Every creator worried about AI eventually asks the same question: is it legal for a company to copy my work to train its software? For years, lawyers answered with a shrug and one word — maybe. It all comes down to fair use, and no US court had actually ruled on fair use in an AI case.
Then one did. On February 11, 2025, a federal judge in Delaware handed down the first real fair-use decision of the AI era in Thomson Reuters v. Ross Intelligence. The AI company lost. Cue the headlines: “Court says AI training isn’t fair use.” Except that’s not quite what happened — and the gap between the headline and the ruling is exactly where your rights live. Here’s the plain-English version.
The 30-second version
Thomson Reuters owns Westlaw, the giant legal-research platform lawyers use to find cases. A startup called Ross Intelligence wanted to build a smarter, AI-powered rival. To train its tool, Ross needed a huge set of legal questions and answers. It asked Westlaw for a license. Westlaw — understandably not keen to arm a competitor — said no.
Track the cases yourself
Every current AI-copyright ruling — and the ones still being fought — lives in our AI Copyright Ruling Tracker. Filter by your creator type to see, in plain language, how each case affects your rights.
So Ross went around the front door. It paid a third-party company to produce training material, and that material was built from Westlaw’s headnotes — the short editorial summaries Westlaw writes for each case — and its Key Number System, Westlaw’s proprietary way of organizing the law. Thomson Reuters sued for copyright infringement. Ross’s main defense was fair use. Judge Stephanos Bibas sided with Thomson Reuters, and in doing so wrote the first AI fair-use opinion on the books.
Wait — what’s a headnote, and can you even copyright it?
This part trips people up, so it matters. A judicial opinion — the actual ruling a judge writes — is a government work and is not copyrightable. Anyone can copy it. So how did Thomson Reuters win?
Because Ross didn’t copy the opinions. It copied Westlaw’s headnotes: the human-written summaries Westlaw editors craft to distill each case. Judge Bibas compared a headnote to a sculptor working a block of marble — the raw material (the opinion) is free for all, but the editorial choices about what to pull out and how to phrase it are original human expression. He found that 2,243 headnotes were original enough to be protected, and that Thomson Reuters owned valid copyrights in them.
The lesson for creators is immediate: copyright doesn’t protect the underlying facts or ideas, but it does protect the original way you select, arrange, and express them. Your unique take on public information can still be yours.
The judge changed his own mind
Here’s a detail that tells you how hard this question is. Bibas had looked at this case once before, in 2023, and reached a different conclusion. Back then he thought fair use was too close to call and should go to a jury. Then he invited both sides to re-brief the issue, reconsidered, and reversed himself — saying his first decision “had not gone far enough.” In February 2025 he granted summary judgment to Thomson Reuters outright, without sending it to a jury.
When a thoughtful federal judge flips his own ruling on the central legal question, that’s your signal not to trust anyone who says the answer is obvious.
How fair use actually works (the four factors)
Fair use isn’t a vibe — it’s a four-part test courts weigh together. Here’s each factor and how it landed in this case.
- 1. Purpose and character of the use. Is the new use “transformative” — does it add something new with a different purpose — or does it just repackage the original? This is usually the big one. Bibas found Ross’s use was not transformative: Ross took Westlaw’s editorial work to build a tool that did the same job Westlaw did — help people find relevant cases. It was also commercial. This factor favored Thomson Reuters.
- 2. Nature of the copyrighted work. More creative works get stronger protection; factual ones get less. Headnotes are fairly functional, so this factor actually leaned Ross’s way — but it’s a minor factor.
- 3. Amount used. How much did you take, and was it the “heart” of the work? This one was mixed and didn’t decide much.
- 4. Effect on the market. Does the copying harm the market for the original — or a market the owner could reasonably license? Bibas called this the single most important factor, and it favored Thomson Reuters decisively. Ross built a direct market substitute for Westlaw, and the court recognized that Thomson Reuters had an interest in the market for its data as AI training material.
Add it up: the two factors that carry the most weight — purpose and market harm — both pointed against Ross. Fair use failed.
Why Ross really lost: it built a competitor
If you take one strategic idea from this case, make it this. The thread running through the decision is that Ross was trying to replace the thing it copied. It wanted the same customers, in the same market, using material lifted from the very company it was fighting.
Fair use has always looked kindly on uses that are different from the original — commentary, criticism, parody, research, teaching, building something genuinely new. It looks harshly on copying that simply lets you compete with the creator using their own work. Ross fell squarely in the second bucket. That, more than any abstract theory about AI, is why it lost.
The huge asterisk everyone skips
Now the part the scary headlines leave out. Judge Bibas was explicit that Ross’s tool was not generative AI. It didn’t write new text or paint pictures. It was a search engine that, given a legal question, pointed users to relevant existing judicial opinions — public documents. And Bibas pointedly noted that only non-generative AI was before him.
That’s a massive limit. The cases keeping artists and writers up at night — image generators, chatbots, music models — involve generative systems whose defenders argue that training is transformative precisely because the output is something new. That argument was never tested here, because Ross’s product didn’t generate anything. So while Thomson Reuters v. Ross is the first AI fair-use ruling, it is not the ruling on whether training a generative model on your art or writing is fair use. That fight is happening in other courtrooms, and this decision doesn’t settle it.
One more piece of context: Ross Intelligence actually shut down in 2021, saying the cost of the litigation had crushed the young company long before any final verdict. And the story isn’t over — the case is now on appeal to the Third Circuit, with outside groups lining up to support Ross. The law here is still being written.
What this means for you
You’re not Westlaw and you’re not a startup, but this ruling hands working creators several genuinely useful lessons:
- Your original expression is protectable even when the facts aren’t. Like Westlaw’s headnotes, your unique summaries, arrangements, and phrasing of public information can be your copyright — even though the raw facts are free.
- “Transformative” is the magic word. Uses that add something new for a different purpose get real fair-use protection. Uses that just repackage your work to compete with you usually don’t.
- Market harm carries the most weight. If someone’s copying substitutes for your work — or for a licensing market you could reasonably tap, including AI training — that’s the argument with the sharpest teeth.
- Don’t over-read this case. It’s about a non-generative search tool that copied a competitor’s editorial content. It does not declare that training image or text generators on your work is illegal. That question is still open.
What to do now
- Register the work that matters. Copyright is strongest — and far easier to enforce — when you’ve registered it. Your original writing, art, and photography qualify.
- Keep records of your originals. Save drafts, files, timestamps, and posting history. If you ever need to prove a work is yours and was copied, that trail is your evidence.
- Think about markets, not just theft. If you’re worried about AI, the strongest framing is economic: here’s the licensing market being harmed. That’s the argument courts weigh most heavily.
- Follow the generative cases. This ruling is the opening chapter, not the verdict. The decisions that will most affect creators are the generative-AI cases still working their way through the courts.
Thomson Reuters v. Ross was billed as the moment AI met fair use — and it was. But it decided something narrow and specific: you can’t copy a rival’s original work to build a tool that competes with them and call it fair use. On the biggest question creators are actually asking, the court left the light on and the door open.
This article is general information, not legal advice. AI and copyright law is moving fast and differs by country and state — for a specific situation, check the current law where you are or talk to a qualified attorney.
IPSurge helps creators and small businesses understand and protect their intellectual property — the ideas, names, and work that make them unique.
Sources & further reading:
- Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., Memorandum Opinion (D. Del., Feb. 11, 2025) — official court PDF
- Court Rejects Fair Use Defense in AI Copyright Case (Goodwin)
- Thomson Reuters v. Ross Intelligence: Copyright, Fair Use, and AI (Round One) (Davis Wright Tremaine)
- Amici Back AI Company’s Third Circuit Appeal of Summary Judgment for Thomson Reuters (IPWatchdog)