How is AI Copyright Litigation constructed?

Procedure and scope

Counts US copyright suits filed against AI developers covering three distinct claim types that are usually reported as one: TRAINING-DATA claims (ingesting works to train a model), OUTPUT claims (the model reproducing protected expression), and LICENSING disputes. They raise different legal questions and are unlikely to resolve the same way. CAVEAT ON PRECISION: the mid-point is a FLOOR, not a count. The cited source says cases "more than doubled" during 2025 to "over 70", so 70 is a lower bound and the true figure is higher. Counts also differ between trackers depending on whether consolidated matters, appeals and state-court filings are counted separately. Treat the shape of this curve as reliable and the individual values as approximate. THE RULING THAT MATTERS: Bartz v. Anthropic (Judge Alsup, June 2025) held that training on copyrighted works can be fair use, but that storing pirated copies of those works is not. That bifurcation is the framework the remaining US cases are navigating, and it is why the September 2025 Anthropic settlement with authors — $1.5bn, among the largest copyright settlements in US history — turned on the piracy limb rather than the training limb. This is an observational index of litigation volume. It is not legal advice and takes no position on the merits of any case. Re-verified 2026-08-11.