1 2 3 4 UNITED STATES DISTRICT COURT 5 NORTHERN DISTRICT OF CALIFORNIA 6 7 Case No. 23-cv-03440-EKL (SVK)

8 IN RE GOOGLE GENERATIVE AI ORDER RE DISCOVERY DISPUTE 9 COPYRIGHT LITIGATION Re: Dkt. No. 302 10 11

12 Before the Court is the fourth discovery dispute in a line of disputes relating to Plaintiffs’ 13 efforts to identify the Named Plaintiffs’ and putative class members’ works that were used by 14 Defendant Google LLC (“Google”) in training its AI models. Dkt. 302; see also Dkts. 140, 159, 15 209. Plaintiffs first sought the provision of source code, which this Court denied. Based on 16 Google’s representations that it was the training datasets, not the source code, that would enable 17 such identification, the Court approved (with some modifications) an inspection protocol for 18 training datasets proffered by Google and agreed to by Plaintiffs. See Dkts. 140, 155. The 19 subsequent disputes related to refinement/enforcement of this protocol. See Dkts. 159, 172, 208, 20 221. Plaintiffs now seek to compel Google, by whatever means possible, to “produce evidence 21 sufficient to identify, or simply identify: (1) all Named Plaintiffs’ copyrighted works in the 22 datasets … used by Google to train the Models, and (2) all class member copyrighted works used 23 for that purpose.” Dkt. 302 at 6. The Court finds this matter suitable for resolution without oral 24 argument. Civil L.R. 7-1(b). Having considered the Parties’ submissions, the relevant law and the 25 record in this action, the Court DENIES the request. 26 The Court denies Plaintiffs’ requests for the same reasons it previously denied source code 27 review: Google then represented and now maintains that “[t]he only record and the only means of ] 302 at 7. Plaintiffs have proffered deposition testimony and other evidence suggesting that 2 || Google knows what works are acquired by its Core Data Acquisition Team (“CDA”) and ingested 3 || through web crawling and that Google maintains records of books ingested into certain other data 4 || corpuses. Dkt. 302 at 4-6. But as Google points out, and as this Court has noted previously, none 5 of this information “reveal[s] what survived the steps of preprocessing, filtering, deduplication, 6 || and construction of final training datasets, which are significantly different from what was 7 || included in [a] source corpus.” Dkt. 302 at 9; Dkt. 280-1 at 2-3. Accordingly, consistent with its 8 || prior decisions, the Court is persuaded that only the training datasets will reveal what putative 9 || class works were ultimately used train Google’s AI models. 10 Even if the Court were to order Google to “produce evidence sufficient to identify” the 11 works used in the training datasets, Google’s could only do so by having an expert work backward 12 || from the training datasets. Yet this is exactly what Plaintiffs acknowledge they have already done: 13 “Plaintiffs submitted a methodology that reliably identifies Class Works in Google’s training 14 || data.” Dkt. 302 at 2. Indeed, Plaintiffs’ true contention appears to be that they “should not have 3 || had to incur the time and expense of developing that methodology.” /d. Plaintiffs’ proposed a 16 || relief, however, would not remedy this harm. Rather, it would have Google duplicate work that 2 17 || Plaintiffs admit was already done. “[A] party is not required to create a document where none Z 18 || exists.” Finjan, Inc. v. Juniper Network, Inc., No. 17-cv-05659-WHA (TSH),

, 19 || at *1 (N.D. Cal. July 3, 2019) (cleaned up). Moreover, a defendant is generally not required to 20 || engage in analysis of records where the burden of deriving the answer from those records would 21 be the same for the plaintiff. Fed. R. Civ. P. 33(d). 22 For all of the foregoing reasons, Plaintiffs request is redundant and not proportional to the 23 || needs of the case and is therefore DENIED. 24 25 SO ORDERED. 26 Dated: December 19, 2025 27 Season val SUSAN VAN KEULEN 28 United States Magistrate Judge