A federal appeals court just handed founders a warning about the data behind their AI products. The Third Circuit affirmed Thomson Reuters’ copyright win over Ross Intelligence and turned down the startup’s fair use defense. It is a big moment for anyone tracking the AI copyright fight.
Too many young founders treat training data like free fuel. Do not make that mistake. This ruling is a reminder to pair your product plans with real intellectual property protection and a clear view of what you borrowed.
What the Third Circuit Decided
The case started when Thomson Reuters sued in 2020. It said Ross copied Westlaw headnotes, which are short summaries of legal points in court opinions, to train a competing AI legal search tool. A federal judge in Delaware sided with Thomson Reuters in 2025, and the appeals court has now agreed.
Ross closed its product in 2021 because the legal fight drained its resources. That detail matters. The company lost its product before it ever got a final answer, and the bill for that fight fell on a small team.
Why This Case Is Narrower Than It Sounds
Before you panic, read the fine print. Reports say the dispute concerns a non-generative system, and the court’s full reasoning is under seal for about ten days. Because of that, lawyers will not know how far the logic stretches until the opinion is public.
Still, it is reported to be the first time a US appeals court has weighed in on AI training and copyright. Even a limited ruling shapes how investors, buyers and partners look at your data. So the smart move is to prepare now instead of waiting for a broader decision.
Audit the Data Behind Your Product
Here is the fast version of a data audit. Start small and work through it in one afternoon.
- List every dataset, document and model your product touches.
- Note where each one came from and what license or contract covers it.
- Flag anything scraped, copied or bought without clear permission.
- Ask your AI vendors what they promise about their own training data.
Then write it down. A written record, plus a basic AI usage policy for your staff, shows a buyer or investor that you took the question seriously. That kind of proof can save a deal during due diligence.
Licensing Beats Litigation for Small Companies
You will not win a copyright fight against a large publisher on a startup budget. Ross could not. Licensing content up front is usually cheaper than defending a lawsuit, and it gives you cleaner books when you raise money.
The market is moving that way already. Google has piloted direct payments to publishers, which we covered in AI Overviews payments. When big platforms start paying for content, smaller companies will be expected to do the same. For official guidance on where the law stands, the US Copyright Office AI page is a good place to start.
Rulings and Filings to Track Next
Watch for the unsealed opinion first. It will show whether the panel limited its reasoning to non-generative tools or left the door open for broader use. Then watch the generative AI cases, because those are the ones that could hit chatbots and image tools directly.
Keep your legal counsel in the loop, and revisit your data audit every quarter. The founders who stay ahead of this issue will move faster, because they will not have to stop and clean up later.
Contracts Your Team Should Review This Month
Your vendor agreements deserve a fresh look. Many AI providers include terms about who is responsible if training data causes a legal problem. Read those clauses, and ask whether the vendor will stand behind its outputs if a rights holder complains.
Do the same with freelancers and agencies that supply content or data. Add a line requiring them to confirm they have the right to share what they deliver. It takes five minutes to add and can spare you a painful dispute later.
Finally, tell your team what is off limits. A simple rule, such as never pasting a paid database into a public tool, prevents the most common mistake before it starts.
One last habit is worth building. Assign a single person to read AI legal news each week and summarize what changed for your product. Ten minutes of attention now keeps surprises from turning into emergencies, and it signals to partners that you run a careful company.
Is AI training on copyrighted material always illegal?
No. This ruling rejected one company’s fair use defense in a specific case, and the reasoning is not fully public yet. Other cases could come out differently.
What should a startup do first?
Inventory your data sources and document the license for each. That single step puts you ahead of most early-stage teams.
