Andersen v. Stability AI: Key Rulings and Fair Use

Andersen v. Stability AI is a class-action copyright lawsuit filed in January 2023 by visual artists against several generative AI companies, alleging that their copyrighted images were copied without permission to train AI image generators. The case is pending in the U.S. District Court for the Northern District of California before Judge William Orrick. As of mid-2026, it is in discovery, with a summary judgment hearing scheduled for February 17, 2027.

Who Is Suing Whom

The lawsuit was filed by three visual artists: Sarah Andersen, Kelly McKernan, and Karla Ortiz. Ortiz is a San Francisco concept artist whose credits include Marvel films such as Guardians of the Galaxy Vol. 3, Doctor Strange, Black Panther, and Avengers: Infinity War, along with games including Final Fantasy 16.¹ A first amended complaint expanded the group to ten artists and added Runway AI as a defendant.² The plaintiffs are represented by attorney Matthew Butterick and the Joseph Saveri Law Firm.

The original defendants are Stability AI, which makes the Stable Diffusion model and the DreamStudio platform; Midjourney; and DeviantArt, which launched a product called DreamUp. Runway AI was added based on allegations that it worked with Stability AI to train Stable Diffusion and then distributed it through its own “AI Magic Tools” product.³⁴

What the Artists Are Claiming

At the center of the case is the LAION-5B dataset, a collection of roughly 5.85 billion image-text pairs scraped from the internet and used to train Stable Diffusion and related models.⁵ The plaintiffs allege these images were taken without permission and that the resulting AI models function as “compressed copies” of the training data. When a user enters a prompt, the plaintiffs argue, the model draws on those compressed representations to produce images that compete with the originals. The complaint described the tools as a “21st-century collage tool.”⁶

The complaint asserts several legal theories:

Direct and vicarious copyright infringement for downloading billions of images, storing them within the models, and producing derivative outputs.
Induced copyright infringement, based on distributing Stable Diffusion to other companies and the public.
DMCA violations under Section 1202 for removing copyright management information from training images and omitting it from outputs.
Right of publicity claims, tied to prompts requesting images “in the style of” specific named artists.
Unfair competition under California law.
Unjust enrichment.

The complaint also seeks class certification and proposes multiple classes, including a “Midjourney Named Artist Class” for artists whose names appeared on a list Midjourney used to guide its model.⁷ As of mid-2026, no class has been certified.⁶

What the Court Has Kept and What It Has Thrown Out

Judge Orrick has ruled on two rounds of motions to dismiss, and those rulings have shaped what remains.

October 2023

The first ruling dismissed most claims against Midjourney and DeviantArt with leave to amend, dismissed copyright claims from plaintiffs who lacked formal copyright registrations, and dismissed the right of publicity and unfair competition claims as “factually inadequate.”⁸ The direct copyright infringement claim against Stability AI for training its model survived.

August 2024

On August 12, 2024, Judge Orrick issued a more detailed ruling on the first amended complaint. Two theories of direct infringement were found plausible. Under the “model theory,” the AI model itself is alleged to be an infringing copy because it embodies transformations of copyrighted works. Under the “distribution theory,” distributing the model amounts to distributing the copyrighted works compressed inside it. The court noted that both theories depend on whether the plaintiffs’ protected works are contained “in some manner” within the AI systems, a factual question to be resolved later.⁷

The induced infringement claim against Stability AI also survived. The court pointed to statements from Stability AI’s own CEO about the model’s ability to compress and reproduce training images, and to academic research showing that outputs could sometimes reproduce training data given precise prompts. Judge Orrick distinguished generative AI from older technologies like VCRs, suggesting that if Stable Diffusion was “created to facilitate that infringement by design,” it could not claim the same safe harbor.³

Claims against Runway AI proceeded as well. The court rejected Runway’s argument that works stored as algorithmic or mathematical representations inside a model are not “fixed” in a copyrightable medium, and it rejected Runway’s reliance on the Kadrey v. Meta ruling, finding that image generators are “materially different” from the large language models at issue there.³

Other claims did not survive. The DMCA claims were dismissed with prejudice. On the Section 1202(b) claims for removal of copyright management information, the court followed Doe 1 v. GitHub, Inc. and held that the statute requires the allegedly infringing copy to be “identical” to the original, which no plaintiff had alleged. The Section 1202(a) claims about false copyright information failed because the court found it implausible that a generic license used by Stability AI would be read as a claim of rights over the plaintiffs’ works. The unjust enrichment claims were dismissed as preempted by the Copyright Act, though the court left the door open to repleading if the plaintiffs could articulate an “extra element” tied to their “artistic personas” that went beyond standard copyright protection.⁷

Where the Case Stands Now

The plaintiffs filed a second amended complaint on October 31, 2024, and a third amended complaint on February 27, 2026, with defendants answering each.⁹

Discovery has produced disputes that reflect how technical the case has become. In March 2026, Magistrate Judge Cisneros granted the plaintiffs’ request for a list of Stability AI’s enterprise customers, ruling the information relevant to the fair use defense and potential market harm. Other disputes have involved Midjourney’s training-data filtering practices and DeviantArt’s discovery custodians.¹⁰

As of January 2026, the parties reported that discovery was “moving along without major disagreement.”¹¹ In February 2026, Judge Orrick granted the plaintiffs’ request to push scheduling deadlines back roughly three months. The summary judgment hearing moved from November 4, 2026 to February 17, 2027, and the trial date originally set for September 8, 2026 was effectively continued.¹² The most recent docket activity occurred on May 7, 2026.¹³ No settlement talks have been publicly reported.

The Fair Use Question

Fair use will likely decide the case. The defendants have not yet formally litigated the defense at summary judgment, but it is expected to be the central battleground. Arguments will focus on whether AI training and the resulting outputs are “transformative” enough to justify unlicensed use of copyrighted works.³ The discovery order compelling Stability AI to produce its enterprise customer list was tied specifically to fair use Factor 4, which examines the effect of the use on the potential market for the original works.¹⁰

How Andersen Fits Alongside Other AI Copyright Cases

When it was filed in January 2023, Andersen was among the earliest lawsuits challenging AI training on copyrighted material. Several related cases have advanced further:

Getty Images v. Stability AI, originally filed in the District of Delaware over more than 12 million photographs, was voluntarily dismissed in August 2025, with Getty saying it would refile in the Northern District of California.¹⁴
Thomson Reuters v. Ross Intelligence produced a partial summary judgment for Thomson Reuters in February 2025, finding that using Westlaw headnotes to train an AI legal search tool was not fair use. The case is on appeal before the Third Circuit.¹⁵
Bartz v. Anthropic, a class action by nearly 500,000 authors over the Claude model, produced a ruling that AI training itself was transformative fair use but that storing pirated copies of books was not. The case settled for $1.5 billion.¹⁵
In re OpenAI Copyright Infringement Litigation consolidates twelve cases against OpenAI, with the court overseeing discovery involving 108 million output logs as of March 2026.¹⁵
Disney and other entertainment companies have separately sued Midjourney over training and derivative outputs.¹⁵

None of these decisions has definitively resolved whether training on copyrighted data is fair use. The Bartz settlement and the Thomson Reuters ruling suggest courts are drawing lines on specific facts rather than issuing sweeping rules. The February 2027 hearing in Andersen will put the same question to a court that has already said the answer turns on whether copyrighted works are, in some form, inside the models themselves.

1
KQED. Artists AI Lawsuit Artificial Intelligence Stable Diffusion DeviantArt
2
Matthew Butterick. The End of the Beginning
3
Copyright Alliance. Andersen v. Stability AI Copyright Case
4
University of California, Irvine. Andersen v. Stability AI Ltd (Redacted)
5
Knowing Machines. Knowing Machines USCO Comment
6
IPWatchdog. Andersen et al v. Stability AI Complaint
7
Justia. Andersen et al v. Stability AI Ltd. et al, Document 223
8
Bloomberg Law. Generative AI Complaint Advisory
9
MeShip Law. Andersen v. Stability AI Litigation Tracker
10
ChatGPT Is Eating the World. Andersen v. Stability AI Category
11
ChatGPT Is Eating the World. Sarah Andersen v. Stability AI Filed Status Report
12
ChatGPT Is Eating the World. Sarah Andersen’s Copyright Lawsuit Gets Pushed Back Again
13
CourtListener. Andersen v. Stability AI Ltd. Docket
14
Baker McKenzie. Getty Images v. Stability AI
15
Norton Rose Fulbright. AI in Litigation Series: An Update on AI Copyright Cases in 2026