AI Fair Use Rejected on Appeal as Chatbot Liability Rulings Mount

A summer that ended with an appellate answer

For three years, businesses building or buying AI have relied on trial-court opinions and settlements to judge their legal exposure. That changed on September 29, 2026. That day, the U.S. Court of Appeals for the Third Circuit upheld Thomson Reuters' copyright win against ROSS Intelligence. It became the first federal appeals court to decide whether training an AI system on copyrighted material can be fair use.2325 Its answer, on the facts before it, was no.25

The ruling came after several months in which courts in the U.S. and Europe also weighed in on who answers for what AI systems say. The courts did not agree on everything. Read together, though, the decisions point the same way. The legal shelter around AI is getting narrower, and the risk is moving toward whoever trains on proprietary content or puts a chatbot in front of customers. Law firms are adopting these tools faster than they are writing the policies to govern them.

What the Third Circuit decided, and what it did not

The panel affirmed Judge Stephanos Bibas's February 2025 ruling on both main questions. Westlaw's headnotes are original enough to be protected by copyright, and ROSS's use of them to train a competing legal search tool was not fair use.25 Judge Montgomery-Reeves wrote the opinion. She said the materials had a "creative spark" and that ROSS wanted to be a direct competitor using them for a very similar purpose.26 The court found that three of the four fair-use factors favored Thomson Reuters, including the purpose of the use and the harm to its market.27

For rights holders, the most useful part may be the court's view of market harm. The panel found harm in two places: the existing legal-research market, and a developing market for licensing headnotes as AI training data.24 One commentator says that second theory reaches well beyond legal research.28 The music industry clearly agrees. The RIAA and NMPA filed a joint brief calling the case the first chance for an appeals court to rule on competing commercial models. To show a licensing market already exists, they pointed to deals the major labels have signed with AI firms such as Stability AI, Udio and ElevenLabs.21

The coverage splits on how far the ruling reaches. Music-industry commentators treat it as a key precedent against training that substitutes for the original work.21 Legal analysts read it more narrowly. ROSS's product retrieved existing judicial text and did not generate new expression, and the court expressly set it apart from generative systems. That leaves open whether training large language models on books or other creative works is fair use.24 One newsletter put it simply: building a direct competitor from someone else's content is not fair use, but training a general-purpose model probably still is.27 Another account notes that the opinion discusses the Justice Department's September 1 filing, which called LLM training "extraordinarily transformative." It adds that the decision does not bind the Ninth Circuit, where most generative-AI cases are being heard.22

The narrower reading is the better one. The opinion is built on competing products and clear substitution, and those facts do not describe OpenAI or Meta. Even so, one change applies to everyone. Developers can no longer argue that training is automatically fair use just because AI is transformative. Courts will look at what the finished product does.27

The training-data map after Bartz and Kadrey

The ruling joins a mixed set of decisions. In 2025, Judge Alsup found in Bartz v. Anthropic that training on books was fair use but keeping a central library of pirated copies was not. The company then agreed to a $1.5 billion class settlement, which Reuters called the largest known U.S. copyright settlement.25 Accounts differ on when final approval came. One tracker says July 2026. Another says a hearing on May 14 and payouts of $3,100 per work, with about 91% of eligible works claimed.2529 In Kadrey v. Meta, the court ruled for Meta on fair use because the plaintiffs had not proved market harm, and it stressed that the holding was narrow.25 Distribution claims tied to Meta's alleged torrent seeding are still alive.30

Abroad, results have been split. Getty Images lost its secondary-infringement claim against Stability AI in the English High Court. A Munich court found that OpenAI infringed GEMA's rights through memorized song lyrics.25 The U.S. government has also filed a statement of interest supporting fair use in the OpenAI multidistrict litigation in New York.23

The practical rule companies are drawing from all this is about where their data came from. One compliance guide puts it this way: how a vendor got its training data is now a supply-chain risk, and deployers should ask what their models were trained on.29 It also warns that courts in the OpenAI litigation have ordered millions of output logs produced in discovery. That makes audit trails a matter of evidence, not just good practice.29 Each new licensing deal makes the fair-use defense harder for the next defendant, because it helps prove a market exists.30 The Third Circuit's two-market analysis turns that point into appellate law, at least within its own circuit.

Liability: the chatbot is you

While the copyright fights turn on training, a second set of rulings over the summer dealt with outputs. The clearest came from Germany. On May 12, the Higher Regional Court of Hamm held that a medical practice was responsible for its website chatbot giving its doctors specialist titles that do not exist. The court rejected the argument that the chatbot was a third party.6 It said the operator bears the risk of hallucinations even when the bot is carefully configured, and it allowed an appeal to the Federal Court of Justice because the question was new.8 About two weeks later, a Munich court issued a preliminary injunction against Google over false AI Overviews about two publishers. It treated Google as a direct infringer because the summaries made new statements of their own.8

DLA Piper links the Hamm decision to Canada's 2024 Air Canada ruling. It sees an emerging pattern in which courts hold companies effectively strictly liable for what their chatbots say, and in which blaming a vendor may not work.6 The firm also notes that neither case came from a top court and that the German case is still on appeal.6 That caution is fair. Still, judges in two systems have reached the same answer, and businesses should treat it as the working rule.

U.S. courts are moving more slowly, and mostly at the early stages of cases. On July 24, a Delaware court refused to dismiss Robby Starbuck's defamation suit over output from Google's AI. That opens the way to discovery but makes no finding of liability.7 The template still comes from the Character.AI case. In May 2025, a federal judge ruled that its app could be treated as a product for liability purposes. The case and four related suits settled in January 2026 on undisclosed terms.15 Since then the pressure has grown. Florida sued OpenAI in June, and on September 8 a federal judge sent that case back to state court. California passed SB 1119, a set of safety rules for companion chatbots. More than 20 suits against OpenAI are pending.9 In July, a Florida pastor sued over ChatGPT health advice, a complaint that combines product-liability and unauthorized-practice-of-medicine claims. It remains an untested filing, not a precedent.4

Legal tech's adoption-governance gap

The legal profession is caught in both trends. It buys AI and also argues these cases. Usage is growing quickly, though the surveys disagree on how quickly. Thomson Reuters reports that 41% of law firms use generative AI, up from 28%.12 ILTA found 94% of 508 firms using or exploring it, and says this is the first time an emerging technology has been the main focus of firm technology spending.13 8am found that individual use of general-purpose tools rose from 31% to 69%.14 Spellbook argues that the figures depend heavily on who was asked.17

Governance is falling behind. According to 8am, 43% of firms have no AI policy and no plans to write one, only 9% enforce a written policy, and 54% offer no training.14 ILTA found that only 6% of firms have the AI in their own document systems switched on and in wide use, meaning purchases are outrunning deployment.13 Lawyers' views are also changing. Law360's survey found that frequent users at firms have become more neutral about AI than they were a year ago.16

The rulings make that gap harder to excuse. If a chatbot's statements count as the company's own, and if the source of training data decides copyright exposure, then a firm with no AI policy carries real risk. That matters most when lawyers rely on consumer tools, which the North Carolina Bar flags as a confidentiality concern.11 It is also worth noting that the company that won the first appellate fair-use case sells one of the legal AI products these firms are buying.

The bottom line

The Ninth Circuit has not yet ruled on generative-AI training, and the German Federal Court of Justice has not ruled on chatbot attribution. Until they do, companies should work from two rules this summer's decisions support. Training on a rival's licensed content to build a competing product is a losing position. Whatever an AI system tells customers will probably be treated as the company's own statement.