Last week a federal appeals court told an AI company that “fair use” is not a magic spell. Ross Intelligence built a legal-research AI by feeding it Westlaw’s copyrighted headnotes, then claimed fair use when Thomson Reuters sued. The Third Circuit said no.
If you are a photographer, a writer, or anyone who owns creative work, this is the most important AI copyright ruling you have probably never heard of. Here is what happened, what the court actually decided, and why it matters for your archive.
The case is Thomson Reuters v. Ross Intelligence, decided by the Third Circuit. It did not involve a chatbot or an image generator. It involved headnotes, the short editorial summaries at the top of published court opinions. That ordinariness is exactly why the ruling matters. The court did not get distracted by the technology. It asked the same questions copyright law has always asked, and the answers were damning.
What Ross Intelligence Did
Westlaw’s headnotes are short editorial summaries that appear at the top of published court opinions. They are written by human editors, and Thomson Reuters owns the copyrights. Ross Intelligence, building a competing AI-powered legal research tool, fed thousands of those headnotes into its system to train it.
When Thomson Reuters sued for copyright infringement, Ross did what every AI company does now: it claimed fair use. The argument, in essence, was that training an AI on copyrighted material is transformative enough to qualify, regardless of what the material was or what the AI was built to do.
What the Court Actually Decided
The Third Circuit rejected that argument, and its reasoning was methodical rather than dramatic. The court ran the use through the same four-factor fair use test that has governed copyright cases for decades.
First, purpose and character. The use was commercial, built to power a product Ross intended to sell. And it was, at best, only minimally transformative. The AI was not commenting on the headnotes, criticizing them, or turning them into something new. It was absorbing them to perform the same function they perform.
Second, the nature of the work. Headnotes are creative editorial expression, not bare facts, which weighs against fair use.
Third, the amount taken. Ross used the headnotes wholesale, not in fragments.
Fourth, and most damaging, market effect. Ross built its product to compete directly with Westlaw. Read that again. They did not just train on the work. They trained on it to replace it.
The Part Everyone Will Get Wrong
Here is the part everyone else will get wrong: this is not a ruling that training AI on copyrighted material is always infringement. The court was careful. It did not declare every AI training use unlawful, and it left room for cases where the use genuinely is transformative or does not threaten the original’s market.
But here is what it did do: it treated an AI case as a plain copyright case. Purpose, nature, amount, market effect. The same four factors that have protected my negatives for 27 years. No special AI exception. No new rule. Just copyright law, applied as written.
That framing is the real precedent. AI companies have spent two years arguing that training deserves its own lenient standard, as if the technology itself were a fifth fair use factor. This court declined the invitation.
The “Transformative” Defense, Translated
The word you will hear most in this debate is “transformative,” so it is worth being precise about what it means. In copyright law, a transformative use is one that adds something new, with a further purpose or different character. A parody that comments on the original is transformative. A thumbnail that helps you find the original is transformative in purpose, even though the image is the same.
What Ross argued was something looser: that the AI process itself is transformative because the output differs from the input. The court did not buy it, and the reason matters for photography. If “the machine changed it” were enough, then any commercial copying laundered through software would qualify. A photocopier with extra steps is still a photocopier.
Apply that to image generators. An AI that ingests a photographer’s portfolio and outputs new images “in the style of” that photographer has not commented on the work, criticized it, or given it new meaning. It has learned to produce a substitute. Under the Third Circuit’s logic, that is Ross’s argument with a different coat of paint, and it fails the same test.
What This Means for the AI Industry’s Legal Playbook
For two years, the industry’s legal strategy has rested on a bet: that courts would treat AI training as categorically different from ordinary copying, and that fair use would stretch to cover it. This ruling shrinks that bet considerably.
It does not end the fight. Other cases with different facts will come out differently, and a genuinely transformative use, one that does not compete with the original, can still qualify. But the blanket defense, the idea that training is fair use as a category, just lost its best test case. Expect AI companies to talk more about licensing now, because the alternative, litigating every use under the four factors, is a war of attrition they cannot be sure of winning.
That shift is already visible in the market. Licensing deals between AI companies and large publishers have been announced across the industry. The companies paying for licenses are implicitly conceding what Ross denied: that the underlying material has owners, and owners get paid.
The Archive as the Test Case
Here is why this case feels personal to anyone who has spent decades building a body of work. My archive is 27 years of frames. Every one of them was a decision: where to stand, when to shoot, how to see. Magazines licensed those decisions. They paid for the use they needed, for the time they needed it, and the rights reverted to me. That is the entire business model of professional photography, and it depends on one premise: the person who made the work controls how it is used.
AI training breaks that premise silently. Nobody asks. Nobody licenses. The work is scraped, ingested, and converted into a product that competes for the same clients. When a publication can generate a passable celebrity portrait “in the style of” a working photographer instead of hiring one, the market effect factor is not theoretical. It is the invoice that never arrives.
The Third Circuit understood this dynamic in the legal research context. A lawyer who stops subscribing to Westlaw because Ross does the same job is the market harm, made concrete. A magazine that stops commissioning portraits because the generator is cheaper is the same harm, wearing different clothes.
None of this requires believing AI image generation has no legitimate uses. It requires believing that the people whose work powers the models get a say in how that work is used. That is not a radical position. It is the position every licensing deal in the history of photography has taken. The only thing new is a set of companies asking to skip the negotiation.
What To Do With This Information
A court opinion does not enforce itself, but it changes the negotiating landscape. A few practical notes for working photographers:
Register your copyrights. The four-factor test protects you in court, but registration is what makes court worth entering. Timely registration unlocks statutory damages, which is what makes infringement cases economically viable. Without it, you are limited to actual damages, which are hard to prove and rarely worth the fight.
License explicitly. Every contract should state what the client may and may not do with the images, including AI training. Silence in a license is an argument waiting to happen. If your existing contracts predate the AI era, and most do, consider updating your standard terms now.
Keep your archive organized. Ownership claims are only as strong as your records. Know what you shot, when, and who licensed it. Metadata embedded at the time of creation is evidence that travels with the file.
Watch the licensing market. As AI companies move from scraping to licensing, the photographers with organized, registered, well-documented archives will be the ones positioned to get paid. The ones with a hard drive full of unnamed files will not.
The Bottom Line
Fair use was never a blanket permission slip, and this court refused to turn it into one for AI. If your business model is training on someone’s work to compete with them, expect the same four-factor test that has always applied. The law did not change. The excuses just ran out.
For photographers, the takeaway is simple. Your archive is an asset, the law recognizes it as one, and this ruling just made that recognition harder to argue around. Act like an owner, because the courts just confirmed that you are one.
I wrote about what this ruling means for photographers, and for anyone sitting for a portrait, on dclee.net.. ..
