In Re Mosaic LLM Litigation

Opinion

In Re Mosaic LLM Litigation. Master Case No. 24-cv-01451-CRB

.

ORDER GRANTING LEAVE TO AMEND

Plaintiffs are five authors who have sued Defendants Databricks and Mosaic ML for alleged copyright infringement arising from Defendants’ large language models, or LLMs, which are used to train artificial intelligence. Plaintiffs initially alleged that Mosaic ML trained some of its LLMs, the MPT models, using a dataset that contained some of their works. Now they seek leave to amend their complaint (1) to add a direct infringement claim against Databricks based on Databrick’s new LLMs, the DBRX models, and (2) to update their list of infringed works. The Court finds this matter suitable for resolution without argument pursuant to Civil Local Rule 7-1(b) and GRANTS Plaintiffs’ motion. In their initial complaints, filed in March and May 2024, Plaintiffs contended that Mosaic ML infringed on their copyrighted works by training their MPT large language models on datasets that included their works. O’Nan Compl. (dkt. 1) ¶ 23; Makkai Compl. (No. 24-cv-2653 dkt. 1) ¶ 22. They therefore brought claims of direct infringement against Mosaic ML. O’Nan Compl. ¶¶ 36–42; Makkai Compl. ¶¶ 35–42. Plaintiffs also brought vicarious liability claims against Databricks, the corporate parent of Mosaic ML. O’Nan In March 2024—several weeks after Plaintiffs O’Nan, Nazemian, and Keene filed their complaint and over a month before Plaintiffs Makkai and Reynolds filed theirs— Databricks released new LLMs, the DBRX models. Proposed Am. Compl. (dkt. 124-2) ¶ 39; Opp. (dkt. 126) at 3. Around that time, Defendants’ employees made several statements that Plaintiffs contend indicate that Defendants trained the DBRX models using the same infringed works. See Proposed Am. Compl. ¶¶ 40–44. Starting in March 2025, Plaintiffs started seeking discovery into the DBRX models, which Defendants opposed. Wakefield Decl. (dkt. 126-1) ¶ 6; Defs.’ Disc. Brief (dkt. 122). Magistrate Judge Cisneros has since ruled that DBRX is relevant to the case as laid out in Plaintiffs’ original complaints. See Disc. Order (dkt. 123) 1 Plaintiffs now move to amend their complaint to add a direct infringement claim against Databricks based on the DBRX models and to update the list of works that they allege that Defendants infringed. See Proposed Am. Compl. ¶¶ 49–56; Proposed Am. Ex. A (dkt. 124-4). Plaintiffs filed their motion for leave on the deadline to do so; on that date, four months remained in fact discovery, and eight months remained before the deadline for summary judgment motions. See Sched. Order (dkt. 111) at 1. Because Defendants have filed an answer in this case and have not consented to amendment, Plaintiffs may amend their complaint only with leave of the Court. Fed. R. Civ. P. 15(a)(2). Though courts “freely give leave when justice so requires,” id., factors such as “undue delay, bad faith or dilatory motive on the part of the movant, repeated failure to cure deficiencies by amendments previously allowed, undue prejudice to the opposing party by virtue of the amendment, [and] futility of amendment” justify denial of leave to amend. Foman v. Davis, 371 U.S. 178, 182 (1962). Of these factors, “the resulting prejudice to the opposing party” is the “crucial factor,” Howey v. United States, 481 F.2d 1187, 1190 (9th Cir. 1973), while “delay alone does not provide sufficient grounds for denying leave to amend.” United States v. Pend Oreille Pub. Util. Dist. No. 1, 926 F.2d 1502, 1511 (9th Cir. 1991) (cleaned up). Defendants contend that four factors weigh against granting leave to amend: undue delay, bad faith, prejudice, and futility. Though the Court acknowledges the substantial delay at issue here, the Court does not find that any other factor warrants denial of leave to amend and accordingly grants Plaintiffs’ motion. Undue Delay. Plaintiffs moved for leave to amend their complaint fourteen months after Databricks released the DBRX models and concomitantly made statements that could have put Plaintiffs on notice that the DBRX models might have been trained on the same allegedly infringing datasets that the MPT models were trained on. That delay is without a doubt significant. For example, in AmerisourceBergen Corp. v. Dialysist West, Inc., the Ninth Circuit affirmed a district court’s denial of leave to amend when the plaintiff had delayed amendment by a similar period of fifteen months. 465 F.3d 946, 951–54 (9th Cir. 2006). And Plaintiffs have not provided a satisfactory explanation for their delay. Their suggestions that the delay was exacerbated by Defendants’ resistance to their discovery efforts into the DBRX models glosses over the fact that those discovery efforts began only in March 2025, and their concern about piecemeal amendments, see Reply (dkt. 128) at 2– 4, while fair enough in the abstract, does not justify over a year of delay in this case. Yet any delay must also be placed in context. The parties are still in the midst of discovery, with several months left to go. See Sched. Order at 1. That differentiates this case from others in which courts have held that undue delay warrants denial of leave to amend. See Texaco Inc. v. Ponsoldt, 939 F.2d 794, 799 (9th Cir. 1991) (denial of leave to amend proper when plaintiffs sought to amend eight months after the court’s summary judgment ruling); Johnson v. Hewlett-Packard Co., 809 F. Supp. 2d 1114, 1120–21 (N.D. Cal. 2011) (denying leave to file fourth amended complaint after summary judgment). And though late, Plaintiffs’ motion for leave to amend fell within the time limits that the In any case, delay alone is insufficient to deny leave to amend, see Pend Oreille, 926 F.2d at 1511, so the Court turns to consider the remaining Foman factors. Bad Faith. Defendants raise two arguments as to why Plaintiffs’ motion for leave to amend should be taken in bad faith. Neither favors denial of leave to amend, though. First, Defendants suggest that Plaintiffs’ delay itself is evidence of bad faith. Opp. at 8 (“Courts find bad faith sufficient to deny leave to amend where the moving party engages in dilatory tactics, such as waiting months to move to amend a complaint.”). But the case they cite, Cabrera v. Residential Credit Solutions Inc., involved a plaintiff who had a demonstrated “history of dilatory tactics.” No. CV 14-9212 DMG, 2015 WL 13916231, at *7 (C.D. Cal. Sept. 10, 2015) (citation omitted). They have not established such a history of delay on Plaintiffs’ part, and delay alone generally does not give rise to an inference of bad faith. See Naranjo v. Bank of Am. Nat’l Ass’n, No. 14-cv-2748-LHK, 2015 WL 913031, at *6–7 (N.D. Cal. Feb. 27, 2015). Second, Defendants argue that Plaintiffs made “misleading representations” in their moving papers as to the basis for their amendment. Opp. at 8. The Court is not convinced by Defendants’ assertion. Plaintiffs seek to make two substantive amendments to their complaint: (1) the addition of a direct infringement claim against Databricks based on the DBRX models and (2) the addition of new allegedly infringing works. See Proposed Am. Compl. ¶¶ 49–56; Proposed Am. Ex. A. The language that Defendants point to as evincing bad faith (for example, Plaintiffs’ statement that they want to update their complaint “to better reflect the evidence to date” and their repeated references to “Defendants’ LLMs” generally) might be misleading if Plaintiffs had only sought the first amendment. But Plaintiffs’ descriptions are accurate given the full scope of their proposed amended complaint. Moreover, Plaintiffs’ statements only carry a risk of confusion if the Court rejects their allegations that Databricks trained the DBRX models on Plaintiffs’ works; if the Court accepts those allegations, even just at the pleading stage, Plaintiffs’ statements are vague at worst. This is certainly not grounds to find bad faith. Prejudice. Defendants bear the burden of showing that they will be prejudiced by amendment. DCD Programs, Ltd. v. Leighton, 833 F.2d 183, 187 (9th Cir. 1987). They assert that they will be prejudiced by Plaintiffs’ proposed amendment because it “will radically alter this case.” Opp. at 9. And to be sure, if an amendment significantly changes the “nature of the litigation,” it may result in prejudice. Morongo Band of Mission Indians v. Rose, 893 F.2d 1074, 1079 (9th Cir. 1990). On the other hand, neither “the burden of having to defend a new claim” nor “the prospect of additional discovery needed by the non-moving party” constitutes a sufficient showing of prejudice to deny amendment. PNY Techs., Inc. v. SanDisk Corp., No. 11-cv-4689-WHO, 2014 WL 294855, at *4 (N.D. Cal. Jan. 27, 2014) (citations omitted). There is especially low risk of prejudice when discovery is still open. Id. at *5; see also Lockheed Martin Corp. v. Network Sols., Inc., 194 F.3d 980, 986 (9th Cir. 1999) (“A need to reopen discovery and therefore delay the proceedings supports a district court’s finding of prejudice.”). Defendants’ conclusory assertions of prejudice do not demonstrate that they will, in fact, be prejudiced by amendment. Discovery is still open and will remain so for several months. The Court has already granted Plaintiffs’ request for discovery into the DBRX models, so Defendants are on notice as to what discovery will be necessary. See Disc. Order. And to the extent that Plaintiffs are correct that there is overlap between the MPT and DBRX models, any discovery into the DBRX models will not represent a radical shift in the nature or scope of discovery. Defendants have not established prejudice, so this factor does not weigh against amendment. Futility. Futility alone can be grounds to deny amendment. Ahlmeyer v. Nev. Sys. of Higher Educ., 555 F.3d 1051, 1055 (9th Cir. 2009). But as a general rule, courts “defer consideration of challenges to the merits of a proposed amended pleading until after leave to amend is granted and the amended pleading is filed.” Clarke v. Upton, 703 F. Supp. 2d 1037, 1043 (E.D. Cal. 2010). Accordingly, the bar for an amendment to be so futile that it warrants denial of leave to amend is “if there is no set of facts that can be proved under the Defendants argue that the allegations as to the DBRX models in Plaintiffs’ proposed amended complaint are largely conclusory and that any allegations of fact do not support an inference of copyright infringement. Opp. at 11–14. Indeed, Defendants contrast Plaintiffs’ proposed allegations about the DBRX models (for example, the allegation that “Defendants have further made copies of Plaintiffs’ Works for use to train the DBRX models,” Proposed Am. Compl. ¶ 49) with their more specific allegations about the MPT models (for example, that the MPT models were trained using specific datasets, id. ¶¶ 27– 37). It may be that Plaintiffs’ DBRX-related allegations are too conclusory to state a claim, but this is not one of the rare cases where it the proposed amendment is so clearly futile that leave to amend should be denied. Defendants can (and no doubt will) brief these issues more fully once Plaintiffs file their amended complaint. For the foregoing reasons, the Court GRANTS Plaintiffs’ motion. Dated: June 25, 2025 CHARLES R. BREYER United States District Judge