For the past three years, the AI industry’s entire legal strategy has basically boiled down to one argument: training on copyrighted stuff is fair use, full stop, don’t worry about it. And it mostly worked. Courts kept nodding along. Then a court actually looked closely at one of these cases and said, well, actually, no. Not this time.

I’m talking about the Thomson Reuters v. ROSS Intelligence case, where a federal judge ruled that ROSS’s use of Westlaw’s copyrighted headnotes to train its competing legal AI tool was not fair use. And look, I know “legal research software” doesn’t sound like the sexiest battleground for the future of artificial intelligence, but trust me, this one matters way more than the name suggests.

Wait, Didn’t AI Companies Keep Winning These Cases?

Yeah, that’s the thing. For a while there it felt like every ruling went the same direction. Judges kept buying the “transformative use” argument – the idea that feeding copyrighted text into a machine so it can learn patterns is fundamentally different from, say, photocopying a book and selling it. Courts in other AI cases leaned toward that logic. It made sense on paper, and companies like OpenAI and Meta built entire defense strategies around it.

But this case was different in one really important way. ROSS wasn’t just training a general-purpose chatbot on a huge pile of internet text. It built a direct competitor to Westlaw, using Westlaw’s own editorial work (those headnotes – basically short, lawyer-written summaries that make searching case law way easier) to do it. That’s not “learning language patterns from the broader internet.” That’s “using your rival’s labor to build the exact product that competes with them.” Judge Stephanos Bibas, who actually sits on the Third Circuit Court of Appeals but was hearing this one at the district court level, didn’t buy the fair use defense. Not even close, from what I can tell.

Okay But Why Does Everyone Keep Saying “Third Circuit”?

Quick aside because this bugged me when I first saw the headlines going around – a lot of people (including, respectfully, whoever wrote that Reddit post) are calling this a “Third Circuit ruling,” and that’s a little misleading. Bibas is a Third Circuit judge, sure, but he was sitting by designation on a district court in Delaware. That’s a meaningfully different thing than the actual Third Circuit Court of Appeals weighing in. It doesn’t carry the same binding weight across the circuit. Does it still matter? Absolutely. Is it the appellate smackdown some headlines are making it sound like? Not quite yet. Give it time, this thing’s probably headed for appeal anyway.

So What Actually Tipped the Scales Here?

A few things, and honestly they’re worth sitting with because they tell you where the next wave of AI lawsuits is going to focus.

• ROSS and Thomson Reuters are direct competitors – that market substitution factor weighs heavy in fair use analysis
• The headnotes aren’t raw facts, they’re original editorial expression, which courts protect more strongly
• ROSS used the material to build a product meant to replace Westlaw subscriptions, not to create something wildly transformative

The court basically said: look, if your AI tool ends up competing in the exact same market as the copyrighted content it trained on, and it’s not doing anything fundamentally new with that content, you don’t get to just wave the “fair use” flag and walk away. That’s a pretty narrow and specific finding, but it’s also exactly the kind of argument plaintiffs in the bigger cases – the New York Times v. OpenAI case, the various author lawsuits against Meta and others – have been making this whole time.

“This is the first real crack in the dam. Everyone assumed fair use would just… hold. It didn’t hold here.”

Why This Isn’t Just a Legal Research Footnote

Here’s the thing that I think people are missing. This ruling probably doesn’t nuke the big general-purpose AI models overnight – training ChatGPT on a huge chunk of the internet is a different fact pattern than training a direct Westlaw competitor on Westlaw’s own editorial content. Different facts, different outcome, that’s how law works, I know, boring. But what this case does is hand every plaintiff’s lawyer in every other AI copyright suit a new favorite exhibit. “See, your honor? A court already said no.”

That matters psychologically even more than it matters legally, not gonna lie. Fair use cases live and die on judges feeling like they’re not the first ones to go out on a limb. Now there’s precedent, even if it’s narrow, even if it’s a district court sitting by designation, even if it’s getting appealed. It’s a crack. And once there’s a crack, more lawyers start poking at it.

I’ve seen this pattern before, actually, in a totally different context – early internet piracy cases. The first few rulings everyone assumed would go one way, and then one weird fact pattern breaks through, and suddenly every argument downstream shifts. Doesn’t mean the whole dam breaks immediately. But the water’s moving differently now.

What This Actually Means

If you’re an AI company right now, this ruling should scare you a little, especially if your product competes directly with whoever’s data you trained on. The “it’s just learning patterns” defense works a lot better when you’re not literally replacing the thing you learned from in the marketplace. ROSS basically built a knockoff using the original’s homework, and the court saw it for exactly that.

If you’re a publisher, a news outlet, an author whose books ended up in some training set without permission – this is the first real win you’ve gotten to point to. It’s not a knockout. It’s barely even a decision, in the grand scheme of all the AI copyright litigation still working through the system. But it’s something. And “something” is more than these companies have had to deal with for the last three years.

My honest prediction? This gets appealed, the appeal takes a while, and in the meantime a dozen plaintiffs’ lawyers cite it in every single filing they make between now and whenever the appellate courts actually settle this. The big question nobody’s answered yet – and won’t be answered for a while – is whether this logic holds up when you’re not dealing with a direct competitor. Does training a general chatbot on news articles count the same way? We genuinely don’t know yet. Anyone who tells you they know for sure is guessing. I’m guessing too, if I’m being honest, I just happen to think the water’s starting to move in a direction the AI industry really didn’t want it to.