Training data and copyright: what is not settled

Imagine a chef who learned to cook by photographing every page of every rival's recipe book and keeping the photographs in a warehouse. Nobody claims the finished dish is a copy of any single recipe. The argument is about the warehouse: the copying that had to happen for the education to happen at all. That is the shape of the copyright fight over generative models, and it is why the debate so often talks past itself. There are two separate questions. First, was the reproduction of works to assemble and train on a dataset infringement, or was it fair use in the United States and fair dealing elsewhere. Second, and independently, does a particular output infringe, either by resembling a specific work or by depicting a protected character or trade mark. A model can be clean on one question and exposed on the other. The honest answer today is that neither question is resolved. The US Copyright Office published Part 3 of its Copyright and Artificial Intelligence report, on generative AI training, in May 2025, and it remains a pre-publication version. The litigation is live and pulling in different directions on different facts: Andersen v. Stability AI is in discovery with Runway added as a defendant, Getty's US case moved districts, Kadrey v. Meta continues, the New York Times matter is consolidated in a multidistrict proceeding, Bartz v. Anthropic ended in a settlement given final approval in July 2026 so its fair-use reasoning was never tested on appeal, and Thomson Reuters v. Ross was argued in the Third Circuit in June 2026 and is undecided. In the UK, the Getty v. Stability judgment turned on territoriality and abandoned claims rather than on whether training is lawful. Anyone who tells you the question is settled is selling something.

What people get wrong

Each of these is a belief worth dropping, followed by what is actually happening.

• a court has already ruled that training on copyrighted work is fair use. No US appellate court has ruled on training fair use. Thomson Reuters v. Ross was argued in the Third Circuit in June 2026 and remains undecided, and the most-discussed district-court analysis, in Bartz v. Anthropic, ended in a settlement approved in July 2026, so it was never appealed and binds nobody.
• Getty lost in the UK, so training is lawful there. Getty abandoned its training claim mid-trial for want of evidence that training took place in the UK, and the court made no finding on how many works were used. That is a narrow judgment about territoriality and evidence, not a ruling that training is lawful.
• if the output does not resemble any single training image, there is nothing to argue about. Output similarity and the copying performed to build the training set are separate questions, and both are being litigated.
• open weights means the training data was licensed. A licence on the weights says nothing about the provenance of the data used to produce them.
• a model trained on licensed data removes the risk. It narrows one risk. Indemnities carry caps and conditions, and prompting for a protected character or a trade mark can still produce an output problem.

Terms this introduces

Fair use

A US statutory defence that can excuse otherwise infringing copying based on a multi-factor analysis, applied case by case rather than as a blanket rule.

Fair dealing

The narrower UK and Commonwealth equivalent, available only for specified purposes rather than as an open-ended balancing test.

Provenance record

The log of model, model version, prompt, seed, parameters and input references behind a generated shot, kept so a delivery can be explained later.

Indemnification

A vendor's contractual promise to cover defined legal costs arising from use of its outputs, usually subject to caps, conditions and correct use of the product.

Multidistrict litigation

A US procedure that consolidates related federal cases before one judge for pretrial purposes, which is why several separate complaints can appear under a single docket.