Disney v. Midjourney: The Copyright Case Aimed at What AI Generates

Disney and Universal are not only suing Midjourney over how its model was trained. They are suing over what it produces when a subscriber types "Darth Vader" or "Minions". That output-side theory carries statutory damages of up to $150,000 per work for wilful infringement, and the Disney Midjourney lawsuit is now 16 months into a schedule with no trial date. If your product embeds image generation, this is the AI copyright lawsuit whose theory reaches you.

The short answer: Disney and Universal v. Midjourney is a pending federal case in Los Angeles, consolidated with a Warner Bros. Discovery suit. It alleges both that Midjourney copied studio films to train its model and that the model's images are themselves infringing copies. No court has ruled on liability, and no trial date has been set.

I have written this for founders and product directors at software companies that put an image generator inside their product. The output-side theory is the part of the case that maps onto your roadmap. The reader I have in mind is deciding which model to call, what to filter, and what to promise customers in a contract.

What the Disney Midjourney lawsuit actually claims

The complaint was filed on 11 June 2025 in the US District Court for the Central District of California, case number 2:25-cv-05275. It runs to roughly 110 pages, much of it screenshots of generated images placed next to the studio originals.

The rhetoric is unusually blunt for a federal pleading. The studios call Midjourney "the quintessential copyright free-rider and a bottomless pit of plagiarism," a line Variety quoted from the filing on the day it landed.

The parties, the court and the consolidation

Seven plaintiffs sit on the Disney side: Disney Enterprises, Marvel Characters, MVL Film Finance, Lucasfilm, Twentieth Century Fox Film, Universal City Studios Productions and DreamWorks Animation. Judge John A. Kronstadt presides, with Magistrate Judge A. Joel Richlin handling discovery.

Warner Bros. Discovery filed its own suit on 4 September 2025, case number 2:25-cv-08376. The court related it to the Disney case the next day and consolidated the two in November 2025, according to AI Lawsuit Tracker's docket summary. One defendant now faces three studios over Superman, Batman and Scooby-Doo as well as Darth Vader, Elsa and Shrek.

The output-side acts the studios plead

The complaint names a long list of characters the service will render on request. Variety's summary includes Deadpool, Wolverine, Iron Man, Spider-Man, the Hulk, Darth Vader, Yoda, Elsa, Olaf, Buzz Lightyear, Lightning McQueen, Shrek and the Minions.

The pleaded acts are the ones copyright law reserves to the owner. Midjourney, the studios say, reproduces their works and publicly displays the results on its own site. It then distributes them to subscribers who pay between $10 and $120 a month "so that they can view and download copies and derivatives" of the characters. That last phrase is the output theory in one sentence.

The complaint also alleges the studios asked Midjourney to stop before suing, and that Midjourney kept releasing new model versions instead. It points to filters Midjourney already runs for violence and nudity as evidence that blocking characters was technically possible. Haldanes partner Anthony Leung, writing on Mondaq in June 2025, read the pleading as placing "a much stronger emphasis on output-side infringement rather than input-side."

Why the output-side theory differs from AI training cases

Most AI copyright litigation asks one question: was it lawful to copy the works into the training set? The Disney case asks that too. It also asks a second, older question, which is whether the thing the machine hands back to a paying customer is an unlicensed copy.

The two questions have different evidence, different defences and different consequences for a vendor. I think the second one matters more to the B2B software market, and the coverage has mostly treated it as a sideshow.

Training claims look backwards, output claims look forwards

A training claim is about a dataset assembled in the past. Courts so far have been more willing to call training transformative than to forgive how a corpus was acquired. That pattern is covered in the analysis of the copyright cases that could reprice training data. Even a loss on training is a one-time liability attached to a historical act.

An output claim attaches to every generation the service runs today and tomorrow. If an image of Darth Vader is an infringing copy, the infringement repeats with each request. That is why the injunction request is the dangerous part. It is aimed at a running product, not a closed chapter.

Famous characters make the output comparison easy

Output claims usually struggle with similarity. A generated landscape rarely matches one specific photograph closely enough to prove copying. Characters change that, because a character is protectable as a recurring set of visual traits, and the studios own the most recognisable set of traits in the market.

So the studios did not have to find one training image and one output that match pixel for pixel. They typed a name and printed what came back. My view is that this is the cheapest infringement evidence any AI plaintiff has yet assembled, and that is exactly why other rights holders are watching it.

The remedies sought, and why the injunction outweighs the money

The studios ask for the full menu: actual damages or Midjourney's profits, statutory damages in the alternative, an accounting, and preliminary and permanent injunctive relief. They also ask for costs and fees, which the Copyright Act allows for registered works.

Remedies as reported from the complaint by Variety and Mondaq in June 2025. The statutory ranges are the statute's own. Which works qualify depends on registration dates, which the public reporting does not resolve.

The per-work arithmetic is smaller than the headlines

Statutory damages are counted per work infringed, not per image generated. A thousand pictures of one character are still one work for this purpose. That keeps the ceiling far below the numbers people quote on social media.

The injunction is the remedy I would model first. A court order requiring character filters would not bankrupt Midjourney. It would, however, write the first judicial specification of what an image service must block, and every vendor in the category would inherit that specification whether they were a party or not.

Midjourney's defence: fair use, user direction and the studios' own AI

Midjourney answered on 6 August 2025. It denied liability, demanded a jury, and pleaded seven affirmative defences, including fair use, licence, unclean hands, de minimis copying and reliance on notice procedures.

Fair use for training, user direction for outputs

On training, Midjourney argues that extracting statistical patterns from images is transformative fair use. On outputs, it argues that subscribers direct what gets made, that its terms forbid infringement, and that it cannot tell automatically whether an image infringes. Parody, fan art and critique, it says, are legitimate uses of the same prompts.

"This case isn't about infringement," the company said in its filing, as Cartoon Brew reported on 7 August 2025. "It's about whether media giants can stop others from using even elements of their works in transformative, expressive, and innovative ways." It also accused the studios of skipping takedown procedures and failing to name specific infringing images.

The fight over how the studios use AI themselves

Midjourney's sharpest move is a hypocrisy argument. "Plaintiffs cannot have it both ways," the answer says, pointing to studio staff and vendors who use Midjourney for concept art and to Disney's own public enthusiasm for generative tools.

Discovery turned that into a procedural war. In mid-June 2026, Magistrate Judge Richlin largely denied Midjourney's motion to compel evidence of the studios' internal AI development. Midjourney asked Judge Kronstadt to review that ruling on 29 June. It argued that studio training on third-party works would be "powerful evidence that the relevant industry considers such conduct to be fair use," per The Art Newspaper. Studio counsel called it a "fishing expedition."

The commercial backdrop gives the argument some texture. On 11 December 2025, Disney agreed to invest $1 billion in OpenAI and license more than 200 characters for Sora and ChatGPT image generation under a 3-year deal, Axios reported. Disney is not against AI images of Elsa. It is against unlicensed ones, which is a pricing position as much as a legal one, and one that sits alongside the deals counted in the tally of AI content licensing deals.

Midjourney lawsuit status: the docket through October 2026

Read the dates below as the shape of the case, not a forecast. Judge Kronstadt's scheduling order of 14 November 2025 set the discovery and motion deadlines and left trial unscheduled until pending motions are decided.

Dates from the 14 November 2025 scheduling order as summarised by AI Lawsuit Tracker, plus the October stipulation reported by a docket-tracking blog. The court docket itself could not be retrieved for this post, so treat each date as secondary until checked against PACER.

The Cox motion and what is left of secondary liability

The complaint pleads in the alternative that Midjourney is secondarily liable for what its users make. That theory took a hit on 25 March 2026, when the Supreme Court decided Cox Communications v. Sony Music unanimously. The Court held that a service provider is contributorily liable only if it intended its service to be used for infringement, through inducement or a service tailored to infringe, as Authors Alliance summarised the opinion.

Midjourney moved for judgment on the pleadings on the contributory claims on 22 July 2026, according to a docket-tracking blog that posted the motion. The studios' August opposition conceded the material contribution theory and kept only inducement. The direct infringement claims are untouched, which means the output theory survives whatever happens to this motion.

Mapping AI output infringement exposure onto your product

You are probably not Midjourney. You call a model through an API, wrap it in a design or marketing workflow, and sell seats. The question is which parts of the studios' theory survive that distance.

The table below is my own mapping of the pleaded acts onto common integration patterns. It is a framework for a conversation with counsel, not legal advice, and no court has applied the theory to any of these patterns.

Author's analysis based on the acts pleaded in the complaint. The final row is the one place where a vendor's own conduct, rather than a user's, most resembles what the Cox standard still punishes.

Two levers recur across every row. The first is filtering. The studios built their wilfulness case partly on the claim that Midjourney already filters violence and nudity, so a vendor that filters some categories and not famous characters invites the same comparison. Warner Bros. Discovery's complaint went further, alleging that Midjourney had blocked video generation from infringing images and then lifted that block shortly before the suit, as Reuters reported in September 2025.

The second lever is contract. Many model providers now offer output indemnities with conditions attached, and the conditions are where the protection lives. The differences between them are laid out in the comparison of model provider terms, and the pattern for code tools is in the AI coding assistant indemnity comparison. I would not ship an image feature to enterprise customers without reading the image-specific indemnity clause line by line.

If you sell into regulated or brand-sensitive buyers, they will ask. Procurement teams that read about this case will add a question about character and logo filtering to the security questionnaire. Having a written answer before the question arrives is cheap, and not having one costs you the deal cycle.

Where this analysis is weakest

Three things cut against the argument above, and you should weigh them before reorganising a roadmap around a pending case.

The primary record is thinner than a docket summary implies

The court docket and the complaint PDF blocked automated retrieval during research for this post. Every date and quotation here comes from press reports and docket-tracking sites that read the filings. Two of those sources disagree on details, including whether the mid-June discovery motion was denied or largely denied. The revenue figure is a party allegation the complaint itself attributes to reports, not a number Midjourney has disclosed.

The case may also never produce the ruling this post anticipates. Judge Kronstadt ordered private mediation with an August 2026 deadline, and no settlement had been reported by 21 October. A confidential settlement would leave the output theory exactly where it is now: pleaded, plausible and untested. Midjourney's user-direction defence may also be stronger than it looks, because Cox narrowed exactly the kind of liability that attaches to a tool used by others. My position remains that the direct output claim is the one to plan around, but a vendor who bets the other way is not being reckless.

Frequently asked questions

Where to start before the November deadlines

Run a 30-minute test this week. Ask your product's image feature for 10 famous characters by name, record what comes back, and send the results to whoever owns your provider contract. That test is precisely what the studios did, and it costs you nothing to see your exposure before a rights holder sees it first.

Then put a reminder on 24 November to check the docket for summary judgment filings. Those briefs will contain the first full argument over whether outputs infringe, and they will tell you more about your filtering roadmap than this post can.

Related reading