Skip to content

Thomson Reuters v. Ross: The Blueprint for AI Theft

Delaware court: training AI on Westlaw headnotes is copyright infringement. Turning millions of copyrighted documents into a competing product isn't fair.

By THEYDIDNTASK

TL;DR

  • —On February 11, 2025, Judge Stephanos Bibas delivered a ruling that sent shockwaves through Silicon Valley: training AI on copyrighted content without permission is not fair use.
  • —At least, not when you're building a competing product.
  • —So Ross did what every AI startup does: they scraped it anyway.

This case may reshape how AI companies source training data.

On February 11, 2025, Judge Stephanos Bibas delivered a ruling that sent shockwaves through Silicon Valley: training AI on copyrighted content without permission is not fair use. At least, not when you're building a competing product.

The Setup

Ross Intelligence wanted to build an AI-powered legal research tool to compete with Thomson Reuters' Westlaw. Small problem: they needed training data. They approached Thomson Reuters for a license. Thomson Reuters said no.

So Ross did what every AI startup does: they scraped it anyway.

They hired a contractor called LegalEase to create "Bulk Memos"—essentially laundering Westlaw's proprietary headnotes (short case summaries) into training data. 2,243 headnotes were copied to train their competing AI.

The Ruling

Judge Bibas analyzed all four fair use factors:

  • Purpose: Commercial use to build a competing product. Loss.

  • Nature: Headnotes are creative compilations. Slight win for Ross.

  • Amount: Headnotes not displayed to users. Slight win for Ross.

  • Market Effect: Ross was building a direct market substitute AND harming the emerging market for AI training data licenses. Loss.

Final Score: Thomson Reuters wins. Fair use does not apply.

Why This Matters

Client counsel describe this as the first major decision regarding copyright liability for AI developers rejecting fair use for AI training data. The court explicitly noted that copying copyrighted works to train a competing AI product is infringement.

"Ross copied headnotes to make development easier, not because it was necessary for innovation."

The implications are massive:

  • The court called the fourth fair use factor "the single most important element of fair use"

  • AI companies may need to license training data

  • The "fair use" defense just got a lot weaker

Concrete Actions

For Developers:

  • Audit your training data sources

  • Consider licensing content from creators

  • Document your fair use analysis

For Content Creators:

  • Monitor AI companies using your work

  • Consider joining collective licensing organizations

  • Document your copyright registrations

The Bottom Line: The free lunch is ending. Pay the creators.

Sources