when a publicity claim overlaps with rights covered by the Copyright Act, federal law preempts the state claim. For style-based claims in particular, the boundaries are unsettled.
A Related Question: Can AI-Generated Art Itself Be Copyrighted?
This question comes up alongside the lawsuits and is often confused with them, but it is a separate issue. The U.S. Copyright Office has held since at least 2023 that copyright registration requires human authorship, so a work generated solely by an algorithm does not qualify.10U.S. Copyright Office. Compendium of U.S. Copyright Office Practices, Third Edition – Chapter 300 Copyrightable Authorship The Office reaffirmed that position in its 2025 report on AI and copyrightability.11U.S. Copyright Office. Copyright and Artificial Intelligence, Part 2 Copyrightability Report
In March 2025, the D.C. Circuit affirmed the denial of Stephen Thaler’s attempt to register an image generated entirely by his “Creativity Machine,” noting that the Copyright Act’s references to an author’s “life,” “widow or widower,” “domicile,” and capacity to sign documents all presume a human being.12United States Court of Appeals for the District of Columbia Circuit. Thaler v. Perlmutter When a human uses AI as a tool, the line is murkier. In its review of the graphic novel Zarya of the Dawn, the Copyright Office protected the human-written text and the overall arrangement of the book but denied copyright for the individual Midjourney-generated images, finding the author had not exercised enough creative control over the specific visual outputs.13United States Copyright Office. Zarya of the Dawn (Registration # VAu001480196) The more a human directs the creative choices in the final work, the stronger the copyright claim.
What Artists Are Asking Courts to Order
Standard copyright remedies include actual damages, disgorgement of profits, and attorneys’ fees. If plaintiffs elect statutory damages, the $150,000-per-work ceiling for willful infringement multiplied across millions of training images produces the eye-popping figures that show up in headlines.2Office of the Law Revision Counsel. 17 U.S. Code 504 – Remedies for Infringement: Damages and Profits
Several complaints go further and ask for model destruction, meaning a court order requiring the company to delete any AI model trained on infringing material. The New York Times complaint against OpenAI, for example, seeks “destruction of all GPT or other LLM models and training sets that incorporate Times Works.” No court has ordered that yet, and the statute’s language on destruction is permissive rather than mandatory, so a judge would apply the usual test for injunctive relief before granting it. The threat alone, however, gives AI companies a strong reason to settle. Anthropic has already reached a class-wide copyright settlement establishing a fund for rightsholders whose books appeared in its training data.
Where the Major Cases Stand in 2026
Andersen v. Stability AI, filed in early 2023 by a group of visual artists against Stability AI, Midjourney, and DeviantArt, is the case most of the art world is watching. The court dismissed weaker theories like vicarious infringement but let the direct copyright claims proceed. The case is now in discovery, and trial is set to begin September 8, 2026.
Getty Images has pursued Stability AI on two fronts. In the United Kingdom, a trial concluded in late 2025 and the judgment was largely unfavorable for Getty: the training-related claim was abandoned during trial, and the court found that most versions of Stability’s model did not reproduce Getty’s watermark frequently enough to meet the trademark infringement threshold. In the United States, Getty filed a separate suit in August 2025 in the Northern District of California. A judge partially granted Stability AI’s motion to dismiss in April 2026, and surviving claims are in discovery, with fact-finding scheduled to close by September 2026.14CourtListener. Getty Images (US), Inc. v. Stability AI, Ltd.
Kadrey v. Meta narrowed to a single direct-infringement claim after other theories were dismissed, and the court’s subsequent fair use ruling for Meta may effectively end the case. New York Times v. OpenAI is still in its early stages, with a second amended complaint filed in May 2025 and the matter consolidated into multidistrict litigation in the Southern District of New York. More than 50 copyright suits against AI companies are now pending in federal courts, and the outcomes of the first trials will shape whether the rest settle, proceed, or collapse.
Courts so far have been more receptive to claims about the copying that happens during training than to claims about what the AI produces afterward. Proving that a specific output is “substantially similar” to a specific copyrighted work has been a high bar, and most derivative-work claims have been dismissed at the pleading stage. The cases that survive focus on the inputs: what was scraped, how it was processed, and whether the source of the training data was legitimate.