What the Midjourney Discovery Fight Actually Tells Us About AI Copyright Law

Disney, Universal, and Warner Bros. Discovery sued Midjourney last year for training its image generator on their characters without permission. That part of the story is familiar by now. The interesting part is what happened next.

Midjourney asked the court to make the studios turn over information about their own internal AI use. Training datasets, model weights, research reports, the works. Its argument is straightforward: if the studios are using AI tools internally to storyboard, ideate, or market content, trained on data of unclear provenance, that undercuts their claim that unlicensed training is inherently unlawful. A magistrate judge limited the request to "consumer-facing" applications in June. Midjourney is now asking a district judge to overturn that limit, and a hearing is set for August 17.

I don't know how that motion comes out. But the fight itself is more useful than most rulings, because it shows you exactly what matters right now in AI copyright law, and what still doesn't.

Output claims are getting real traction

Set the Midjourney case aside for a second and look at a separate one: several studios sued MiniMax over its Hailuo tool, alleging both direct and secondary infringement. In May, the judge denied MiniMax's motion to dismiss. The reasoning matters. This case is built on outputs, not training. Hailuo will generate a recognizable Spider-Man or Darth Vader off a simple prompt, and that's a much easier claim to win than "you scraped our library at some point."

That's the theory to watch. Proving what a model was trained on is expensive, technical, and often impossible without discovery you'll spend a year fighting for. Proving that a tool spits out your copyrighted character on request is a much shorter path to a courtroom win. If you're a rights holder trying to figure out where you actually have leverage against an AI platform, output infringement is where the current case law is heading, not the training question everyone talks about.

Training-data claims cut both ways

Here's the part worth sitting with if you're on the production side rather than the litigation side. The studios suing Midjourney over training data are themselves using AI internally, the same way most production companies are at this point. Storyboarding tools, marketing generators, editing assistants. Nobody in the industry has fully audited what those tools were trained on, because until recently nobody needed to.

Midjourney's discovery request is a preview of a question that's going to come up more often: what happens when a company that uses AI internally also wants to sue someone else for training on unlicensed material? The studios' lawyers are calling this a "fishing expedition" to distract from Midjourney's own conduct, and maybe it is. But the fact that the argument exists at all tells you the standard being fought over in this case could eventually apply to anyone using AI tools in production, not just AI companies.

If you're running a production company, an animation studio, or an indie shop and you've adopted AI tools for any part of your workflow, the practical takeaway isn't to wait for this case to resolve. It's to start asking your vendors what their tools were trained on and getting that in writing now, before it becomes a discovery request in someone else's lawsuit.

Licensing is becoming the actual answer

While this litigation plays out, the more instructive development might be the deal-making happening alongside it. Disney signed a three-year, roughly billion-dollar deal licensing more than 200 characters from Disney, Marvel, Pixar, and Star Wars for use in OpenAI's Sora video generator. That's not a concession. It's a recognition that licensing, not litigation, is where this settles for the parties with enough leverage to negotiate.

Independent producers and smaller rights holders won't get billion-dollar licensing deals. But the underlying lesson applies at any scale: if AI companies are willing to pay for clean, licensed access to IP, there's a market forming for exactly that. Rights holders who get their chain of title and licensing terms in order now are positioned to participate in that market. Rights holders who don't are just exposed, with no upside to offset the risk.

None of this means the law is settled. Human authorship still has to anchor any copyright claim, courts are still working out what counts as meaningful creative control over AI-assisted output, and the big training-data question hasn't been resolved by anyone. But the shape of where the leverage sits, output claims over training claims, licensing over litigation, is becoming visible. Worth asking yourself where your own AI tools and your own IP actually sit in that picture.