The Motion Picture Association (MPA) and ByteDance just signed a major AI copyright agreement, but it's not what most people think it is. On August 17, 2026, the two organizations established a cooperation framework to protect intellectual property on ByteDance's Seedance video model and Seedream image model, covering platforms including TikTok, CapCut, and Dramina. However, the agreement focuses entirely on controlling what these AI systems produce, not on how they were trained or what data they used. This distinction reveals something crucial about how the entertainment industry actually views AI copyright disputes.

The MPA and ByteDance agreement is what experts call "output-side governance," meaning it sets guardrails on the products AI systems create rather than addressing whether using copyrighted material to train AI models violates copyright law. No financial terms, license fees, or film library authorizations were disclosed. This matters because it contradicts media reports suggesting the deal represents a "truce" or opens up a "copyright library" for AI training. The reality is far more strategic.

"Responsible AI innovation is inseparable from the practical protection of right holders," stated John Rogovine, General Counsel of ByteDance.

John Rogovine, General Counsel of ByteDance

Why Are Studios Avoiding the Training Data Question?

The answer lies in recent court cases and legislative efforts that have created a legal minefield for everyone involved. When you look at what Congress is actually proposing and what courts are actually ruling, a pattern emerges: the input side of AI copyright (whether training on copyrighted material is legal) remains deeply uncertain, while the output side (what AI systems produce) is far easier to control and monitor.

In June 2026, the U.S. House Judiciary Committee held a hearing on AI and intellectual property. Even scholars who are skeptical of AI companies, like Bhamati Viswanathan from Suffolk University Law School, proposed "light-touch regulation" rather than outright bans on AI training. The proposed CLEAR Act (S. 3813) and TRAIN Act focus on one core mechanism: requiring AI companies to disclose their training data sources and allowing copyright holders to request that information through administrative subpoenas. Neither law actually determines whether using copyrighted works for training constitutes infringement.

This uncertainty is intentional. The courts have not settled the question of whether AI training qualifies as "fair use" under copyright law. The Anthropic case (Bartz v. Anthropic) ended with a $1.5 billion settlement in 2026, but the settlement only addressed compensation for "pirated data",it did not establish legal precedent on whether the training itself was lawful. The gap remains open.

What Happened When Studios Sued Midjourney Over Copyright?

The Disney and Universal v. Midjourney case, filed in June 2025, reveals why studios are now choosing output-side agreements over input-side litigation. The studios filed a 110-page complaint documenting how Midjourney generates images of copyrighted characters like Princess Elsa, Buzz Lightyear, and Shrek that are "almost indistinguishable from official artworks." They described Midjourney as a "bottomless pit of plagiarism" and argued the system produces unauthorized copyrighted characters on demand.

But then something unexpected happened. In June 2026, Midjourney counter-sued, demanding that Disney, Universal, and Warner Bros. disclose their own internal AI training datasets, model weights, and board decision-making documents about AI use. Midjourney argued that these studios also use generative AI internally and likely train with unauthorized materials themselves. District Judge A. Joel Richlin partially granted the request, requiring studios to disclose information about their consumer-facing AI applications. Midjourney has appealed to expand the scope further.

This counter-suit changed the entire game. If studios are also training AI systems without authorization, they face the same legal exposure they are suing others for. This explains why the MPA and ByteDance agreement deliberately avoids discussing training data: once a court establishes that "AI companies must fully disclose training data sources," Hollywood itself becomes vulnerable to the same scrutiny.

How Are Studios Actually Deciding What to Protect?

The real insight comes from Disney's own behavior. In December 2025, Disney signed an investment and licensing agreement with OpenAI, authorizing OpenAI to use hundreds of characters from Disney, Pixar, Marvel, and Star Wars franchises for the Sora video generator. Disney sued Midjourney over copyright infringement while simultaneously licensing its content to OpenAI. This reveals that studios' copyright positions are not absolute principles but flexible commercial judgments.

Studios care far less about whether their training data is used in AI systems than about whether AI-generated outputs compete with their products or damage their brands. Litigation over training data is expensive, takes years to resolve, and copyright holders do not necessarily win even when they do sue. Most settlements result in damages and injunctions, not destruction of the AI model itself. It is technically impossible to remove training data from an already-trained AI system.

Steps Studios Are Taking to Protect Their Interests

• Output-Side Monitoring: Studios are negotiating agreements that require AI companies to implement guardrails preventing the generation of copyrighted characters, storylines, and visual styles, focusing on what the AI produces rather than how it was trained.
• Selective Licensing: Rather than fighting all AI training, studios are licensing their content to AI companies willing to pay and implement output protections, creating revenue streams while maintaining control over brand usage.
• Disclosure Demands: Studios are using litigation discovery to force AI companies to reveal training data sources, creating legal risk for competitors while protecting themselves through licensing agreements that provide safe harbor.

The ByteDance-MPA agreement represents the future of AI copyright disputes: not courtroom battles over whether training was legal, but commercial negotiations over what AI systems are allowed to create. This approach protects studios' actual business interests far more effectively than litigation over training data ever could.

" }