Training Data & Intellectual Property in Generated Content

Beijing AI Text-to-Image Case

The court found the disputed image reflected the user’s intellectual input and individual expression, considering prompts, parameter adjustments and the generation process. Protection depended on the specific human contribution in this case; the decision does not make every AI-generated image copyrightable.

Effective first-instance judgment

• The disputed image satisfied the originality requirement.
• The model lacked legal personality; the human contribution in this specific generation process supported the finding of authorship.

Guangzhou Ultraman AI Image Case

The defendant connected to a third-party AI image service through an API. Users could generate images substantially similar to the protected Ultraman works. The court classified the defendant as a generative AI service provider and found infringement of the reproduction and adaptation rights.

First-instance judgment

• The generated images infringed the reproduction and adaptation rights in the Ultraman works. The service provider had not itself trained the model at issue.
• The court ordered technical measures to prevent further infringing outputs and RMB 10,000 in damages and reasonable costs.

UK Getty Images Case

Getty challenged training, model distribution, outputs and trademark use involving Stable Diffusion. It abandoned direct copyright and database-right claims at trial without a merits determination. The High Court rejected secondary copyright infringement and found limited trademark infringement in watermark outputs of specified older versions.

High Court judgment; Getty permitted to appeal secondary copyright ruling

• On the particular model and evidence, weights were not found to be infringing copies for the secondary copyright claim. This is not a technical finding about every AI model.
• Limited watermark outputs in specified older versions infringed trademarks.

U.S. ROSS Training Data Case

ROSS used materials derived from Westlaw headnotes to train a competing legal research tool. On September 29, 2026, the Third Circuit affirmed the district court’s rulings on originality and fair use. The decision concerns a non-generative AI system and an interlocutory appeal; it does not resolve every remaining issue in the litigation.

Third Circuit affirms partial summary judgment

• The Third Circuit affirmed originality of the 2,243 headnotes at issue and the conclusion that ROSS’s use was not fair use.
• The holding depended on the non-generative research tool and specific competitive and market evidence; it is not a universal ruling on all AI training.

U.S. Anthropic Book Training Case

On the specific facts, the court distinguished model training, one-for-one digitisation of purchased books with destruction of the originals, and a library acquired from pirated sources. It later finally approved a USD 1.5 billion class settlement of defined past input and copying disputes; approval did not license future conduct.

Class settlement finally approved on 20 July 2026; implementation underway

• The 2025 partial judgment found fair use for training and replacement digitisation on those limited facts, without extending that defence to the permanent pirated-source library.
• On 20 July 2026, the court finally approved USD 1.5 billion for 482,460 listed works. Implementation continued; approval does not establish that distributions are complete.

U.S. Meta Book Training Case

On evidence submitted by 13 authors, the court entered summary judgment for Meta on copying books to train Llama. It did not decide every distribution, contributory-infringement or proposed class issue.

Summary judgment on training-copying claims; distribution and contributory claims remain

• On this record, the court found transformative training use and insufficient evidence of relevant market harm from the 13 authors, granting fair-use summary judgment on training-copying claims.
• ECF700 of 25 March 2026 allowed amendments to contributory and distribution claims and the proposed class definition, while denying class discovery at that time.

AI & Data Compliance

Map data sources, processing roles, purposes, flows and controls across AI systems.

Generated Content & IP

Assess training materials, inputs, outputs, ownership, licensing and infringement response.

AI Contracts & Disputes

Support procurement, development, licensing, investment and disputes involving technology, data and outputs.