Litigation — Thomson Reuters v. ROSS Oral Argument Held Today #
- ROSS, Westlaw appellate arguments tentatively set for June 11 (MLex) — The Third Circuit heard oral argument today (June 11, 2026) in Thomson Reuters v. ROSS Intelligence — the first AI training data fair-use case to reach US appellate court level. The court is deciding: (1) whether ROSS’s use of Westlaw headnotes to train its AI legal search engine was transformative fair use; (2) whether Westlaw headnotes meet the originality threshold for copyright protection. No ruling is expected at the argument itself — Third Circuit opinions typically follow weeks to months after argument. The record is now complete; the waiting period begins.
- AI Lawsuits in 2026: Settlements, Licensing Deals, Litigation (AI Business, 2026) — Bartz v. Anthropic settled for $1.5 billion: Judge Alsup’s ruling held AI training on copyrighted books constitutes fair use, but maintaining a separate “central library” of pirated copies does not. Estimated $3,000 per work. This is the most important settled case to date: it bifurcates the fair-use question — training use is transformative, but the acquisition method matters separately. Meta partial dismissal: court found LLM training to be fair use regardless of whether underlying materials came from legitimate or illegitimate sources — a more expansive fair-use holding than Bartz.
- AI Copyright & Training Data — The Lawsuits That Matter for Developers (2026) (AI Made Tools, 2026) — Current state of the litigation map: 80+ active suits; NY Times case still proceeding (April 2026 status); Bartz settled at $1.5B; Meta partial dismissal granted. The Bartz/Meta divergence on the acquisition-method question means two courts have now reached opposite conclusions on whether training from pirated sources affects the fair-use analysis. This circuit split (if it persists) is the question Thomson Reuters v. ROSS is positioned to address at the appellate level.
Regulation — GAAIA’s Training Data Disclosure Provisions #
- Unpacking the Great American AI Act (DLA Piper, 2026-06) — GAAIA’s Frontier AI Governance title requires large frontier developers (>$500M revenue, models trained on >10²⁶ FLOPs) to submit training data disclosures through Independent Verification Organizations (IVOs). This is a parallel US compliance mechanism to the EU GPAI training data summary Template (August 2 deadline), but structured fundamentally differently: US uses third-party audit organisations rather than a Commission submission platform; US threshold is revenue + compute (not just model capability); US focus is whistleblower-protected disclosure rather than public summary filing. If GAAIA passes, frontier labs will face dual compliance obligations — EU GPAI Template by August 2, 2026, and IVO audits under a new US framework.
Cross-links #
- AI Impact on Society GAAIA’s IVO audit requirement for training data is politically significant in the US context: it creates a private-sector compliance infrastructure (IVOs) rather than a government registry — consistent with the Trump administration’s preference for industry-led governance while still enabling enforcement.
- Open vs Closed AI Ecosystems Bartz’s bifurcated ruling (training = fair use; pirated central library = not) creates a different risk profile for open-weight labs vs. closed labs: open-weight developers typically don’t maintain a central training library for post-deployment queries, whereas closed-source labs with retrieval-augmented systems may maintain searchable document stores that look like the Anthropic “central library” in the Bartz fact pattern.
- Quality signal: The Bartz/Meta acquisition-method divergence is the most legally significant development in the AI copyright space since the Thomson Reuters Delaware ruling. Two courts have now reached opposite conclusions on whether training from pirated sources changes the fair-use analysis — the circuit split that Thomson Reuters v. ROSS will now partially address at the appellate level.
- Emerging pattern: The litigation is bifurcating into two distinct tracks with different risk profiles: (1) training use (converging toward fair use — Bartz, Meta both partial grants); (2) acquisition method (unresolved — Bartz says pirated acquisition is separate liability; Meta says source doesn’t matter). Labs with clean data acquisition but transformative training use are in a better position than labs with mixed acquisition histories.
- Gap: No reporting yet on whether GAAIA’s IVO concept has any existing regulatory models to draw from. If IVOs are a novel institution that requires creation from scratch, the timeline for implementation could extend well beyond any three-year preemption clause.