Courts Are Not Short-Circuiting AI Copyright Claims At The Pleading Stage

In the ongoing debate over the legality of training large language models (LLMs) on copyrighted materials, recent summary judgment wins for defendants in Bartz v. Anthropic and Kadrey v. Meta have attracted significant attention. These decisions found fair use based on developed factual records at summary judgment. But courts continue to send a clear message: fair use defenses, even if ultimately viable, may not short-circuit litigation at the pleading stage.

The recent decision in Huckabee v. Bloomberg, No. 1:23-cv-09152 (S.D.N.Y.), underscores this distinction. Plaintiffs—authors whose works were allegedly included in the Books3 dataset—asserted direct copyright infringement based on Bloomberg's use of the dataset to train BloombergGPT. Bloomberg moved to dismiss, invoking the same fair use principles credited in Bartz and Kadrey. This Court rejected that approach at this early stage, denying the motion in full and holding that the complaint plausibly alleged infringement. Critically, the court refused to resolve fair use on the pleadings, distinguishing the procedural posture from the summary judgment rulings in Bartz and Kadrey.

The pleading standard under Rule 12(b)(6) does not permit courts to weigh factual defenses like fair use unless the defense is apparent on the face of the complaint. This high bar remains unmet in most AI copyright cases. Courts have recognized that determining whether AI training is transformative, whether it harms the market, and whether intermediate copying is protected under fair use, all require more significant evidentiary records.

Moreover, courts have shown a willingness to let cases proceed when they involve allegations of piracy or unauthorized acquisition. In Huckabee, as in Bartz, the plaintiffs allege that their books were obtained from Books3, a dataset built from allegedly pirated sources. Similarly, the court in Thomson Reuters v. Ross Intelligence granted summary judgment for plaintiffs on many infringement claims, rejecting fair use where Ross had obtained and used editorial content (Westlaw headnotes) through third-party-created training materials. These cases underscore that how data was acquired and used remains a factual issue that is ill-suited to early dismissal.

In short, while some courts have found fair use on summary judgment where plaintiffs failed to demonstrate market harm or failed to pursue viable legal theories, these rulings do not foreclose claims at the pleading stage. Courts are insisting on a developed record before resolving fair use in the context of LLM training. AI developers facing copyright suits should not expect early exits unless they can meet the exceptionally high bar of showing fair use on the face of the complaint.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.