A managing partner asks a reasonable question: we already pay for the treatises, the form books, and the research platform, so why can we not point our own model at them? The answer is that the subscription you bought is not the license you would need, and the case law now says so from two directions.
Two Gates, Not One
Every treatise-training question passes through two gates in order. The first is contract. The second is copyright. Most firms argue about the second and lose at the first.
Gate one is your subscription agreement. Research platform and treatise licenses commonly restrict use to authorized users for internal research, and prohibit systematic downloading, bulk extraction, redistribution, and the creation of derivative databases. Ingesting a licensed corpus into a model or a vector index is, in most agreements written before 2023, exactly the activity the terms were drafted to prevent. Fair use is not a defense to breach of contract.
Gate two is copyright, and that is where the 2025 and 2026 decisions land.
Bartz: Lawful Acquisition Was the Hinge
In Bartz v. Anthropic, the Northern District of California separated two things that get conflated in the trade press. Training a large language model on books that were lawfully acquired was treated as fair use. Building and retaining a library of pirated copies was not, and the price of that distinction was the largest copyright settlement on record.
The court granted final approval on July 20, 2026: a $1.5 billion fund covering 482,460 works, roughly $3,000 per work, with output claims expressly preserved rather than released. That last detail is the one practitioners keep missing. The settlement resolved how the copies were obtained. It did not resolve what the model produces.
The Bartz takeaway
Provenance is a line item, not a footnote. A model trained on lawfully obtained material sits in a fundamentally different legal position than one trained on material someone downloaded from a shadow library, even if the training method is identical.
Ross: The Case That Actually Answers the Treatise Question
Thomson Reuters v. Ross Intelligence is closer to home because the material at issue was legal research content and the product at issue was a legal research tool. Ross sought a Westlaw license, was refused, and obtained training material through a third party in the form of Bulk Memos built from Westlaw headnotes. The court held that training a competing legal research product on that material was not fair use, reasoning that "the effect on a potential market for AI training data is enough" to weigh against the user on the market-harm factor.
Read that alongside Bartz and the shape of the rule emerges. It is not "AI training is fair use." It is: lawful acquisition helps you, and competing with the source in its own market hurts you. A firm training a research assistant on a publisher's editorial content is squarely in the second category.
The appeal is pending in the Third Circuit, argued in June 2026. A decision is expected within months and will move this analysis, in either direction. Treat any current position as provisional.
Kadrey: The Market Dilution Warning
Kadrey v. Meta did not deliver a plaintiff win on the record before the court, but it delivered the theory that should worry anyone building on licensed publisher content: that flooding a market with machine-generated substitutes can itself be the cognizable harm, even without verbatim output. The plaintiffs there failed on evidence, not on logic. The next set of plaintiffs will have read the opinion.
What This Means for Your Firm
| Pathway | Posture | Why |
|---|---|---|
| Negotiated AI rights | Generally workable | An amendment or rider that expressly permits ingestion, indexing, embedding, and retrieval for internal use, with defined retention and deletion terms. |
| Vendor-native AI features | Generally workable | The publisher's own AI layer over its own corpus. The rights question is already resolved inside the subscription you bought. |
| Retrieval inside licensed seats | Fact-dependent | Retrieval that keeps content behind per-seat authentication and returns short, attributed passages to licensed users is a materially different posture than building a training corpus. It is still governed by the contract. |
| Your own work product | Strongest ground | Firm briefs, memos, forms, and closed-matter files that the firm owns or controls, subject to confidentiality, privilege, and client consent obligations. |
| Scraping or bulk-copying a licensed database | Do not | This is where both gates close at once: breach of the subscription terms, plus the fair use analysis that Ross lost. |
So the honest answer to the title question is: not on the license you already have. You can get there with negotiated AI rights, with the vendor's own AI features, with retrieval that stays inside licensed seats, or with your own work product. Those four routes cover most of what firms actually want, and none of them require betting the firm on a fair use argument that a district court has already rejected in this exact context.
Before you ingest anything
- Pull the actual subscription agreement and read the use restrictions and any AI or machine-learning clause.
- Document provenance for every corpus you touch, including who acquired it and under what terms.
- Ask whether the output competes with the source. If it does, expect the market-harm factor to run against you.
- Separate retrieval from training in your architecture and in your paperwork. They are not the same act.
- Run confidentiality, privilege, and client-consent review before firm work product enters any index.
This is also a COUNSEL problem, not only a copyright problem. The provenance of a corpus, the authority to use it, and the record that proves both are governance artifacts. Firms that can produce them on demand will license content on better terms than firms that cannot.
Source Ledger
| Source | Type | Article use |
|---|---|---|
| Bartz v. Anthropic PBC docket (N.D. Cal.) | Court docket | Settlement approval posture and case history. |
| Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence, No. 25-2153 (3d Cir.) docket | Court docket | Appeal status of the headnote training ruling. |
| Authors Guild — final approval of the $1.5 billion Anthropic settlement | Industry organization report | Approval date, class size, and per-work figure. |
| Pearl Cohen — analysis of the approved settlement | Law firm analysis | Second source on settlement terms and preserved claims. |
| JURIST — record AI copyright settlement approved | Legal news | Independent confirmation of approval. |
| Baker Botts — Third Circuit oral argument in Ross | Law firm analysis | Argument posture and issues on appeal. |
| Third Circuit sets June 11 oral argument in the first AI fair use appeal | Legal commentary | Second source on the argument date. |
| AP News — Anthropic copyright settlement coverage | Wire reporting | General background on the settlement. |
| Kadrey v. Meta Platforms, Inc. (N.D. Cal.) | Court decision | Market dilution reasoning discussed in the analysis. |
Related LegalTek.ai reading
Matthew A. Mishak, Esq. is the Managing Attorney of Mishak Law LLC and the Founder and CEO of LegalTek.ai (SilverTung), an AI powered legal practice management and governance platform. He brings twenty years of Ohio legal practice across domestic relations, criminal defense, and municipal law, and is the architect of the COUNSEL framework operationalizing ABA Formal Opinion 512.
Disclaimer: This article is for general informational purposes only and does not constitute legal advice. Case law in this area is developing and the Ross appeal remains pending. Attorney review required before reliance. LegalTek.ai is a technology company, not a law firm.
