On Nov. 13, 2025, Judge McMahon issued her ruling denying Cohere’s partial motion to dismiss Advance Local Media’s Complaint. Summarized here:
Today, I will take a closer look at the copyright part of the court’s ruling.
A. Cohere attempted to dismiss the “substitutive summaries” theory of infringement
Judge McMahon rejected Cohere’s argument on “substitutive summaries“: “Cohere contends that many of Command’s summaries do not copy any protectable expression because Command “incorporates the abstracted facts into new and original sentences.” Dkt. No. 50, at 16. Even where the summaries do copy some of Publishers’ works, Command argues, they do so only minimally, rendering them non-infringing.”
The Judge relied on Exhibit B, a compilation of 75 examples of alleged infringement by Cohere’s AI model: (1) 50 examples were regurgitated news articles verbatim and (2) 25 examples were a “mix of verbatim copying and close paraphrasing.”
Judge McMahon pointed to the example on p. 21:
As Judge McMahon concluded, “Publishers have adequately alleged that Command’s outputs are quantitatively and qualitatively similar.” “Cohere’s contention that the only similarities to Publishers’ works are Command’s use of the same facts is belied by Publishers’ allegations and examples showing that Command’s outputs directly copy and paste entire paragraphs of Publishers’ articles verbatim.”
My take: Can’t disagree with that. Indeed, without reading Cohere’s motion to dismiss, the argument here seems almost frivolous in light of Exhibit B. Perhaps Cohere was speaking about non-regurgitated outputs for its motion on substitutive summaries. But the Judge did not interpret it that way.
It’s interesting to see the output theory of infringement now getting scrutiny by courts. Judge Stein also recently allowed an output theory of infringement against OpenAI based on plot summaries:
B. Direct infringement of Cohere users for secondary liability based on evidence of Plaintiffs’ investigators’ use
The more interesting of the opinion is Judge McMahon’s analysis of the secondary liability claim. To prove secondary liability of Cohere, whether vicarious or contributory infringement, the plaintiffs must prove direct infringement by users of Cohere’s AI.
For that, Judge McMahon held that it was sufficient to rely on the plaintiffs’ investigator’s own use of Cohere AI for the alleged examples of direct infringement at the pleading stage.
Judge McMahon reasoned: “In New York Times, the court rejected the precise argument that Cohere now makes, dismissing defendants’ reliance on Bender because at the pleading stage, it was sufficient that plaintiffs provided more than 100 pages of examples and alleged “widely publicized” instances of end-user copyright infringement. 777 F. Supp. 3d at 307. I reach the same conclusion here. Publishers’ allegations and 75 examples of allegedly infringing outputs spanning 125 pages are sufficient to “raise a reasonable expectation” that discovery will reveal additional evidence of third-party infringement. See Bell At!. Corp. v. Twombly, 550 U.S. 544, 556 (2007). Contrary to Cohere’s assertions, New York Times does not require Publishers to allege “widely publicized” instances of alleged third-party infringement in order to survive a motion to dismiss.”
My take: Hard to quibble at the pleading stage. But query whether, on summary judgment, the plaintiffs must come up with examples of real-world infringement by users or evidence that infringing outputs occurred “in real life” or “in the wild.” This very issue came up in the UK case, Getty Images v. Stability AI. And the Justice there did recognize that what I call (i) investigatory evidence of infringement by Plaintiffs’ investigators has less probative value than (ii) evidence of real world infringement by users in normal use of an AI model.
At least as the pleading stage, Judge McMahon didn’t find it relevant whether “Publishers’ investigators had to breach its Terms of Service in order to receive the allegedly infringing outputs, which it claims no bonafide user would ever do.”
However, I think that issue is highly relevant at the summary judgment stage and any trial. It goes to whether the AI company has implemented reasonably effective guardrails to minimize infringing outputs. If normal use of AI does not lead to infringing outputs and the only use that does requires paying a skilled investigator to deploy adversarial techniques in violation of terms of use, the probative value of the evidence of alleged investigatory evidence plummets.
C. Is Cohere’s knowledge of specific infringement required for contributory infringement?
Judge McMahon rejected Cohere’s argument that Plaintiffs must allege Cohere had actual knowledge of specific acts of infringement to support a contributory infringement claim, or willful blindness.
Here’s what Cohere argued:
Judge McMahon rejected the Ninth Circuit’s approach and even embraced the 2d Circuit lower standard of “know or have reason to know” of direct infringement.
My take: I thought the “have reason to know” standard was erroneous after the Supreme Court’s decision in Global-Tech Appliances v. SEB S.A., 563 U.S. 754, 766 (2011). But apparently it’s still the law in the 2d Circuit.
This issue came up in the BMG Rights Management v. Cox Communications case, a part of which is before the Supreme Court now. There, the 4th Circuit held:
It is well-established that one mental state slightly less demanding than actual knowledge—willful blindness—can establish the requisite intent for contributory copyright infringement. This is so because the law recognizes willful blindness as equivalent to actual knowledge. See Global–Tech Appliances, Inc. v. SEB S.A., 563 U.S. 754, 766, 131 S.Ct. 2060, 179 L.Ed.2d 1167 (2011) (“[P]ersons who know enough to blind themselves to direct proof of critical facts in effect have actual knowledge of those facts.”); Aimster, 334 F.3d at 650 (“Willful blindness is knowledge, in copyright law … as it is in the law generally.”).
Whether other mental states—such as negligence (where a defendant “should have known” of infringement)—can suffice to prove contributory copyright infringement presents a more difficult question.4 The notion that contributory liability could be imposed based on something less than actual knowledge, or its equivalent, willful blindness, is not entirely without support. See Aimster, 334 F.3d at 650 (“[I]n copyright law … indeed it may be enough that the defendant should have known of the direct infringement ….”) Nonetheless, we believe for several reasons, that, as Cox contends, negligence does not suffice to prove contributory infringement; rather, at least willful blindness is required.
The Second Circuit approach would seem to create a circuit split, increases the chances that the Supreme Court will eventually have to review this issue.
D. The Active Inducement Claim
Judge McMahon credited the Plaintiffs’ allegations of Cohere touting and marketing its AI to conduct “a web search to access the latest news about trends and competitors[.]”
The Judge ruled: “Publishers have plausibly pleaded that Cohere takes affirmative steps to foster infringement by advertising Command as a tool to access news in order to solicit customers. The Complaint provides numerous examples of Cohere promoting Command’s news reporting function. For example, in a September 2023 marketing pitch, Cohere advertised Command’s ability to conduct “a web search to access the latest news about trends and competitors[.]” Compl. ,r 67. Similarly, in September 2024, Cohere announced the Cohere AI app, touting its ability to “keep you up to date with the latest news.” Id. ,r 68. Publishers have further alleged that Cohere’s Chat interface suggests to consumers that its models provide summaries ofthe latest news articles and, in the free online demo of Command, even “prepopulates the interface with a request to summarize recent technology news, inviting prospective customers to use the models to access news stories.” Id. ,r 69. These allegations are sufficient at the pleading stage.”
The Judge rejected Cohere’s view of what “latest news” meant: “According to Cohere, Publishers’ allegations are insufficient to show that Cohere took “active steps” to foster infringement because “merely touting products as tools to stay up to date with ‘the latest news’ is not a culpable inducement to infringe.” Dkt. No. 50, at 14. Once again, this misstates Publishers’ theory, which is that Cohere intentionally programmed Command to generate and deliver copies ofPublishers’ works to third parties.”
My take: Viewed in light of the non-moving party, the Plaintiffs, the court’s conclusion sounds right. If “latest news” means copyrighted news articles and not just facts, then the marketing could be interpreted in the way the Plaintiffs argue.