Adobe Wins Dismissal of Nemotron Copyright Claim

The Same Day Suno Lost, Adobe Won — on the Words "Upon Information and Belief"

Judge Corley dismissed the Nemotron half of an author's copyright claim as merely possible rather than plausible. The Books3 half was not challenged and stands.

Hours after a Massachusetts court kept a music-training case alive, Judge Jacqueline Scott Corley in the Northern District of California went the other way in Tanzer v. Adobe Inc., No. 26-cv-04712. Following oral argument held on 20 August, she granted Adobe's motion to dismiss the portion of the author's single cause of action relating to NVIDIA's Nemotron product, with leave to amend; any amended complaint must be filed by 2 October 2026.

What failed, precisely

The complaint alleged, "upon information and belief," that to develop the Nemotron language models Adobe copied a corpus of 1,881,250 English-language books obtained from Anna's Archive, described in the pleading as the most active and notorious shadow library. Corley held that this was "a bare conclusion unsupported by factual allegations that make it plausible rather than merely possible," citing the Ninth Circuit's Century Aluminum standard, and that the plaintiff had therefore failed to allege either Article III standing or direct infringement on the Nemotron theory.

What the common framing gets wrong

Three corrections. Adobe did not win the case — it moved against one portion of one cause of action. The allegations concerning Adobe's own SlimLM series, trained on the Books3 dataset whose creator has publicly confirmed it represents "all of Bibliotik," were not challenged and remain live. Second, dismissal with leave to amend is an invitation, not an ending: the plaintiff has until 2 October to plead facts rather than inference. Third, the ruling is not about whether training on shadow libraries is lawful. It is about whether this plaintiff pleaded enough to make it plausible that Adobe was the one who took the books from that library.

The pairing

Two federal judges, the same day, on adjacent questions, reaching opposite results. Saylor held that a plaintiff need not produce the infringing output; Corley held that a plaintiff must do more than assert on information and belief where the training data came from. Both are pleading-stage rulings, and read together they mark where the line currently sits: you may infer the output, but you must substantiate the input.

Why It Matters

AI copyright plaintiffs rarely have direct evidence of what went into a training set, so they plead it on information and belief. A court refusing to accept that shifts the burden to a stage before discovery can supply the answer.