Midjourney can't force Disney to disclose the prompts it used that didn't generate material in complaint

Disney Enterprises, Inc. v. Midjourney, Inc., No. 2:25-cv-05275-JAK-AJR, 2026 WL 2055488 (C.D. Cal. Jun. 15, 2026)

This is what seems like a significant discovery dispute; I’m not an evidence scholar, but there’s interaction with the fair use analysis that copyright folks should attend to.

Plaintiffs sued Midjourney over its image and video generation “diffusion models,” arguing that they infringed both by copying their characters in training and by distributing copies of those characters in images and videos to Midjourney subscribers.

Midjourney sought to compel plaintiffs to produce: “(1) documents concerning their development, use of, and policies regarding generative artificial intelligence (‘AI’) tools for image and video creation, and (2) the complete set of Midjourney prompts and outputs that Plaintiffs (or their agents) used” for the operative complaint. The court granted the request in part for (1) but found (2) was privileged.

Plaintiffs objected to producing documents about third-party datasets they used in connection with any AI tool and similar documents. They agreed to produce nonprivileged responsive documents “sufficient to show instances in which [Plaintiffs] authorized its employees or contractors to use generative AI to generate images and/or video outputs intended for consumers, featuring the asserted works.” As for the prompts, they agreed to produce nonprivileged responsive documents “sufficient to identify the prompts used to create the image[s] generated by Midjourney shown in the complaint and put at issue in this action, and the side-by-side outputs contemporaneously generated in response to such prompts,” but not the prompts that they didn’t choose to include.

Midjourney argued that these discovery requests were relevant to both its defense of fair use, as well as its equitable defense of unclean hands.

Fair use: Midjourney argued that its requests bore on multiple elements of fair use, including transformativeness. But the court didn’t buy its argument that, “if Plaintiffs are developing and deploying diffusion models (employing the same technology, training techniques, and public data as Midjourney), that is a powerful concession that such models produce something fundamentally new and useful—relevant evidence of the transformative nature of that (and Midjourney’s) technology.” After all, Warhol says “the same copying may be fair when used for one purpose but not another.” [But this would be the same purpose, just a different user, no?] The SDNY has agreed that “the relevant inquiry under the first fair use factor concerns a defendant’s use of a plaintiff’s copyrighted material, not a downstream use of defendant’s allegedly infringing material by a copyright-holder plaintiff.” In re OpenAI, Inc., Copyright Infringement Litig., 800 F. Supp. 3d 602, 608 (S.D.N.Y. 2025). Likewise, New York Times Co. v. Microsoft Corp., 757 F. Supp. 3d 594 (S.D.N.Y. 2024), held that the fair-use factors “do not require a court to examine statements or comments a copyright holder may have made about a defendant’s general industry, whether the copyright holder has used tools in the defendant’s general industry, whether the copyright holder has admitted that other uses of its copyrights may or may not constitute fair use, or whether the copyright holder has entered into business relationships with other entities in the defendant’s industry.”

Public benefits (part of factor four after GvO): But GvO “made clear” that balancing the public benefits against the losses to copyright owners would not “always [be] relevant to the application of fair use,” and focused on the consequences of Google’s copying, not anything that the plaintiff (Oracle) did in its own business.

Nor was this evidence relevant to market harm. Relevant discovery would include the loss to plaintiffs and how the challenged use might “kill demand for the original,” as well as discovery directed to Midjourney concerning the public benefits from the copying.

What about industry custom and practice? Wall Data Inc. v. Los Angeles Cnty. Sheriff’s Dep’t, 447 F.3d 769, 778 (9th Cir. 2006), says that courts “should bear in mind that fair use is appropriate where a reasonable copyright owner would have consented to the use, i.e., where the custom or public policy at the time would have defined the use as reasonable.” But that’s dicta. [This is contrary to what one might have thought was happening two decades ago.] Sure, “industry custom and practice may be relevant to the analysis of fair use in a particular case,” but “courts consistently reject the argument that ‘everybody else is doing it’ as a defense to copyright infringement.” This is really about the (narrower) defense of unclean hands.

But the equitable defense of “unclean hands is recognized only rarely, when the plaintiff’s transgression is of serious proportions and relates directly to the subject mater of the infringement action.” Indeed, “the alleged wrongdoing of the plaintiff does not bar relief unless the defendant can show that he has personally been injured by the plaintiff’s conduct.”

Thus, plaintiffs’ development, use of, and policies regarding generative AI were not relevant to the defense of unclean hands because none of the discovery would establish inequitable conduct that is both directly related to plaintiffs’ claims and injured Midjourney. [Other than through preventing competition with this lawsuit.]

Still, the requested discovery was potentially relevant to establishing: (1) the potential market for or value of the copyrighted work; (2) industry custom and practice; and (3) the defense of unclean hands. But plaintiffs agreed to produce documents sufficient to show their business plans, roadmaps, research reports, other studies, and approvals of their actual or proposed development or training of generative AI intended for consumers, including their contractors’ authorized use of generative AI intended for consumers, as well as documents about some related matters. That was enough. Documents related to the actual or proposed development of generative AI not intended for consumers were not relevant, or not enough to be proportional. However, the court granted the motion to compel plaintiffs to also produce documents sufficient to show plaintiffs’ approval of the use of generative AI to generate images and/or video outputs intended for consumers, featuring the asserted works, which is relevant to the market-harm factor.

Requests for training-related documents had some relevance to establishing industry custom and practice, but had to be narrowed to focus on actual or proposed development of generative AI intended for consumers. “By contrast, Plaintiffs’ training of generative AI tools not intended for consumers would not provide evidence of industry customs and practices that would be relevant to Defendant’s defense of fair use,” or, if relevant, not enough to justify the burden of production. Thus, plaintiffs needed to produce non-privileged responsive documents sufficient to show their “development, training, or contemplated development or training of any generative AI to generate images and/or video outputs intended for consumers, featuring the asserted works, including training datasets, datasources, or model weights.”

Midjourney prompts: Plaintiffs agreed to produce documents sufficient to identify the prompts used to create the images generated by Midjourney shown in the operative complaints, as well as the side-by-side outputs contemporaneously generated in response to such prompts.

Midjourney argued that “withholding prompts and outputs related to images not used in the operative complaints would allow Plaintiffs to artificially inflate the universe of allegedly infringing outputs, distort the damages calculus, or misrepresent their own engineered images as examples of third-party infringement.” But “the volume of prompts and outputs related to images generated for potential use in the operative complaints, but not actually used, is infinitesimal compared to the true scope of this case which involves tens of millions of subscriber prompts associated with Plaintiffs’ copyrighted works.” Given the current statistical sampling protocol that the parties are finalizing, withholding the prompts wouldn’t distort a damages award.

The non-used prompts and outputs were protected work product.  The work-product doctrine protects “from discovery documents and tangible things prepared by a party or his representative in anticipation of litigation.” This was “classic” pre-suit investigation and efforts to prepare the operative complaints. Such unused prompts and outputs were “core” work product because they necessarily reveal counsel’s “mental impressions, conclusions, opinions, or legal theories developed in anticipation of litigation.” This kind of opinion work product “is virtually undiscoverable.” 

Midjourney argued that plaintiffs waived protection by submitting prompts to Midjourney in the first place because Midjourney’s Terms of Service make prompts and outputs public by default and grant Midjourney a license to reproduce, prepare derivative works of, publicly display, publicly perform, sublicense, and distribute their inputs. But “waiver of attorney work-product protection requires more than the disclosure of confidential information, it requires an act inconsistent with the adversary system.” Pre-suit investigation was not inconsistent with the adversary system. Nothing stops Midjourney from conducting its own investigation of prompts submitted by plaintiffs and their counsel, though.

“Accepting Defendant’s view of selective disclosure would virtually eliminate the protection for attorney work product in the context of any court filing because every court filing inherently reflects strategic choices of counsel in what facts to include and even what legal arguments to make.”

Comment: If you believe, as many people seem to, that “how hard was it to get the model to generate an allegedly infringing output?” is a relevant question, then the rulings here make it harder to conduct that inquiry. I’m no evidence expert, but the ruling seems to make probabilistic/guardrails inquiries off-limits to fair use, which seems directionally wrong to me. (Burden-shifting might help—one could say that if the defendant puts in evidence that it tried to make it hard to generate infringing outputs, then the plaintiff has to do more to show that those guardrails didn’t work, which wouldn’t necessarily require disclosure of attorney work product. This is my optimistic reading of the reference to the millions of prompts at issue in this case.)