And what do I care? I’m not a lawyer. Thomson Reuters v. ROSS is a case about legal research software, and nobody in it wrote a note of music. So why would I read thirty pages about Westlaw headnotes? Because it’s the first time a federal appeals court has ruled on whether training an AI on copyrighted material is fair use, and every music AI case now pending will be argued with this opinion in hand. In whatever circuit.
Super basics
Westlaw publishes “headnotes,” little summaries of key points in court opinions. ROSS wanted to build a search engine that answered legal questions in plain English. To train it, ROSS wrote about 25,000 practice questions, and used Westlaw headnotes to write them. Thomson Reuters sued. The court granted summary judgment in February 2025. On September 29, 2026 the Third Circuit affirmed, and the opinion was unsealed the next day. (that’s today.)
The court held that 2,243 of the headnotes are original enough for copyright, and that ROSS’s copying failed three of the four fair use factors. Those, by the way:
- The purpose and character of the use, including whether it’s commercial or nonprofit educational.
- The nature of the copyrighted work.
- The amount and substantiality of the portion used in relation to the whole.
- The effect of the use upon the potential market for or value of the copyrighted work.
ROSS won on the second, probably the least interesting, right?
How can AI training be fair use, ever?
Fair use is a defense for copying without permission. A court weighs the purpose of the copying, the nature of the work, how much was taken, and the effect on the market for the original work. So far it’s been somewhat split from my accounting and it may stay murky for a while. No court has answered the AI training question in general; that would be tough. These things are case by case. Two trial judges in San Francisco found training fair use in Bartz v. Anthropic and Kadrey v. Meta. Here the district court and now an appeals court have said no:
“In truth, this is no more than an ordinary copyright case.”
“Unlike necessity, ease is not a justification for copying.”
“This does not give ROSS carte blanche to violate copyright law merely because it incorporates AI.”
Whew.
So, I’m looking ahead to musicology-related issues. ROSS’s model “would not create any new expression; it would only return text passages from preexisting judicial opinions.”
Why a musicologist cares
This is fascinating!
The court separated ROSS from the generative cases because ROSS “cannot generate original expression” and because ROSS “trained its AI for the purpose of creating a commercial substitute for Westlaw.” Suno generates. And factor four, Suno is also, by the labels’ account, a substitute for the recordings it trained on.
And we’re down the road a ways. Warner has a deal with Suno. UMG has one with Udio. Suno’s v6 runs on catalog from Warner, BMG, and Believe. But we’re far from done.
ROSS said it copied the headnotes the way Accolade copied Sega (I do not know what Accolade was, but I sure remember Panzer Dragoon on my Sega Saturn) and the way Google copied Java, to get at something nobody owns, and that’s a significant part of what a forensic musicologist looks at in copyright infringement cases. Here, the court said those cases only work when there’s no other way in. ROSS had ten million public court opinions sitting right there. It used the headnotes because that was easier. “Unlike necessity, ease is not a justification for copying.” If you scraped recordings you could have licensed, that sentence sounds like it’s applicable to your cases too.
Let’s consider too, that every headnote was its own copyrighted work — of course it was. Originality is a low bar. So ROSS took whole works every time, and with no real transformation the court called that unreasonable. And here, evidently, bad faith counts against you quite a bit. ROSS got into Westlaw on borrowed and disguised accounts and the court held it against them. We’ve seen that in the music cases too. It doesn’t look good to have gone around anti-piracy protections to get the recordings.
ROSS also said the ruling would stall AI and that AI is a national security matter. The court asked for evidence. There wasn’t any.
The music AIs are going to talk about how they are “democratizing music” and there’s a ton to that, I can tell you. So many empowered musicians out there. But it’s hardly a national security matter.
Case by case, I say once again, so this doesn’t bind other courts very much. They’ll read it though. And it’ll be very interesting, I suspect, as each of these decisions leaves its mark on the ethereal and hard-to-get-at general thinking. In a month, Judge Eumi K. Lee in San Jose hears a motion in Concord v. Anthropic on October 21. It’s got a new exhibit now, probably.
So, that’s why I read it.
