Illustration of a blindfolded justice figure with scales facing an AI head silhouette, with a copyright symbol between them — Illustration: LawSites (LawNext), from its coverage of Thomson Reuters v. ROSS Intelligence.

AI copyright law just got its first answer from a US federal appeals court on training data. The Third Circuit ruled that ROSS Intelligence broke copyright law when it used Westlaw headnotes to train a competing legal research tool.

I coordinate AI and software projects for clients in India and the USA. Almost every project now touches someone else's data. This ruling gives us a clearer picture of where the lines sit, and it should change how agencies source training data.

Below I explain what the court decided, what it left open, and the steps I think every agency should take this quarter.

The AI Copyright Case in Plain Words

Thomson Reuters owns Westlaw, one of the biggest legal research platforms. Westlaw editors write headnotes. These are short summaries of the key legal points in a court opinion.

ROSS Intelligence was a startup founded by three University of Toronto computer science students. It built an AI search engine that answered plain-language legal questions with passages from about ten million public court opinions.

To train that engine, ROSS hired a company called LegalEase Solutions. LegalEase wrote about 25,000 training memos, each pairing a legal question with four to six opinion passages. According to the court, its writers used thousands of Westlaw headnotes to frame those questions.

Thomson Reuters sued in 2020. Courthouse News reports that LegalEase was used to get access after Westlaw refused ROSS access. ROSS has since shut down.

The timeline so far

2020: Thomson Reuters files suit in the US District Court for Delaware.
February 2025: Judge Stephanos Bibas grants partial summary judgment for Thomson Reuters on 2,243 headnotes.
June 11, 2026: A Third Circuit panel hears the appeal.
September 29, 2026: The panel affirms. The full opinion is now public, and law firms are publishing their readings of it this week.

What the Third Circuit Decided

Judge Tamika Montgomery-Reeves wrote for a unanimous panel that also included Judges L. Felipe Restrepo and Emil J. Bove III. You can read the full Third Circuit opinion on the court's website.

The court framed it simply. In its words, "this is no more than an ordinary copyright case." It answered two questions, and Thomson Reuters won both.

1. Headnotes are protected by copyright

Copyright only needs a small "creative spark." The court found that editors make creative choices about which legal points to include and how to word them. That was enough.

The court also said judicial opinions stay free for anyone to use. Protecting the headnotes does not give anyone a monopoly over the law.

2. Training on them was not fair use

US fair use law weighs four factors. Here is how the court scored them:

Purpose of the use: Against ROSS. Its use was commercial and "minimally transformative, at best," because both companies used the headnotes to build legal research tools.
Nature of the work: Slightly for ROSS, since headnotes are published and fairly factual.
Amount taken: Against ROSS. It argued it took only 0.08% of Westlaw's 28 million headnotes. The court treated each headnote as its own work, so each copy was a full copy.
Market harm: Against ROSS. The court found harm to Westlaw's market and to a "rapidly developing" market for licensing headnotes as AI training data.

One line from the opinion stands out: "Unlike necessity, ease is not a justification for copying." ROSS could have worked from the free court opinions. It used headnotes because they were easier.

Close-up render of a computer chip labeled AI on a dense circuit board — A conception of artificial intelligence. Photo: Igor Omilaev/Unsplash via Courthouse News.

What This AI Copyright Ruling Does Not Settle

This is not the final word on generative AI. The court went out of its way to say so.

ROSS was not generative AI. The opinion says its system "would not create any new expression." It only returned passages from existing opinions. The court said concerns raised in the pending OpenAI copyright litigation "do not apply here."
It was a direct competitor. ROSS priced its product "in line with" Westlaw and hoped to replace it. A general-purpose model has a weaker overlap with any single source.
Output was not at issue. The court did not decide whether AI-generated output can infringe.
Bad facts hurt ROSS. In a footnote, the court said ROSS at times acted in bad faith. One example was trying to access Westlaw with law-firm investor credentials despite notice that Westlaw's terms prohibited it.

The fight is also not over. Yar Chaikovsky of White & Case, who represents ROSS, told Courthouse News that ROSS will seek Supreme Court review. A Thomson Reuters spokesperson said the company was pleased that the court agreed Westlaw's editorial content is protected.

The ruling binds courts only within the Third Circuit. Still, other courts will read it. Fox Rothschild's Melissa E. Scott noted that plaintiffs suing Anthropic have already filed it as supplemental authority.

Why Agencies in India and the USA Should Pay Attention

Most agencies are not training giant foundation models. We build smaller things for clients. Think search tools, classifiers, recommendation engines, chatbots and fine-tuned models.

That kind of work looks a lot more like ROSS than like a general-purpose chatbot. A narrow tool trained on a rival's content, for a similar purpose, sits close to the facts of this case.

The vendor angle is the part I keep thinking about

ROSS did not copy the headnotes itself. LegalEase and its subcontractor, Morae Global, did. For the appeal, ROSS did not dispute that their actions were attributable to it.

That matters for us. Many Indian agencies do data labeling, annotation and training-set work for US clients. If the people writing your training data copy from a protected source, your client may carry the risk. So may you, depending on the contract.

A licensing market now counts

Thomson Reuters had never licensed headnotes to anyone for AI training. The court said that "does not disprove that a market exists."

In practice, that means a data owner can argue harm even before it sells training licenses. Any content with value as training data, from product catalogs to course material, can now point to this ruling.

A quick note for Indian teams

US fair use is not the same as Indian law. India's Copyright Act, 1957 uses fair dealing under Section 52, which lists specific allowed purposes. But if your client is in the USA, US law will usually shape the risk, so plan for it.

An AI Copyright Checklist for Agency Projects

Here is the checklist I am adding to our AI project kickoffs. None of this is legal advice. It is the process side that helps your lawyers do their job.

Map every data source. List where each training, fine-tuning and retrieval dataset came from. Note the license or terms for each.
Ask the "could we use something else?" question. The court cared that ROSS had free alternatives. Write down why you need a specific dataset, not just that it was convenient.
Check for competitor overlap. If the data comes from a product your client's tool will compete with, stop and get legal review.
Put rules in vendor contracts. Annotation and data vendors should confirm they will not copy from paid or restricted sources. Ask for written source notes.
Respect terms of service. Never use shared, borrowed or investor credentials to reach a paid database. That detail hurt ROSS badly.
Prefer licensed or first-party data. Licensing remains the most defensible path when your tool and the source serve the same purpose.
Keep a provenance log. Store dataset versions, dates and approvals. It helps in audits and in disputes.

Where generative AI projects fit

If you build on top of large language models, the open questions still matter. The US Copyright Office's AI reports take a case-by-case view of training and fair use. Courts are still working through the big generative AI cases.

So keep your own fine-tuning data and retrieval content clean, even when the base model is someone else's. Your client controls those layers, and so do you.

If you are also thinking about how AI companies collect web data, our post on AI crawlers and Wikipedia's bot findings covers the other side of the same problem.

How Debugged Pro Handles Training Data

On our AI development projects, data sourcing now sits in the discovery phase, not at the end. We agree on sources, licenses and vendor rules before anyone writes a line of training code.

It adds a few days at the start. In return, clients get a model they can actually defend, sell and scale. If you want a second opinion on an existing setup, our software development consulting team can review your data pipeline and contracts with your legal counsel.

Key Takeaways

First appellate ruling: The Third Circuit held that ROSS's AI training on 2,243 Westlaw headnotes was not fair use.
Convenience is not a defense. Using a source because it is easier, when free options exist, weighs against you.
Emerging licensing markets count, even if the owner has not licensed its content yet.
Generative AI is still open. ROSS's tool did not generate new text, and ROSS plans to seek Supreme Court review.
Agencies should map data sources, tighten vendor contracts and keep a provenance log on every AI project.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He manages development, QA and data teams delivering web platforms, ecommerce systems and AI integrations for clients across India and the USA, with a focus on clean processes for data sourcing and delivery.