Legal AI Training Data Ruling Lands as Harvey Hits $15.5B

Two signals in one month

Two developments landed weeks apart this fall, and they point in different directions for legal technology. On September 9, Harvey raised $550 million at a $15.5 billion valuation. That is up from $11 billion in March, which makes Harvey the highest-valued and best-funded startup in legal AI.14 Three weeks later, on September 29, the Third Circuit became the first federal appeals court to decide whether training an AI system on copyrighted material is fair use. It ruled against the AI developer.1

The two events are related. Investors are betting that AI will run through the daily work of law firms. The courts are drawing the first clear lines on what data those systems can lawfully learn from. A third pressure sits underneath both: lawyers keep filing briefs with citations an AI made up, and judges are punishing them harder each time. The way I read it, legal AI's biggest risk right now is not the technology. It is whether data is sourced legitimately and whether humans check the output. The companies and firms that treat those two issues as core strategy, not compliance paperwork, are the ones positioned to benefit.

The money

Harvey's rise has been steep. The company was valued at $3 billion in February 2025, then $5 billion in June 2025 and $8 billion in December 2025. The $11 billion mark followed in March 2026.13 The September round was co-led by Lightspeed Venture Partners and Diffusion, a new firm. Sapphire Ventures and Whale Rock joined as new investors.16 Harvey says more than 100,000 lawyers at about 1,300 organizations use its software, including most of the AmLaw 100.17 Another account puts usage at 80 of the 100 highest-grossing U.S. law firms.20

Coverage disagrees on what the round is worth relative to revenue. Revenue is reported at more than $400 million, which one analysis translates into a revenue multiple of roughly 39x.16 Another deal summary puts the multiple at about 14x annual recurring revenue. It also dates the round to October 1 rather than September 9.12 Most coverage says September 9, so the second account's figures should be treated with caution.1416 The gap matters because it decides whether this is a stretched valuation or a defensible one. With revenue above $400 million, 39x is the figure the arithmetic supports.

The new money is also going somewhere new. Harvey plans to spend it on building its own AI models instead of relying only on frontier labs like OpenAI and Anthropic. That puts it in more direct competition with the companies whose models it has licensed.16 The round came shortly after Harvey released its first post-trained open-weight model.14 Earlier analysis had flagged the obvious weakness: if Harvey does not own a frontier model, the gap between its product and a general chatbot loaded with legal prompts could shrink.18 Building its own models is Harvey's answer to that. It also means Harvey now has to decide what data its models train on, which is exactly the question the Third Circuit just addressed.

Harvey is not the only beneficiary. Stockholm-based Legora raised a $600 million Series D at a $5.6 billion valuation.13 Clio is valued at $3 billion and acquired vLex for about $1 billion.19

The ruling

The case, Thomson Reuters v. ROSS Intelligence, began in 2020. ROSS shut down in 2021, citing litigation costs.6 The dispute was over Westlaw headnotes, the short summaries Thomson Reuters editors write to flag points of law in court opinions.1 ROSS had asked to license Westlaw content and was refused because it was a competitor. It then obtained about 25,000 "Bulk Memos" from a contractor, LegalEase, and used them to train its search tool.6 Judge Stephanos Bibas found that ROSS infringed 2,243 headnotes. The Third Circuit panel unanimously affirmed in an opinion by Judge Tamika Montgomery-Reeves.9

The court ruled against ROSS on three of the four fair-use factors. It called the use commercial and only minimally transformative, because ROSS used the headnotes for the same purpose Westlaw does: helping lawyers find relevant law.1 It found that ROSS copied each headnote in full.4 It also identified harm in two markets: legal research itself, and a developing market for licensing headnotes as AI training data.1 The court also said the evidence showed ROSS at times acted in bad faith, including using login credentials to reach Westlaw after being told not to.9 And it noted that ROSS had about ten million uncopyrighted judicial opinions it could have trained on instead.9

The licensing-market reasoning reaches furthest. The court accepted that Thomson Reuters already trains its own AI products on headnotes, and that ROSS took away its chance to license them for the same purpose.1 In effect, a rights holder's own AI plans now count as evidence of market harm. Publishers, database operators and other owners of curated content can rely on that argument.

Where coverage disagrees on scope

Most commentators agree the ruling is narrow, but they emphasize different things. Ballard Spahr notes that ROSS's tool returned existing court passages rather than generating new text, so the central question about generative AI is still open.1 Plagiarism Today goes further. It argues that unless general-purpose AI can be shown to compete directly with its training material, OpenAI and Anthropic have little to fear from this part of the decision.7 Tech Times reads it more sharply. In its view the decision makes generative-versus-non-generative the main dividing line in U.S. AI copyright law, and it stresses that fair use for generative training is still unsettled.30 It also points out that the decision binds federal courts in Delaware, New Jersey and Pennsylvania but not the Ninth Circuit, where most generative-AI cases are being heard.30

For legal tech specifically, the generative carve-out offers less protection than it seems. Many legal AI products retrieve, summarize and rank existing law for the same lawyers Westlaw serves. One analysis puts it plainly: if your training data and your product's output serve the same function for the same user, you are in ROSS's position.8 In my view, a legal AI startup that trained on a publisher's editorial layer cannot count on the carve-out just because its interface generates text.

The wider litigation picture

The ROSS decision lands among trial-court rulings that point in different directions. In Bartz v. Anthropic, Judge William Alsup found that training on lawfully purchased books was fair use. He also found that keeping a library of pirated copies was not. Anthropic then paid $1.5 billion to settle, the largest copyright settlement in U.S. history.28 Coverage disagrees on when the settlement became final. One account says final approval came on May 14, 2026.23 Others say the May 14 date was the hearing and that the judge entered final approval and judgment on July 20.2721 The July date is better supported. Some summaries also describe the settlement as if it decided fair use, but approving a settlement is not a ruling on the merits.21

In Kadrey v. Meta, the court ruled for Meta on fair use, but claims about Meta's torrenting are still alive.24 The consolidated OpenAI litigation in Manhattan has reached cross-motions on fair use, with evidence that ChatGPT reproduced articles verbatim at the center of The New York Times' case.22 The common thread is that how training data was obtained, and what the system produces, now matter as much as the training itself.

The adoption problem

While courts debate what goes into models, judges are dealing with what comes out of them. Researcher Damien Charlotin maintains a public database of cases involving AI-fabricated citations. One tracker reports that as of September 21 it had logged 2,046 cases, up from about 200 a year earlier.32 Other counts taken earlier in the year were lower, at about 1,490 decisions as of May.31 The direction is the same either way.

Penalties are rising too. The Sixth Circuit ordered two lawyers to pay $15,000 each in Whiting v. City of Athens.31 In Mississippi, Judge Sharion Aycock canceled a trial after both sides filed fake citations. She barred the lead lawyers from appearing in her district for two years, writing that the duty to verify cannot be outsourced to technology.38 Reports on the Oregon winery case Couvrette v. Wisnovsky conflict. Several put combined sanctions at around $110,000.3240 Another account cites $15,500.39 In England, a disciplinary tribunal struck off a lawyer over AI misuse for the first time in September.39

Specialized tools do not remove the risk. A widely cited Stanford study found that legal-specific research tools still hallucinated on more than one in six queries.34 One sanction in the Fifth Circuit involved a lawyer using commercial legal AI platforms.34

What it adds up to

The investment case for legal AI is built on adoption, and adoption is happening. Courts are not punishing lawyers for using AI. They are punishing them for filing output nobody checked.35 Meanwhile, the first appellate ruling on training data protects curated legal content, which strengthens incumbents like Thomson Reuters that own large editorial archives. For startups, that makes legitimately sourced training data a competitive advantage rather than a cost. For law firms, the practical rule has not changed: whatever tool a lawyer uses, the lawyer who signs the filing is responsible for it.