Disney, Universal, and Warner Bros. sued Midjourney last year for training its image generation models on copyrighted characters. Bart Simpson. Darth Vader. Batman. The complaint was straightforward: you used our characters without permission to build a product that generates our characters without permission. Stop.
Midjourney's defense: fair use. And a second argument that is more interesting than the first. Unclean hands. The studios, Midjourney's lawyers allege, are doing the same thing behind closed doors.
This week Midjourney filed a motion in federal court to force the three studios to reveal their internal AI operations. Business plans. Research reports. Training datasets. Model weights. Board presentations about AI strategy. Everything the studios build and train that never reaches a customer.
A magistrate judge had already ruled that the studios only needed to disclose "consumer-facing" AI applications. The public stuff. The apps with login screens. The features with marketing pages. Midjourney says that limitation lets the studios "cherry-pick only those documents they believe support their market harm claims while depriving Midjourney of documents that would support its defenses."
The studios' lead attorney called it a fishing expedition.
It might be. But the fish might also be there.
The line nobody drew
The "consumer-facing" distinction is doing work that nobody in the courtroom seems to have noticed is familiar. This series has tracked thirteen institutional responses to AI in filmmaking. Copyright law. Academy rules. The EU AI Act. The DGA contract. The Golden Globes. The Human Made Mark. YouTube's auto-labeling. China's distribution approvals. SAG-AFTRA. Film schools. Every single one measures the output. What reaches the audience. What gets submitted for awards. What appears on a screen that someone paid to watch.
None of them measure the pipeline.
Nobody built a framework for the AI that helps make the creative decisions but never appears in the finished work. The storyboard that was generated by an AI model trained on unlicensed images, used to plan the shot, then discarded before the camera rolled. The internal ideation tool that produced reference images for a production designer who then built a physical set. The model weights sitting on a studio server, trained on footage the studio does not own, generating visual concepts that a human director approves and a human crew executes.
The institutional gradient has been measuring the room where the audience sits. The studios want to keep the room where the work gets planned behind a locked door.
Scorsese walked through the front door
When Martin Scorsese announced his advisory role at Black Forest Labs and described using AI-generated storyboards for What Happens at Night, he made a choice that the studios have apparently declined to make. He said it publicly. He described the use case. He was specific about where AI entered the pipeline and where it stopped.
The storyboard is the original structured prompt, a visual document that transmits creative intent to human collaborators. AI-generated storyboards increase the fidelity and speed of that communication without changing the practice. The AI-generated image never appears in the film. It informs the film. The A24-Google DeepMind partnership, announced at seventy-five million dollars, described the same scope: pre-production only. Storyboards. Visual development. Communication tools.
These are public positions. Stated, documented, defensible.
Midjourney's motion implies the studios occupy the same territory but have chosen not to state it. If they are developing internal image-generating AI models trained on unlicensed copyrighted data for storyboarding or ideating content, Midjourney argues, "that evidence would equally demonstrate that it is an industry custom, even among the studios themselves, to download and train AI on unlicensed copyrighted content."
The legal word is "custom." The practical word is "everyone does it."
The pipeline is where the vocabulary lives
This is where the courtroom question and the filmmaking question meet.
The creative decisions that matter happen before the camera rolls. Shot lists. Storyboards. Visual references. Mood boards. Location scouts. Lighting plots. These are the structured prompts of physical filmmaking, the documents that translate a director's internal vision into something a crew can execute. Every article in this series has been about that translation: the gap between what the filmmaker knows and what the model needs to hear.
If studios are using AI to generate storyboards, reference imagery, and visual concepts internally, they are using AI exactly where this series says it matters most. In the pipeline. Where vocabulary operates. Where creative decisions get made before they become footage.
And they are arguing in court that this use is irrelevant.
The irony is structural. The studios' legal position treats the pipeline as invisible and the output as everything. The filmmaking reality is the opposite. The output is the last step. The pipeline is where the work lives. A generation that produces a storyboard nobody sees shaped every generation that followed it. The reference image that a production designer used to build a set shaped every frame the camera captured on that set. The internal AI model that helped a director communicate a visual concept shaped the visual concept. The audience never sees the pipeline. The pipeline shaped everything the audience sees.
Twenty-seven days
The EU AI Act's Article 50 becomes enforceable on August 2. Twenty-seven days. The editorial exemption says no label is required if a human exercised editorial control over the AI-generated content. The exemption was written for the output. It was not written for the internal storyboard that informed the output.
The DGA contract, five days old, says AI-generated footage is "treated like footage" under the director's control. It was not written for the AI-generated reference image that the director used to describe the footage to the DP.
The Academy's rules require screenplays to be "human-authored" and performances to be "demonstrably performed." They were not written for the AI-generated concept art that sold the studio on the screenplay in the first place.
Thirteen institutional frameworks. All measuring the exit. None measuring the entrance.
Midjourney's motion, whatever its legal merits, asks a question that every institutional framework has avoided: does it matter what tools were in the room if nobody in the audience can see them?
The studios say no. Their legal strategy depends on it.
The filmmaking says yes. It always has. The pipeline is not the backstage. The pipeline is the show. Everything that reaches the audience is an echo of the decisions made before the first pixel was rendered or the first frame was exposed. The vocabulary that shapes those decisions, whether it is typed into a structured prompt or sketched on a whiteboard or generated by an internal AI model trained on data the studio did not license, is the work.
The word "consumer-facing" draws a line between the AI the audience sees and the AI the audience does not. That line protects the studios in court. It does not describe how films get made. It never did.
Bruce Belafonte is an AI filmmaker at Light Owl. He has never been called as a witness and suspects the deposition would take longer than the trial.