ByteDance-MPA signs AI copyright governance framework: output-side governance, "re-discussion" on the input side

On August 17, 2026, ByteDance signed a Memorandum of Understanding (MOU) with the Motion Picture Association (MPA) to establish a "cooperation framework" for intellectual property protection on platforms including ByteDance's Seedance video model and Seedream image model. MPA Chairman Charles Rivkin stated that the two sides will "implement meaningful guardrails", while John Rogovine, General Counsel of ByteDance, noted that "responsible AI innovation is inseparable from the practical protection of right holders".

It is worth noting that this is a "governance framework agreement on the output side", not an authorization contract for training data (input side). The official MPA release clearly states that the purpose of the agreement is to "maintain strong guardrails on Seedance and Seedream", covering user-facing product matrices including TikTok, CapCut, Dramina and others. No financial terms, license fees, or scope of film library authorization have been disclosed in the publicly available information about the agreement. Therefore, claims of a "truce" or "open copyright library" reported by the media are both misinterpretations.

This subtle difference is critical, as it reveals a more authentic commercial demand and even the future direction in AI copyright governance: output-side guardrails rather than input-side blocks more directly represent the core interests of film and television companies, and also help AI companies adopt effective protection measures while developing their businesses. This point can be supported by and draw lessons from the recent Disney v. Midjourney case, and is consistent with the overall trend of current AI copyright governance in the United States.

I. Legislative Path of the U.S. Congress: Transparency Rather Than Prohibition

On June 30, 2026, the U.S. House Judiciary Committee held a hearing themed "40-Year-Old Midlife Crisis", which conducted multiple rounds of expert demonstrations focusing on intellectual property in the internet sector. In her testimony, Bhamati Viswanathan, a professor at Suffolk University Law School, raised systematic doubts about AI training data issues: she not only questioned the broad interpretation of "fair use" made by Judge Alsup in the Anthropic case, but also emphasized that "excessively high licensing costs" is by no means a valid defense.

However, even scholars like Viswanathan who hold a critical attitude towards AI companies proposed a solution of "light-touch regulation" (a light hand of regulation) — to incentivize the formation of a licensing market rather than impose mandatory intervention. This line of thinking corresponds to the recent legislative path of the U.S. Congress around AI training data copyright.

The proposed CLEAR Act (S. 3813) only sets one core obligation for AI companies: mandatory disclosure; while the core mechanism of the TRAIN Act is to "allow copyright owners to obtain administrative subpoenas". On the one hand, the two acts require AI companies to make their training data "transparent", and on the other hand, grant copyright holders the right to know to request disclosure from AI companies. However, neither of them attempts to resolve or determine whether the use of copyrighted works for training in generative AI models constitutes copyright infringement.

The substantive legality of training data use still seems to need to be determined through judicial judgments on a case-by-case basis.

II. Judicial Practice: Compensation, But Also "Backlash"

The Anthropic case (Bartz v. Anthropic) ended with a $1.5 billion settlement. It seemingly promotes the industry to evolve towards an "authorization framework", but in fact, the final compensation is only the economic compensation for the part of "pirated data". Although the interpretation of "fair use" has not become a precedent, it has left a huge gap. The only parties that are most satisfied with the outcome of the entire case are probably the plaintiff's lawyers and Anthropic itself.

The Disney and Universal v. Midjourney case reveals a more complex industrial game. In June 2025, Disney and Universal jointly sued Midjourney. The 110-page complaint is almost entirely dedicated to detailed side-by-side comparisons of "user prompts vs. Midjourney-generated images", demonstrating how the model generates characters such as Princess Elsa, Buzz Lightyear, and Shrek, and that these images are "almost indistinguishable from official artworks". The production companies described Midjourney as a "bottomless pit of plagiarism" and "copyright free-riding". The complaint emphasizes that Midjourney outputs unauthorized copyrighted characters on demand. This is also one of the largest legal battles over generative artificial intelligence and intellectual property to date.

However, the most noteworthy development in this case so far is a turnaround that occurred in June 2026: Midjourney counter-sued to demand that Disney, Universal, and Warner Bros. disclose their own internally used AI training datasets, model weights, and board of directors' AI decision-making documents, on the grounds that these companies also extensively use generative AI internally and are very likely to train with unauthorized materials as well.

District Judge A. Joel Richlin has partially granted Midjourney's request, requiring the studios to disclose information related to consumer-facing AI applications — the studios' own AI practices are at least somewhat relevant to the issues in this case, while rejecting the broad disclosure requirements for internal tools, training datasets, and model weights. Midjourney has appealed to the district court to expand the scope of disclosure.

The power of this evidence disclosure order lies in that it redirects the spear of "input-side infringement" back to the plaintiffs. If the film studios are also training AI and using data without copyright authorization, it will first increase Midjourney's chance of winning this case, and more importantly, this litigation strategy will go beyond the scope of this case and increase the risk of the film studios being sued by other copyright owners.

This explains why the MOU between MPA and ByteDance deliberately avoids the training data issue: once the rule that "AI companies must fully disclose the sources of training data" is established in any individual case, Hollywood itself will face the same review.

III. Copyright Demands of Film Companies: Not Principles, But Commercial Judgments

From the above legislative and judicial trends, we can distill the real "aspiration" of film companies: whether the input side constitutes copyright infringement is not an absolute, truthful stance for them, but a judgment that can be flexibly adjusted according to commercial interests.

In December 2025, Disney signed an investment and licensing agreement with OpenAI, authorizing OpenAI to use hundreds of characters from Disney, Pixar, Marvel, Star Wars and other franchises for the Sora video generator. This shows that Disney's attitude towards training data is not principled: it can sue Midjourney, and it can also license OpenAI, everything depends on the price the other party is willing to pay and the guardrails it provides.

A large number of previous cases also show that litigation on the input side is not only long in cycle, but also copyright holders do not necessarily have an advantage under the defense of "fair use". Even if they win the case, the mainstream remedy is "damages + injunction" rather than "destruction of the model", and it is almost technically infeasible to make the training data "spit out" from the AI model.

For production companies, what really affects their commercial interests is not whether their training data is used for AI training, but whether the product-level output of the AI model affects their commercial interests. Therefore, the MOU between ByteDance and MPA is a pragmatic strategy of "locking core commercial interests at the lowest cost".

It should be noted here that as an industry organization covering seven Hollywood giants, the MOU signed with ByteDance represents the collective attitude and industry consensus of the entire Hollywood on AI output-side governance — the scope of this consensus is even wider than the plaintiffs in the Midjourney case (Disney system + Universal system), because it also includes the positions of members such as Warner Bros., Paramount, Sony, Netflix, and Amazon MGM. However, the MOU cannot replace the independent positions of each member company in specific litigations and specific commercial licenses.

IV. Practical Significance for Chinese AI Enterprises

Based on the above analysis, this MOU implies three specific legal meanings for Chinese AI enterprises — especially those planning to enter Western markets:

First, output-side guardrails are the "visa" to enter the market, with almost no room for bargaining;

Second, the "original sin" of input-side training data is a long-term variable rather than a current liability;

Third, the real competitive barrier will be "licensing negotiation capability" rather than "compliance defense capability".

The significance of the ByteDance-MPA agreement is that it once again proves that the relationship between copyright holders and AI companies is not a simple zero-sum game. Hollywood did not "unconditionally surrender" to accept AI's use of its IP, and technology companies did not withdraw from the market — the two sides found a third path: neither excluding AI from the production system, nor accepting the logic of "use first, then negotiate rights as long as the technology is advanced".

However, different from the narrative of "great trial of training data" some time ago, this MOU further reveals the real rhythm of AI copyright governance: what is immediate, urgent and engineerable at present is the "output side", which is governed immediately through guardrail agreements, which is the core commercial demand of film companies; while the "input side" is long-term, gradual, and negotiable, which will be "gradually resolved" through transparency mechanisms and market-oriented licensing. Moreover, on this point, film companies themselves also face the risk of "backlash".

AI copyright compliance has entered a new stage: input-side compliance and output-side compliance are still running in parallel logically, but there has been a tilt in commercial demands — what film companies want is not a principled declaration of "training data must be cleared of rights", but the commercial result that "IP cannot be arbitrarily generated". Whoever can take the lead in understanding this and bring this understanding into negotiation or litigation games will occupy a high ground in the next round of competition.

This article is from the WeChat public account "Internet Law Review", written by Zhang Ying, and authorized for release by 36Kr.