What the 2025 AI fair-use rulings mean for companies licensing data

What did the 2025 AI fair-use rulings decide?#

The 2025 AI fair-use rulings were three US trial-court decisions on whether copying published works to build AI tools was fair use, and they reached different results on different records. Thomson Reuters v. Ross Intelligence rejected fair use for a competing legal-research tool, while Bartz v. Anthropic and Kadrey v. Meta Platforms found training on books to be fair use on the records before those courts, with important limits.

Each ruling bound only its parties, and the judges weighed the same factors differently. A general counsel should treat them as signals of how courts approach the questions, not settled law, and should read the opinions or ask counsel for case-level detail before relying on any one.

What has happened since the rulings?#

Since the rulings, one case has reached an appellate court and another has settled, which changes how the 2025 decisions should be read. In late September 2026, a Third Circuit panel affirmed in Ross that the headnotes are copyrightable and that copying them to train a legal-research AI was not fair use; LawNext reported it as the first federal appellate decision on fair use in AI training. Because Ross involved a non-generative tool, how far it reaches generative model training is still debated.

Bartz did not go to trial on the pirated library. The parties agreed to a $1.5 billion class settlement that received final approval on July 20, 2026. It requires destruction of the downloaded pirated files and copies derived from them, and its release covers only past conduct through August 25, 2025, not claims about AI outputs or future conduct. A settlement is not a ruling on fair use, but it shows what unlawfully acquired training material can cost.

Many other AI copyright suits are still moving through the courts, and decisions on other kinds of works, on market dilution and on licensing markets could shift the picture again. For operating companies those developments matter mostly at the margin, for the reasons below.

Where the rulings left questions open: licensing markets and pirated copies#

The 2025 rulings left two questions open that matter to data suppliers: whether lost licensing fees count as market harm, and how much the source of the training copies matters.

On licensing, Judge Chhabria in Kadrey rejected the argument that lost fees for licensing books as AI training data were a cognizable market harm, calling that reasoning circular. The Copyright Office's May 2025 report took the opposite view where a licensing market exists or is likely, and in Ross the market-effect factor favored Thomson Reuters because Ross meant to build a substitute product. Judge Chhabria also wrote that it is hard to imagine fair use where copying enables a potentially endless stream of competing works, a signal rather than a holding that plaintiffs with better evidence of market dilution could win.

On sources, Judge Alsup held that building a central library from pirated books was not fair use even though training itself was. In Kadrey, Meta acknowledged downloading books from shadow libraries, and the court granted Meta only partial summary judgment on the training claim on the record presented; the ruling did not resolve every claim in the case. Because the courts did not apply one rule on acquisition, developers increasingly ask suppliers for documented, lawful provenance instead of relying on a defense.

The fair-use factors, read from a data supplier's seat#

The four fair-use factors are the purpose and character of the use, the nature of the copyrighted work, the amount used, and the effect on the potential market. Read from a data supplier's seat, each factor maps to a practical question about operational records.

The second factor deserves attention. Facts such as a part number, a delivery time or a defect code are not protected by copyright, while the expression in a support agent's reply or an engineer's design note can be. Much of the value of operational data sits in that mix, and its protection often comes more from confidentiality and contract than from copyright.

Three implications for companies licensing data#

The rulings carry three practical implications for operating companies that license data, and none of them make licensing less relevant.

Confidentiality is often the stronger shield. Records kept confidential and released only under contract are better placed to keep trade secret protection where it applies than records that have been published or shared loosely.

• 1Fair use does not open your systems. The doctrine can excuse certain copying of works someone already has; it does not give a developer access to your Zendesk instance, Salesforce history or private GitHub repositories. Private records reach developers through a license, and unauthorized access raises separate legal issues.
• 2Provenance matters more. Courts looked at how training copies were obtained, and the Bartz settlement required destruction of pirated files, so developers ask suppliers to document where data came from, who held rights in it and what was removed before delivery.
• 3Contracts carry the controls. Copyright doctrine will not limit retention, model release or downstream use of your licensed records; the license terms do, so permitted use, deletion and distribution clauses deserve careful drafting.

Which license terms matter more now?#

The license terms that matter more after the rulings are the ones that describe where data came from and limit what happens to it after delivery. Because copyright doctrine leaves so much to case-by-case judgment, a supplier gets certainty by writing its boundaries into the agreement.

Illustrative: a distributor's general counsel reads the headlines#

Illustrative: the general counsel of a fictional industrial distributor reads coverage of the fair-use rulings and asks whether the company's records could simply be used for training without its involvement. The records in question are years of order exceptions, supplier delay notes and customer service resolutions stored in Epicor and a shared helpdesk.

Counsel concludes that the rulings do not reach these records, because no developer can access them without the company's permission. The decision shifts to whether to license them at all. The company chooses a narrow, non-exclusive license covering prepared exception records with customer names removed, a permitted-use clause limited to model training and evaluation, no redistribution of the dataset, and deletion at the end of the term.

How SourceX approaches rights after the rulings#

SourceX treats every package as a licensed transaction rather than a fair-use question. Each package goes through the Rights step of the SourceX five-step transaction, Supply, Rights, Preparation, Approval and Delivery, where ownership, contract restrictions and privacy limits are reviewed with the supplier.

For each approved package, a SourceX Evidence Packet sets out provenance, licensing rights, permitted use, the privacy record and release authorization. That documentation answers the provenance questions the rulings brought forward, and the supplier approves each step before anything is delivered.