Summary of Key Points
ByteDance has signed a cooperation agreement with the Motion Picture Association (MPA), the industry organization representing the seven major Hollywood studios. However, this agreement is not about purchasing copyrights or addressing the issue of how much Hollywood content is used in AI training processes. Instead, it stipulates that ByteDance’s AI models (such as Seedance for video generation and Seedream for image processing) must include certain “safeguards” when producing content for platforms like TikTok and CapCut. These safeguards prevent the AI models from randomly generating copyrighted materials, such as Disney’s Princess Elsa or Universal’s Shrek. The agreement does not involve any financial transactions or the opening of a shared copyright library, so media reports suggesting a “truce” or an “open access to content libraries” are misinterpretations.
The underlying logic is that Hollywood’s real concern is not the use of their content in AI training, but rather the potential for AI-generated content to directly compete with their businesses (for example, user-created free images replacing licensed merchandise). Moreover, Hollywood itself may also use unauthorized data in its own AI models and does not want to make a big deal out of “input-side infringement” issues. This agreement represents a new approach to AI copyright governance: focus on regulating the “output” (current business activities) first, and then gradually address the “input” (long-term compliance) issues.
I. The Agreement Between ByteDance and MPA: Regulating Output, Not Input
Simply put, the “output side” represents Hollywood’s source of revenue, while the “input side” is a more complex and problematic area.
- Output-side safeguards: For instance, if a user uses CapCut to create a video featuring Spider-Man, ByteDance’s AI should be able to recognize and prevent it or require the user to obtain authorization. This directly protects the commercial value of Hollywood’s IP, such as sales of licensed films, merchandise, and related products.
- Input-side avoidance: The agreement does not specify how much Hollywood content is used in AI training. If Hollywood were to get too involved with this, it could face issues itself, as it may also use unauthorized materials (such as online images or videos) in its own AI processes. Therefore, the agreement represents a practical compromise: ByteDance does not need to resolve the “historical” issues related to training data immediately and can continue operating in the Western market, while Hollywood ensures protection of its core IP interests.
II. The US Legal Approach to AI Copyright
Recent US legislation and jurisprudence do not consider the use of copyrighted content in AI training as an infringement. Instead, they follow a “lighter regulation” approach:
- Legislation: Laws like the CLEAR Act require AI companies to disclose their training data, and the TRAIN Act allow copyright holders to request such disclosure, but these laws do not determine whether the use of copyrighted material in training is illegal.
- Judicial cases: For example, Anthropic settled a lawsuit with a copyright holder for $1.5 billion, but only for the portion involving unauthorized data; the issue of “fair use” (legally permitted free use) remains unresolved. When Disney sued Midjourney, Midjourney argued that Hollywood was also using unauthorized data in its AI models, and the judge supported some of Disney’s disclosure requests. This indicates that Hollywood is reluctant to push input-side issues too far for fear of counterproductive consequences.
In summary, US law currently favors allowing AI to develop while gradually ensuring compliance, with a focus on transparency rather than outright prohibition.
III. Hollywood’s Real Intentions
Hollywood’s attitude towards AI is not one of principled opposition but of prioritizing commercial interests:
- They can sue companies like Midjourney for using their IP without payment or grant licenses to companies like OpenAI (which have paid and implemented safeguards).
- Litigation regarding input-side issues is costly and time-consuming, and even a win may not result in the destruction of the AI model (since it’s technically impossible to remove training data from the model). Therefore, it’s more effective to regulate the “output” directly, as this directly affects box office revenue and merchandise sales.
The MPA’s signing of this agreement represents Hollywood’s collective decision to focus on building safeguards for the output side rather than engaging in a lengthy battle over training data with AI companies.
IV. Implications for Chinese AI Companies
This agreement sends three important signals to Chinese AI companies looking to enter the Western market:
1. Output-side safeguards are essential: Without them, they will likely face lawsuits from Hollywood (as seen with Midjourney), similar to being denied entry without a visa.
2. Input-side issues do not need to be resolved immediately: They represent long-term risks that can be managed later; focusing on regulating the output first is crucial for getting started in the Western market.
3. Competitive advantage lies in negotiation skills: The key is not how many lawsuits a company can avoid but whether it can negotiate agreements with industry organizations like the MPA or obtain licenses from copyright holders (like OpenAI) to gain access to Western markets.
In conclusion, the focus of AI copyright compliance has shifted from determining whether training data is illegal to ensuring that output content protects IP rights. Companies that can grasp this shift will have a competitive advantage in the Western market.
Final Summary
The relationship between AI and copyright holders is not one of mutual exclusion but of cooperation for mutual benefit. Hollywood wants to prevent its IP from being freely used, while AI companies need to be able to continue developing. The balance found between the two parties is to regulate the “output” first and then address the “input” issues. Chinese AI companies seeking to enter the Western market must understand and adapt to these rules.