The First Appellate AI Copyright Ruling Arrived Under Seal.

The result is public. The rule is not yet public.

Reuters reported that the Third Circuit affirmed the Delaware district court's ruling for Thomson Reuters and rejected ROSS's claim that its use of Westlaw material was fair use. Reuters called it the first U.S. appellate ruling to address copyright in AI training. That makes the holding important. It does not make every description of it accurate.

The opinion itself is not currently available to the public. The court said it had filed the opinion temporarily under seal, directed the parties to submit proposed redactions and a reason for each within ten days, and stated that it would then decide what to accept before unsealing the opinion in an appropriate form. That is not a declaration of permanent secrecy. It is a redaction process. The distinction matters because confidential facts can justify short-term protection while the legal analysis still belongs in public view.

Until that public version appears, we know the disposition but not the appellate court's path. We do not know which parts of the district court's reasoning the panel adopted, narrowed, or declined to reach. We should resist filling the gap with the most dramatic available interpretation. An affirmance can rest on a precise factual record without converting every sentence below into circuit law.

This was not a chatbot learning from the internet

The district-court record is unusually specific. Thomson Reuters owns Westlaw, including headnotes written by lawyer-editors to distill points of law from judicial opinions. Judicial opinions themselves are not protected by Thomson Reuters's copyright. The added editorial layer can be. ROSS wanted a question-and-answer dataset for an AI legal-search tool. After Thomson Reuters declined to license Westlaw content to a competitor, ROSS obtained about 25,000 so-called Bulk Memos from LegalEase. The district court found that thousands of the memo questions had been built from Westlaw headnotes and granted summary judgment on actual copying and substantial similarity for 2,243 headnotes, subject to a remaining copyright-validity issue for some works.

ROSS's system was not generative AI. A user entered a legal question and the system returned relevant judicial opinions. The headnotes did not appear in the user-facing result. ROSS therefore argued that any copying was intermediate: the material helped train relationships among legal words, while the finished product delivered public law rather than Westlaw text.

That is an important argument, but not a universal description of AI training. This case did not involve a large language model producing prose, a general-purpose model trained across the open web, or a developer that happened to ingest a protected work among billions of files. It involved a direct competitor, a refused license, a purpose-built training set, and editorial annotations closely tied to the function of the competing product. Anyone calling the affirmance a complete answer to generative-AI copyright litigation is claiming more than the public record supports.

The district court drew the line at substitution

Section 107 requires courts to weigh four fair-use factors: the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality used, and the effect on the potential market. It does not contain an AI exception or an AI prohibition. The district court ruled for ROSS on the second and third factors because the headnotes were not highly creative and did not appear in the final output. It ruled for Thomson Reuters on the first and fourth, which it considered more important.

The court viewed ROSS's purpose as commercial and nontransformative because the headnotes made it easier to develop a legal-research product that would compete with Westlaw. It distinguished software reverse-engineering cases where intermediate copying was needed to reach otherwise inaccessible functional elements. ROSS could have built its own legal questions from the public opinions, the court reasoned, without copying Westlaw's expression.

Market effect carried the most weight. The court considered both the market for legal-research platforms and a potential derivative market for legal-AI training data. It did not require proof that Thomson Reuters was already licensing the exact dataset at issue. ROSS bore the burden on fair use and, in the court's view, had not shown that the relevant markets did not exist or would not be affected.

That analysis presents the real appellate question. If a company copies protected editorial work at an intermediate stage to build a substitute for the owner's product, how much transformation is enough? When is a claimed market for training data a cognizable derivative market, and when would recognizing it let a copyright owner define fair use out of existence? The sealed opinion may answer those questions narrowly or broadly. The difference is worth waiting to read.

A public court should not create a private compliance rule

The Third Circuit has good doctrine for handling this problem. In its precedential Avandia decision, it said that the common-law analysis begins with a strong presumption of access to judicial records. A party seeking continued sealing must identify the kind of information courts protect and a clearly defined, serious injury that disclosure would cause. Courts must make a particularized assessment rather than use broad categories or routine confidentiality as a substitute for findings.

That does not mean every training example, contract term, source-code description, or sealed exhibit must appear in the public opinion. Redaction can protect actual trade secrets. The court's order appropriately asks for a reason for each proposed redaction and reserves the final decision to the judges. The public-interest question is whether the released opinion will contain enough facts and reasoning for lawyers, developers, copyright owners, and lower courts to understand the rule.

The legal system asks AI vendors to explain provenance, limits, and decision pathways. Courts should impose a comparable discipline on themselves when their decisions govern those systems. A bare affirmance would leave responsible actors guessing while advantaging the parties and insiders who saw the sealed text. That is the opposite of what appellate law is for.

The practical lesson starts with provenance, not slogans

Legal-technology teams should not wait for a slogan such as 'training is fair use' or 'AI training is infringement.' Neither is a workable diligence standard. Start with the actual inputs. Judicial opinions, statutes, agency materials, briefs, treatises, headnotes, citators, annotations, docket metadata, and licensed databases do not have the same copyright status or contractual restrictions merely because they all look like legal text.

Then map what the system does with each source. A tool that searches public opinions, one that reproduces protected annotations, one that learns from a licensed corpus, and one that uses a rival's editorial layer to build a substitute present different facts. Preserve source records, license terms, transformation steps, evaluation datasets, and deletion controls. Indemnity language cannot repair provenance that nobody documented.

For law firms buying AI tools, the useful vendor question is not simply whether a model was trained on copyrighted material. Ask which materials, under what authority, for what product function, and with what risk allocation. That is the same distinction this case forces into view: public law is available to everyone, but the private editorial work layered on top of it may not be.

The Third Circuit may have reached the right result. We should be able to evaluate why. The court has promised an appropriately unsealed version after a targeted redaction process. In the first appellate decision of its kind, appropriate should mean genuinely usable. A landmark without a public boundary is not guidance. It is only news.

The law does not become clearer because a court names the winner. It becomes clearer when the court explains the boundary.