Training-data lawsuit news: week of July 20, 2026

Anthropic's $1.5B settlement got final approval

The single biggest event of the week: Judge Araceli Martínez-Olguín granted final approval on July 20, 2026 to Anthropic’s settlement with authors over training Claude on pirated books — the largest known US copyright settlement, at roughly $3,000 per work. The order expressly sets no binding precedent.

The settlement followed Judge William Alsup’s June 2025 ruling that training on lawfully acquired, purchased-and-scanned books was fair use, while downloading and retaining a pirated library was not. That split — acquisition method decides the outcome, not the training itself — is the thread running through nearly every other case this week touches.

“$1.5 billion — about $3,000 per book — and expressly no binding precedent.”— Order granting final approval, N.D. Cal., July 20, 2026

Publishers filed a new class action against Google over Gemini

On July 14, 2026, a new publisher class action landed against Google, alleging Gemini was trained on copyrighted works belonging to Hachette, Cengage, Elsevier, author Scott Turow, and the group S.C.R.I.B.E. That plaintiff list matters as much as the filing date: it pulls education and scientific publishers into the exposed content classes, alongside the trade-book and news plaintiffs who have dominated this wave of litigation so far.

It is Google’s latest addition to a growing docket that already includes the consolidated In re Google Generative AI Copyright Litigation covering visual artists and author Jill Leovy. The Gemini suit is a separate action, but it extends the same basic question — what was in the training set, and was it licensed — into new content categories.

Suno heads toward a summary-judgment hearing

The record labels’ case against AI music generator Suno was headed for a July 2026 summary-judgment hearing in the District of Massachusetts, before Chief Judge F. Dennis Saylor IV. Audio-fingerprinting evidence in the case indicated millions of copyrighted recordings inside Suno’s training data — turning "what is in the training set" from an allegation into forensic evidence.

That fingerprinting technique is the detail worth remembering past this week. It is the same diligence question any buyer should be asking an audio-data vendor directly, and it shows courts now have a technical method to answer it independent of what either side claims in a brief.

Still pending: the Third Circuit in Thomson Reuters v. Ross

Oral argument in Thomson Reuters v. Ross Intelligence was heard June 11, 2026 — the first appellate test of fair use for AI training anywhere in the federal system. The panel pressed both sides on transformativeness and market harm. As of this week, the decision is still pending.

This is the single biggest catalyst on the calendar for how US law treats training on a rival’s licensed or compiled content, and it will bind more broadly than the February 2025 district-court ruling it reviews — which was itself limited to a non-generative tool. Nothing this week changed that timeline, but nothing has diminished its importance either.

NYT v. OpenAI grinds through discovery

As of mid-2026, the New York Times’ case against OpenAI remained in discovery with no trial date set. OpenAI was ordered to produce roughly 20 million ChatGPT conversation logs, an order affirmed January 5, 2026, and in April 2026 was ordered to re-produce a corporate designee the court found unprepared for deposition.

No merits ruling exists in this case. That is worth repeating plainly this week, given how often headlines elsewhere in AI training litigation get read as verdicts before any court has actually decided anything.