Mathew PaquetChecklist · 4 min read
The AI Training Data Shortcut Check
Run the four questions from the first appeals ruling on AI training against every source you train or fine-tune on, before the training run.
On September 29, 2026, a federal appeals court ruled that ROSS Intelligence's use of Westlaw's headnotes was not fair use. Fair use is the part of copyright law that allows some copying without permission. ROSS's AI never wrote anything: every answer was a passage from a public court opinion. It lost over how its training data got made.
The court gave two reasons that travel to any business: ease is not necessity, and a market the owner never sold can still count. The next three sections turn them into four questions.
Narrow ruling. This was not a chatbot that writes new text, and the court called it no more than an ordinary copyright case. ROSS says it intends to ask the Supreme Court to review it. This is not legal advice.
Question one. Is it the raw material, or someone's layer on top of it? Raw material is the source itself. A layer is the work someone added: choosing, summarising, tagging, rewriting.
Where it came from: [OWN / LICENSED / BOUGHT / SCRAPED / VENDOR]
Raw material or someone's layer: [RAW / LAYER]
Whose layer: [WHO ADDED THE WORK]
Run these two on every source you marked as someone's layer.
Worked example. You want a support bot that answers like the best help center in your industry. Writing your own answers from your product manuals is slow. Copying a competitor's help articles is fast, and it is the path this court looked at.
Question four. Could they sell it for training, and who actually built your dataset? ROSS argued there was no market, because Thomson Reuters had never licensed headnotes to anyone. The court said the market for licensing headnotes to train AI is rapidly developing, and Thomson Reuters trains its own AI search on them. If the owner licenses content for AI, or plausibly could, that is a market.
ROSS did not write its memos: a contractor did, and on appeal ROSS did not dispute that the contractor's work counted as its own. If a vendor builds your data, ask in writing.
Before we train on the dataset you're building for [PROJECT], could you confirm in writing:
1. Where did each part of the data come from?
2. What license covers training an AI model on it?
3. Will you stand behind that answer if someone asks?
Thanks,
[YOUR NAME]
Do it before the run. ROSS's platform was gone by January 2021, and the case kept going for more than five years after that. Ask these questions before the training run, not after the letter.
Sources
Mathew's LibraryLibrary card
Keep reading free with your library card
One email opens this checklist and all 3 resources in Mathew's library. No account, no password.
List every source your model or custom GPT learns from, and mark each one raw material or someone's layer.
Saving works best in your phone's browser. A copy is also in your inbox.