What Does a Copyright Ruling Actually Establish?

On September 29, 2026, a federal appeals court decided a copyright case. By the time the opinion was unsealed the next afternoon, two opposing communities had each announced that the ruling was on their side.

One said it vindicated copyright law. The other said it implicitly confirmed that highly transformative generative-AI training is likely fair use.

They were reading the same document.

---

What the court held

The case was Thomson Reuters v. Ross Intelligence. ROSS was a legal-research startup that answered plain-language legal questions by returning passages from judicial opinions. To train it, the company used headnotes — the short summaries Westlaw’s editors write for individual points of law.

The Third Circuit held that the headnotes were copyrightable and that ROSS’s use of them was not fair use. Its reasoning turned mainly on purpose and market harm. ROSS’s tool “shares the same ultimate purpose” as Westlaw, “making ROSS’s use minimally transformative, at best” — a use the court called “highly commercial and minimally transformative.” The court also found that ROSS’s copying would harm two markets: Westlaw’s existing legal-research market, and a “rapidly developing” market for licensing headnotes as AI-training data.

That is a real ruling with real consequences. It is also, in the court’s own words, deliberately small.

---

What the court fenced off

The opinion goes out of its way to say what it is not deciding. “This is no more than an ordinary copyright case,” it says on page two. ROSS’s AI “was not a generative AI” — it “would not create any new expression.” In a footnote, the court discussed a Justice Department filing from separate OpenAI litigation — not a filing in this case — and said the generative-AI concerns raised there “do not apply here,” because ROSS “cannot generate original expression.”

The court did not decide whether training a generative model on copyrighted work is fair use. It said so, repeatedly. The generative-AI question is exactly the one the opinion leaves open — and it is the question everyone wanted answered.

---

What competing actors made it mean

What happened next is the point.

Within forty-eight hours, Mary Rasenberger, the CEO of the Authors Guild, said the ruling “vindicates copyright law. It clarifies that one cannot steal someone else’s works to train an AI system that displaces the very works the AI was trained on.” Maria Pallante of the Association of American Publishers framed it in terms of “a robust licensing market.” Thomson Reuters’s own spokesperson celebrated the decision.

On the other side, Adam Eisgrau of the tech group Chamber of Progress posted that the decision “implicitly confirms that highly transformative gen AI training … is likely fair use.”

Two opposite conclusions. Same opinion.

---

Why the gap matters

The interesting fact is not that the two sides disagree. It is that both of their claims reach past what the court actually said — though not to the same extent.

The rights-holder reading takes a ruling about a non-generative, same-purpose competitor — a tool that returned other people’s text, built by a company set up to compete directly with Westlaw — and turns it into a statement about generative AI in general. One part of that reading has firmer ground: the court did find harm to a developing market for licensing headnotes as training data. But it found that harm for a non-generative competitor whose use it had already called minimally transformative, and it said nothing about what such a market would mean for generative models. The AI-side reading does the mirror image: it takes the court’s silence on generative training and reads it as permission.

A single figure shows how easily precision slips. By ROSS’s own count, Thomson Reuters holds roughly 28 million headnotes. The opinion is not consistent about the next number. Its account of the facts says ROSS’s contractor wrote about 25,000 memos, drawing on thousands of headnotes; its fair-use analysis describes 25,000 headnotes copied into the memos. The two descriptions don’t obviously fit together. The district court examined a batch of 2,830 headnotes and held 2,243 of them infringing, on partial summary judgment.

The number that travels is the precise one — 2,243 — and it travels attached to the wrong noun, so that “the court found 2,243 headnotes infringing” becomes “ROSS trained on 2,243 headnotes,” which is not what happened. The precision is the whole discipline. Lose it, and a technically accurate number becomes a false claim.

---

Not equally wrong

The two over-readings are not symmetrical, and the difference is not that one side is more wrong.

They have different evidentiary status, and different downstream stakes.

The AI-side reading is an argument built on a single footnote — footnote 7, where the opinion turns to generative AI only to set it aside. An advocate points at that carve-out and says, in effect, the reasoning leaves room. That is rhetorical work, in a debate about a future case that has not yet been decided.

The rights-holder reading is already doing heavier work, and its own authors say so. Pallante’s statement casts the ruling as an input to “a robust licensing market” — that is, as something to be used in negotiations, not just reacted to. The Authors Guild, whose CEO called the ruling a vindication, is itself a plaintiff in ongoing copyright litigation against OpenAI. Several of the organizations doing the extrapolating are the organizations who will put the ruling to use.

So the question is not which side is more wrong. It is which extrapolation is still a talking point and which is already becoming an instrument. And on that question the two readings are not equal — not because one argument is worse, but because one is attached to a licensing agenda and a lawsuit, while the other is an opinion about a case that has not happened.

That is a distinction to watch, not an outcome to declare. Whether the rights-holder reading actually migrates from rhetoric into legislation or binding terms is an empirical question, and the right thing is to keep it open rather than settle it.

---

There is a principle underneath all of this, and it is worth naming once. Consequential rulings acquire political meaning faster than they acquire legal scope. The court spent pages narrowing what it was deciding. The public spent hours broadening it. The gap between the two is not a failure of communication; it is where the actual stakes live.

We do not yet know whether generative-AI training is fair use, and we may not learn it in a single stroke. The answer is more likely to arrive in fragments — a holding about one kind of model, one use, one market, in one circuit — each as narrow as this one, and each broadened the moment it lands.

So watch for the same two events each time. What a ruling actually holds, and what it gets made to mean. Those are two different events, and they will not arrive at the same time.

---

Dale Joseph is the author of Thought Partners: Preserving Cognitive Sovereignty in the Age of AI and founder of the Emergence Institute. He worked for years as a consultant helping install hospital networks before turning to writing and systems thinking. He lives in Boynton Beach, Florida.

Correction — October 5, 2026: After checking this essay against the text of the Third Circuit opinion, I revised several passages. The Justice Department filing discussed in footnote 7 was made in separate OpenAI litigation, not in this case. The district court granted partial summary judgment on 2,243 headnotes; it did not enter final judgment. The opinion describes the 25,000 figure inconsistently, as training memos in one place and as copied headnotes in another, and the essay now says so. The essay also now notes that the court found harm to a developing market for licensing headnotes as AI-training data, and the discussion of the rights-holder reading has been adjusted to reflect that. The essay’s central point, that the opinion expressly distinguishes generative AI, is unchanged.