The Supreme Court’s newest copyright decision is already being used to argue that an AI video generator trained on major studio content should be treated the same as an internet service provider.
Three weeks ago, the Supreme Court decided a case about an internet service provider that refused to disconnect subscribers who pirated music using its network. Now, the developer of a Chinese AI image and video generator has invoked that opinion to argue it has no obligation to stop producing Darth Vader on demand—never mind that the model can only generate Darth Vader because the developer trained it on unauthorized copies of Star Wars.
Welcome to life after Cox v. Sony Music.
Disney, Universal, and Warner Bros. are suing the operators of Hailuo AI, a Chinese AI image and video generator marketed as “a Hollywood studio in your pocket” that can produce short clips of copyrighted characters from simple text prompts. This is one of two major AI copyright lawsuits the studios have joined forces on. The other targets image generator Midjourney, where the defendant has answered and the case is now in discovery. That makes MiniMax the only studio AI copyright case to trigger a pleading-stage motion.
On Friday, Nanonoble Pte. Ltd. (the Singaporean entity that operates Hailuo AI for U.S. users on behalf of Chinese AI company MiniMax) filed a motion to dismiss the complaint in its entirety (read it here). A separate motion challenges personal jurisdiction over two co-defendants (read that one here). A hearing on both motions is currently set for May 29 before Central District of California Judge Stanley Blumenfeld, Jr. in Los Angeles.
Hailuo AI already employs guardrails to prevent users from generating violent or pornographic content. But the platform has decided not to apply any of those guardrails to copyright, readily producing picture-perfect renditions of Spider-Man, Homer Simpson, Shrek, and Bugs Bunny from simple five-word prompts. None of that matters, MiniMax argues. Knowing your users might infringe and choosing not to stop them is nothing more than inaction. And after Cox Communications, Inc. v. Sony Music Entertainment, inaction isn’t enough for contributory liability.
Before We Get to Cox
The motion raises a host of arguments—the studios’ copyright registrations don’t cover their individual characters, the complaint’s 52 showcase outputs were manufactured by the studios’ own lawyers, the training of Hailuo AI occurred entirely outside the United States, and it’s the users, not MiniMax, who select and generate the allegedly infringing content. Some of MiniMax’s arguments are pretty strong. Others fall flat—starting with character copyrightability.
While the studios’ complaint has focused on individual characters within their copyrighted works, it attaches exhibits mapping those characters to hundreds of registered works—Exhibit A catalogs over a hundred Disney titles alone. The theory is that by reproducing those characters, Hailuo AI necessarily infringes the registered works those characters appear in. There’s no need to have a separate copyright registration for Darth Vader if the model’s output infringes Star Wars. MiniMax also asserts the characters aren’t separately copyrightable under the Ninth Circuit’s character-protection tests, but if the defendants are really trying to argue that Darth Vader and Batman are as generic as Sam Spade, I suspect Judge Blumenfeld will have other ideas.
The reveal that all 52 examples from the studios’ complaint were generated by their own lawyers at Jenner & Block is an interesting rhetorical move, but it doesn’t really shift the needle. The complaint alleges plenty of other organic evidence, including subscriber-posted Instagram videos featuring Spider-Man, the Simpsons, Batman, the Joker, and Superman, as well as third-party posts across Reddit, TikTok, and YouTube.
MiniMax’s extraterritoriality challenge to the training claims is worth taking seriously, and as I noted in an earlier post on Seedance 2.0, it’s a hurdle that Chinese AI developers may increasingly rely on. It’s true that the studios don’t (and likely can’t) allege that scraping or model training occurred in the United States. That said, the complaint’s theory isn’t limited to the act of scraping. It alleges that the copyrighted works are embodied in the model itself, and that model is being distributed to and operated for millions of U.S. users. If the copy lives in the model, the copy goes wherever the model goes.
The volitional conduct argument raises a question that will recur in every generative AI case. The brief asserts that Hailuo AI produces an output only when a subscriber submits a prompt, making the user, not the platform, the one responsible for any infringement. But generative AI complicates that framework. The user types “Darth Vader” and the system, not the user, decides whether and how to render it. That’s one of the reasons the Copyright Office has taken the position that AI-generated outputs aren’t copyrightable, because users lack meaningful control over how the system interprets even highly detailed prompts. There’s a certain tension in arguing that the user has enough control to be the infringer but not enough control to be the author. Courts haven’t resolved that tension yet.
But the real fight is over what happens after Cox.
Cox’s Two Lanes of Liability
In Cox, the Court held that contributory infringement requires intent, and that intent can be shown in only two ways. A copyright owner must prove either that the provider induced the infringement by actively encouraging it through specific acts, or that the provider’s service was “tailored to infringement,” meaning it was not capable of substantial noninfringing uses. The Court was emphatic: “mere knowledge that a service will be used to infringe is insufficient to establish the required intent to infringe.” That holding reversed a billion-dollar verdict premised on the theory that an ISP’s continued service to known infringers was enough. Whether the same framework applies without friction to a system that doesn’t merely transmit copyrighted content but was built from it is a question I’ll return to below.
Tailoring is likely a nonstarter. The complaint’s own screenshots show Hailuo AI generating animated owls working as hospital administrators, panoramic cityscapes at sunset, and women offering flowers to the camera. It’s hard to argue a product is “tailored to infringement” when its homepage features an owl in middle management—unless there’s an obscure title in Disney’s vault that I’m not aware of.
The widest gap between the motion and the complaint is on inducement. MiniMax dismisses the studios’ evidence of its own promotional activity as “scattered references.” But the complaint alleges paid partnership collaborations on Hailuo AI’s official Instagram account featuring Spider-Man, Batman, the Hulk, Elsa, the Minions, Shrek, Superman, and the Joker. It alleges a video on MiniMax’s official YouTube channel titled “Superheroes will heal your soul” that features a parade of copyrighted characters from Deadpool to Wonder Woman.
And one paid partnership video walks viewers through the process: open Hailuo AI, type what you want, wait a moment—”Boom—your world, your way”—and out comes Spider-Man and Supergirl, with a call to action to download the app.
The brief tries to neutralize the “Hollywood studio in your pocket” tagline by comparing it to Sony’s Betamax ads—and the comparison isn’t frivolous. Both taglines promise access to professional-quality content without naming specific copyrighted works. But there’s a difference between a device that records whatever happens to be on television and a service that generates content drawn from copyrighted expression it was trained on. Hailuo AI’s tagline tells on itself.
This may not be Grokster, where the defendants built their entire business around piracy. But at the pleading stage, the court isn’t weighing the evidence—it’s asking whether the complaint plausibly alleges inducement. Paid ad campaigns, step-by-step tutorials, and a tagline invoking Hollywood by name should clear that bar. (Fun fact: MiniMax’s lead counsel, Bobby Schwartz of Quinn Emanuel, was on the other side of the inducement question in Grokster, representing Time Warner.)
The Deeper Problem
But even if the studios get past this motion to dismiss (and they should), the reality is that Cox has made the road ahead considerably harder.
The motion cites Grokster’s footnote 12, reaffirmed in Cox, for the proposition that a court can’t find contributory liability “merely based on a failure to take affirmative steps to prevent infringement.” The defendants read this to mean that Hailuo AI’s selective filtering (guardrails for violence and pornography, but not for copyright) is just another form of inaction.
I’d argue that there’s a meaningful difference between an ISP that, as Justice Thomas put it, “simply provided Internet access” and a company that built a content-filtering system, actively uses it for violence and pornography, rewrites user prompts through its “Prompt Optimizer” before they’re even submitted, and made a deliberate engineering decision to exclude copyright from those tools—all while running paid ad campaigns showcasing exactly the kind of content it chose not to filter. That’s not exactly the absence of action. It’s a series of affirmative choices about which content to police and which to let its own system produce.
Here’s the deeper problem though, and it’s one that will hang over every AI copyright case that follows. While the Cox majority spoke in absolute terms about the limits of contributory infringement, the case it decided involved a fundamentally different technology. Cox never built its network out of Sony’s copyrighted music. The infringing content traveled across Cox’s infrastructure, but the infrastructure itself was neutral. Here, the model internalized the studios’ copyrighted expression during training and can now regenerate it on demand. When MiniMax declines to filter for copyright while filtering for everything else, it isn’t merely declining to police what users do with a neutral tool—it’s leaving wide open the copyrighted expression it already ingested and embedded in its own product. Cox’s network didn’t know what a song sounded like. Hailuo AI knows exactly what Yoda looks like only because it was trained on Yoda.
Generative AI isn’t a pipe. The user contributes five words. The model contributes everything else. Justice Sotomayor’s concurrence in Cox warned that the majority’s two-lane framework “unnecessarily limits secondary liability.” But even under the majority’s own framework, the distinction between a neutral conduit and a system built from copyrighted expression may prove decisive.
The Bottom Line
The complaint’s inducement allegations are specific, detailed, and more than sufficient to survive 12(b)(6). I don’t think MiniMax’s arguments will carry the day at the pleading stage.
The bigger takeaway is what this motion signals for AI copyright litigation after Cox. If the looser paths to contributory liability are off the table—at least for providers whose infrastructure is truly neutral—then plaintiffs will have to plead inducement with unusual specificity. The studios did that here by alleging paid Instagram partnerships, step-by-step instructional ads, a tagline that practically dares the other side to explain it away. An AI company that markets itself as “a creative tool for everyone” and doesn’t run ad campaigns featuring Batman is going to be a much harder inducement case to plead, even if the underlying technology works exactly the same way.
A hearing on MiniMax’s motion to dismiss is currently set for May 29. I’ll keep you posted. In the meantime, as always, I’d love to hear your thoughts. Drop a note in the comments below or @copyrightlately on social media.
View Fullscreen
Credit: Source link