The First Appellate AI Training Ruling Is About Headnotes. Your Vendor Contracts Should Care Anyway.

The first federal appeals court ruling on fair use and AI training landed on September 30. It isn't about chatbots, image generators, or a frontier lab. It's about Westlaw headnotes.

LawSites reported that the Third Circuit ruled for Thomson Reuters in its long-running copyright case against ROSS Intelligence, the now-shuttered legal research startup. IPWatchdog reported the same day that the court affirmed the District of Delaware's rulings that ROSS infringed copyrights in more than 2,000 Westlaw headnotes and that its fair use defense failed. One detail matters a lot here: the opinion was still sealed pending redactions when those reports came out. Much of what we know comes from the parties and early coverage, not the full text.

On October 2, LawSites reported that ROSS plans to ask the Supreme Court to review the ruling. The fight isn't over. But in-house teams don't have to wait for a cert petition to learn something useful.

Why a headnote case should be on a GC's radar

You might say you don't train AI models and you don't sell legal research. That's fair. But your department now buys AI, feeds it your documents, and increasingly builds small tools on top of it. All three put you near the questions this case raised. MLex reported that the court treated the case largely as an ordinary copyright dispute about one competitor copying another to build a substitute product. That framing may keep the ruling narrow. A narrow ruling still gives you a useful checklist.

Five lessons for the legal department

1. Ask vendors where their training data came from (and get the answer in writing)

"Proprietary" isn't an answer. If a vendor's model or retrieval layer was built on licensed content, they should say so. If it was built on scraped or "publicly available" material, you want to know that before procurement signs. Put a data-provenance question in every AI vendor questionnaire. You can find tools to compare in the generative AI legal assistants directory.

2. Read the IP indemnity like you'll actually need it

Many AI contracts include an IP indemnity that sounds generous and turns out to have a lot of exceptions. Check whether it covers claims about the model's training data, not just the outputs. Check whether it's capped at fees paid. Check whether "modifications" or "customer prompts" void it. A ruling like this one is a good reason to send the indemnity section back for another pass.

3. Your playbooks are training data too

Your clause library, negotiation playbooks, and annotated precedents are carefully edited work product. That's the kind of material at the center of this case. Make sure every vendor agreement says plainly whether your data can be used to train or improve models other customers use. "We may use aggregated, de-identified data to improve our services" should prompt a follow-up question, not a nod. The contract lifecycle management and drafting and clause management categories are good places to compare how vendors handle this.

4. Watch your own builders

Legal ops teams are wiring up workflows and copying third-party content into knowledge bases faster than ever. That's mostly great. But "we pasted a research service's summaries into our internal bot" is exactly the kind of fact pattern you don't want discovered later. Licensed content usually comes with license terms. Read them before anyone builds on top of them.

5. Don't over-read it either

Coverage has repeatedly stressed that ROSS was a non-generative system, that the copying happened before generative AI, and that ROSS was building a direct competitor to Westlaw. Those facts may limit how far the ruling reaches into the generative AI cases. Don't rewrite your AI policy on the strength of one sealed opinion. Do flag it for your next AI governance update.

What to do this week

• Add two questions to your AI vendor intake: "What content was used to train or ground this product, and under what license?" and "Is our data used to train models other customers use?"
• Pull your top three AI contracts and check whether the IP indemnity covers training-data claims.
• Remind internal builders that licensed research content comes with license terms.
• Set a calendar reminder for the unsealed opinion and any cert petition. The reasoning will matter more than the headline.

The first appellate AI training ruling ended up being about headnotes, not chatbots. Your vendor contracts should still assume a bigger ruling is coming.

This article is general information, not legal advice.