The first US appeals ruling on AI training went against the AI company. European courts are further along.

The US Court of Appeals for the Third Circuit has upheld the verdict against Ross Intelligence, a legal research start-up that used headnotes from Thomson Reuters’ Westlaw to train its AI search tool. The court agreed that Ross could not rely on fair use. It is the first US appeals ruling on fair use in AI training, and every AI developer being sued in the US will read it closely once it is public. For now, the public version of the judgment is one word: affirmed. The court’s reasoning is under seal while the parties agree on redactions.

What Ross did and why it lost

Westlaw headnotes are short summaries of the legal points in court decisions, written by Thomson Reuters editors and available only to subscribers. Ross wanted to build a competing legal search engine and used thousands of headnotes to teach its system which court opinions answer which questions. Thomson Reuters sued in 2020. Ross shut down its product in 2021, citing the cost of the litigation, and has been fighting a case about a company that no longer operates ever since.

In February 2025 Judge Stephanos Bibas ruled that the use was not fair use. His reasoning was simple: Ross copied the headnotes to build a product that competed with the one they came from, so the use did not transform anything, and it harmed Thomson Reuters’ market. The Third Circuit heard the appeal on June 11 and has now affirmed.

Narrower than the headlines

The outcome sounds bad for OpenAI, Anthropic, Meta and the other companies facing dozens of copyright suits in the US. But on the facts, Ross is a weak model for those cases, as a review in the American Bar Association’s Business Law Today set out before the appeal was decided.

Ross built a direct substitute for the product it copied from. A large language model trained on news articles does not replace a newspaper in the same obvious way. Headnotes sat behind a paywall, which showed Thomson Reuters meant to control their use, while most generative models were trained on material that was freely available online. And Ross’s system was not generative. It found existing court opinions. It did not write new text.

So what matters is why the Third Circuit affirmed, and that is exactly what we cannot read yet. Copyright Lately lists the questions generative AI defendants care about: whether a merely potential market for licensing training data is enough to count as harm, and whether copying as an intermediate step is excused only for computer code. If the court answered broadly, the pending cases get harder for the AI companies. If it stuck to the facts of a direct competitor, they barely move.

My read: even the narrow version helps rights holders. Every AI company that builds a product competing with its data source, which is a fair description of AI legal research, AI news summaries and AI stock photography, now has an appeals court ruling against it.

In Europe the question is being fought on different ground

European copyright has no fair use. It has specific exceptions, and the one that matters here is the text and data mining exception in the 2019 copyright directive, which lets anyone mine lawfully accessible content unless the rights holder has reserved the right in a machine-readable way. The European fights are about whether that exception covers training, and what happens when a model can reproduce what it was trained on.

The clearest answer so far came from Munich. In November 2025 the Regional Court ruled against OpenAI in a case brought by GEMA, the German music rights society, over song lyrics. The court held that lyrics memorised in a model’s parameters are copies, that outputs reproducing them infringe, and that the mining exception does not cover this. It granted an injunction and damages. OpenAI has appealed.

Denmark has its own case. In February the Danish press publications’ collective management organisation DPCMO, which represents 99% of the Danish news industry, took OpenAI to court. It says OpenAI trained on Danish news until at least the summer of 2024, while publishers had no way to opt out before the summer of 2023, and that ChatGPT reproduces Danish articles. The suit followed a failed mediation that OpenAI declined to take part in, which the culture minister publicly regretted.

On top of the courts sits the AI Act. Providers of general-purpose models must have a policy for respecting copyright reservations and publish a summary of their training data, and since 2 August the AI Office can fine them up to 3% of global turnover or €15 million, whichever is higher, for not doing so. The code of practice that most large labs signed commits them to crawlers that honour robots.txt and similar opt-out signals.

What it means if you use or own data

If you only use AI tools, the direct legal risk still sits mostly with the providers. The exception is output. The Munich ruling found infringement in what the model produced, and if your marketing team publishes a ChatGPT text that reproduces someone’s article, you published it. Check whether your AI contracts include a copyright indemnity, and read what it excludes.

If you fine-tune or train models yourself, stop assuming that anything reachable online is fair game. Keep a record of where training data came from and whether the source reserved its rights. The AI Act expects providers to be able to show this, and the Ross ruling suggests US courts will not treat careless copying kindly either.

If you own valuable content, such as a database, a product catalogue, specialist articles or support documentation, your position just got stronger. Put a machine-readable reservation on it (robots.txt and your terms, at minimum) if you have not already, because under EU law the mining exception applies unless you have opted out. Then decide whether you would rather license it. Thomson Reuters spent six years in court to prove that its summaries were worth paying for. A company that sets its terms now will not need to.