District Court Denies Plaintiffs Motion to Compel Google to Identify Copyrighted Works in Training Data
Bot Mutiny |
U.S. District Judge Eumi K. Lee has upheld a ruling that Google does not have to provide a specific list of copyrighted works used to train its generative AI models.
Google won't be forced to generate a list of every copyrighted work it used to train its generative AI models. On August 24, 2026, U.S. District Judge Eumi K. Lee issued an order in the Northern District of California denying a motion for relief filed by plaintiffs in the ongoing case, In re Google Generative AI Copyright Litigation. The decision, documented in a filing first reported by CourtListener, marks a significant moment in the discovery battle between copyright holders and the tech giant. Plaintiffs had moved to compel Google to produce documents sufficient to identify, or simply to identify, all copyrighted works from the named plaintiffs and class members within the datasets used to train the models at issue. Google fought the request. The company argued that no such list exists and that there's no way for it to generate one. According to Google's opposition, the only record and the only means of determining what materials were used for training is the training data itself. ## The Redundancy Ruling
Magistrate Judge van Keulen previously presided over what the court described as an exhaustive and contentious discovery process. She concluded that only the training datasets would reveal which putative class works were used. Because the plaintiffs already have access to sample datasets, Judge van Keulen ruled that the request would force Google to duplicate work that the plaintiffs' own expert had already started. The plaintiffs challenged this, filing a motion for relief under ECF No. 328. They argued that Judge van Keulen relied solely on Google's claims that "excavating" the data was the only way to identify works. They also claimed the order improperly limited discovery into how Google copies works during the training process. ## Clear Error Standard
Judge Lee rejected both arguments. Under federal law, a district court can only grant relief from a magistrate judge's pretrial order if it's "clearly erroneous or contrary to law." Judge Lee found that Judge van Keulen didn't just take Google at its word. The magistrate judge had evaluated deposition testimony and other evidence before concluding that Google couldn't produce a list more efficiently than the plaintiffs' own methodology. The court's stance is that the request was redundant and not proportional to the needs of the case. By denying the motion, the court is essentially telling the plaintiffs that if they want to know exactly what was scraped, they'll have to find it in the datasets themselves. This ruling keeps the burden of identification on the plaintiffs. It accepts Google's technical claim that its internal records don't include a neat inventory of every copyrighted item fed into the machine. For the class members, this means the path to proving infringement remains tied to the raw, massive datasets rather than a convenient list provided by the defendant.