In Re Mosaic LLM Litigation

Background

Five authors sued Databricks and Mosaic ML for alleged copyright infringement involving large language models used to train artificial intelligence. Their initial complaints alleged that Mosaic ML trained its MPT models using datasets containing some of their works. They asserted direct-infringement claims against Mosaic ML and vicarious-liability claims against Databricks, which the opinion identifies as Mosaic ML’s corporate parent.

Databricks released its DBRX models in March 2024. The authors pointed to statements by the defendants’ employees that they contended indicated the DBRX models were trained using the same works. The authors later sought discovery about DBRX, and Magistrate Judge Cisneros ruled that DBRX was relevant to the case as pleaded in the original complaints.

The authors moved for leave to amend their complaint to add a direct-infringement claim against Databricks based on the DBRX models and to update the list of allegedly infringed works. They filed the motion by the amendment deadline, while fact discovery and the deadline for summary-judgment motions were still several months away.

Legal Standard

Because the defendants had answered and did not consent to amendment, Federal Rule of Civil Procedure 15(a)(2) required the authors to obtain the court’s permission. Courts generally freely allow amendments when justice requires, but may deny permission because of undue delay, bad faith, repeated failure to correct deficiencies, unfair prejudice, or futility. The court treated prejudice to the opposing party as the most important factor and noted that delay alone is not enough to deny amendment.

Court’s Analysis

Undue delay. The court found the authors’ fourteen-month delay after the DBRX models’ release to be significant and found that they had not provided a satisfactory explanation. But the court also considered the timing of the case: discovery was still ongoing, several months remained, and the motion was filed within the scheduling order’s time limits. Because delay alone was insufficient, this factor did not require denial.

Bad faith. The court rejected the defendants’ arguments that the delay itself showed bad faith or that the authors had made misleading statements about the proposed amendment. The court found no established history of dilatory conduct. It also concluded that the authors’ descriptions of the amendment were accurate when considered together with their effort to add both the DBRX claim and new allegedly infringed works.

Prejudice. The defendants argued that the amendment would radically alter the case. The court rejected that argument as conclusory. Discovery remained open, the court had already allowed discovery concerning DBRX, and the defendants were on notice of the discovery that would be needed. The court also stated that additional discovery or the burden of defending a new claim, by itself, was not enough to show legally sufficient prejudice.

Futility. The defendants argued that the proposed DBRX allegations were largely conclusory and did not support an inference of copyright infringement. The court acknowledged that the allegations might ultimately be too conclusory to state a claim, but held that this was not one of the rare cases in which the proposed amendment was clearly futile. The court deferred fuller consideration of those merits challenges until after the amended complaint was filed.

Disposition

The court granted the plaintiffs’ motion for leave to amend. The order allowed them to add the direct-infringement claim against Databricks based on the DBRX models and update the list of allegedly infringed works. It did not decide whether the amended allegations establish copyright infringement.