110 incidents on record · 2026 Headlights Incident reports by Ellie Harris · Melbourne
10 new this week Library last updated 27 July 2026
← The incident library
HD-INC-093
Technology · United States · 2025 · Training-data provenance

Ross trained an AI legal-research tool on Westlaw headnotes obtained after being refused a licence, and a US court ruled it was not fair use, in what commentators called the first such decision

By Ellie Harris · Filed Conduct from around 2015; suit filed 2020

Alleged: Ross Intelligence; Thomson Reuters developed or deployed the AI system implicated in this incident. Details are drawn from public reports; parties are presumed innocent of any wrongdoing not established by an official finding.

Ross trained an AI legal-research tool on Westlaw headnotes obtained after being refused a licence, and a US court ruled it was not fair use, in what commentators called the first such decision

What happened

It was reported that in February 2025 a United States federal court ruled that Ross Intelligence, an AI legal-research startup, had infringed the copyright of Thomson Reuters by training its tool on material drawn from Westlaw, and that the use was not protected as fair use. It was described at the time as the first US decision to reject a fair-use defence for using copyrighted work as AI training data. The judge, Stephanos Bibas, had reached a different view in 2023, when he left the fair-use question for a jury, then reconsidered, invited new briefing, and changed his mind.

Ross had set out to build a search tool to compete with Westlaw. According to the ruling, it first asked Thomson Reuters for a licence to use Westlaw content and was refused because it was a competitor. It then obtained training material through a third party, LegalEase, in the form of “Bulk Memos” that the court found were built from Westlaw’s headnotes, the short notes that distil points of law from court opinions. The court found that the headnotes were original enough to be protected, that Ross had directly copied more than 2,200 of them, and that its use was not fair use. Weighing the four fair-use factors, the court found the first and the fourth, the purpose of the use and its effect on the market, favoured Thomson Reuters. It held the use was not transformative, because Ross took the headnotes to make it easier to build a competing legal-research tool, and that the clearest harm was to the market, both for Westlaw itself and for a potential market in licensing material to train legal AI. Ross had already shut down in 2020, saying it could not carry the cost of the litigation. The court agreed to let the decision be appealed to the Third Circuit, and stressed that only non-generative AI was at issue.

What an auditable version would have shown

The question the case turned on was simple to ask and hard to answer after the fact: what went into the model, and on what authority? An auditable version is designed to record the provenance of each source used to build a system, the dataset, the document, and the licence or permission behind it, as a signed entry made at the time of training. With such a record, material that traced back to Westlaw headnotes, gathered, the court found, after a licence had been refused, could have been identified before release, rather than reconstructed later by a court working through thousands of headnotes. In an auditable system, provenance is not left to litigation; it is meant to be recorded at the outset.

Where the gap was

On the court’s findings, the model was built on material the developer was not licensed to use, and where that material came from, and whether it was used lawfully, were questions a court had to resolve rather than facts already sitting on the record. A ConductRecord keeps the lineage of every training source, where it came from, how it arrived, and the permission it rests on, so a company can answer “what did you train on, and were you allowed to” straight from its own files. A ConstraintGate checks that permission before training starts, so unlicensed data is far less likely to slip quietly into the model. The dispute that took years to resolve was, in large part, an argument about records of provenance that were missing or contested.

What governance should have looked like

An organisation that trains or builds a model owes a clear answer to a basic question: where did the data come from, and did we have the right to use it? That answer should exist as a record made before the model is trained, not as an argument assembled once a lawsuit begins. Where a source cannot be tied to a licence or a lawful basis, the safer default is not to use it, and the record should make that decision, and its reason, visible. The reported lesson of this ruling is that provenance is a governance obligation, and that not keeping track of what a model was trained on is not a position a court is obliged to accept.

Failure Pattern: a model was trained on copyrighted material the developer was not licensed to use, and the developer could not show, from its own records, where the training data came from.

Governance Principle: the provenance and licensing of every source used to train or build a model should be recorded before use, so that a developer can show what went in and on what authority.

The reference implementation of ConductRecord and ConstraintGate is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.

Sources

The mailing list

Fresh incident reports every week. One email to match.

We add new incidents to the library regularly, and send a single short email each week with what's new. The library stays free and open; this is just how you keep up with it.

No tracking. Unsubscribe in one click.

The record

An auditable system would have produced a signed, tamper-evident record the moment this happened: what the system did, the version that did it, the basis it acted on, and the action taken, and Ross Intelligence; Thomson Reuters could have produced it on demand.

This is the record the system as deployed did not produce in a signed, auditable form.

What this teaches
Capture what happened when it happens
What the system did, the version that did it, the basis it acted on, and the action taken, recorded at the moment, not reconstructed after.
Sign it, so no one has to trust the record-keeper
A tamper-evident entry. Edit it later and the signature breaks. The record does not ask for the benefit of the doubt.
Make it verifiable by anyone
A court, a regulator, a customer's lawyer can check the record themselves, without taking the company, or us, at our word.

Headlights summarises publicly reported AI incidents. All summaries are independently written, attributed to their original sources, and intended for research and educational purposes. Allegations are identified as such until established through official findings.

Last reviewed June 2026. This report is based on the sources listed above and reflects information available at the time of review; later developments may not be captured. Where a person is described as charged with or alleged to have done something, that allegation is unproven unless a conviction or a court or regulatory finding is stated. Headlights publishes journalism and commentary, not legal advice.

Want to write back?

Direct to my inbox.

ellie@useheadlights.com →