Active

Getty Images v. Stability AI

Exploring the legal boundaries of training datasets and the commercial viability of fair use when generative output substitutes for licensed imagery.

Editorial illustration — pixelated imagery fading into a cobalt grid

Getty Images v. Stability AI · editorial illustration

Background

I · The facts

Getty alleges Stability AI copied more than 12 million photographs, captions, and metadata from its collection to train Stable Diffusion. The case tests whether mass ingestion of licensed imagery for commercial model training can ever qualify as fair use.

“The unauthorized use of Getty's intellectual property represents a calculated risk that threatens the very foundation of the commercial photography ecosystem.”

— Getty Images complaint, 1:23-cv-00135, ¶ 4

The Claims

II · What the plaintiff argues
  1. Copyright Infringement. Systematic, commercial-scale copying of more than 12 million original photographs for training Stable Diffusion without license or attribution.
  2. Trademark Dilution. Outputs that contain distorted but recognizable Getty Images watermarks, damaging the marketplace's reputation and the distinctive mark itself.
  3. Removal of CMI. Alleged deletion of Copyright Management Information (metadata, creator attribution) during the ingestion process — a standalone violation under 17 U.S.C. § 1202.

The Defense

III · What the defendant argues

Stability AI contends that training on publicly available web images is a transformative use under 17 U.S.C. § 107. They argue that the model does not store or reproduce the images but rather extracts statistical relationships between pixels and concepts — mathematical abstractions, not copies. On trademark, Stability argues the watermark artifacts are unintentional memorization that the company has actively filtered against.

Analysis

IV · Our read

Unlike Authors Guild v. Google, where snippets served an informational purpose, Stable Diffusion produces outputs that compete directly with the licensed marketplace Getty built. That shifts the fourth fair-use factor — market impact — squarely against Stability.

The CMI removal claim is the most quietly dangerous to Stability. Unlike copyright infringement (which requires showing substantial similarity to specific works), § 1202 is triggered by the act of stripping metadata at ingestion, with statutory damages of $2,500–$25,000 per violation.

The Getty watermark memorization argument collapses Stability's "mathematical abstraction" framing. If the model can reproduce a distinctive mark, it can reproduce more than mathematics.

What this means for you

A Getty win would establish that commercial-scale ingestion of licensed works requires a license — and would likely force every major AI lab into retroactive licensing negotiations with stock-image marketplaces. A Stability win on fair use but a loss on § 1202 would still reshape how training pipelines handle metadata.

See our compliance guides

Timeline

V · Key dates
2023-02
Complaint filed

Getty files in US District Court for the District of Delaware.

2023-05
Motion to dismiss

Stability AI moves to dismiss on personal jurisdiction grounds.

2023-12
Jurisdictional ruling

Court denies most of the motion; the case proceeds.

2024-Q3
Discovery begins

Parties exchange training-data documentation and model weights under protective order.

2025-11
Summary judgment briefing

Cross-motions on fair use and § 1202 filed.

2026-Q3
Trial (expected)

Jury trial anticipated on remaining claims.

Back to case tracker