Sora's "Overt Strategy": Solving the Deadlock of AI Copyright with a Revenue-Sharing Model
On September 30th, OpenAI released its most powerful video generation model to date, Sora 2.0. Its amazing effects shocked the tech circle. The number of downloads exceeded one million in less than five days after its launch, which was even faster than ChatGPT's speed when there were no restrictions. However, it also quickly ignited the long - standing AI copyright crisis.
For a while, a large number of users on the Internet used Sora to generate AI fan - made videos of well - known IPs (such as Pikachu and Mario). The relevant video content has been widely spread and used.
The three major Hollywood talent agencies joined forces to fight back against Sora. They uniformly withdrew their client resources and demanded the establishment of an authorization and revenue - sharing mechanism. Giants like Disney and industry associations also put pressure simultaneously, demanding that OpenAI assume liability for infringement.
Behind this phenomenon is the "Opt - out" mechanism initially adopted by Sora. This mechanism allows the generation of copyrighted content by default, unless the copyright holder actively requests its removal. This policy completely shifts the responsibility of safeguarding rights to the copyright holder and is accused of potentially causing systematic infringement.
On the cusp of the storm, on October 4th, OpenAI changed its "Opt - out" model. Its CEO, Sam Altman, announced a key strategic shift: Sora will adopt a new "Opt - in" policy. In the future, it will introduce a revenue - sharing mechanism for IP, sharing the platform's revenue with copyright holders who authorize the use of their characters. This move is like an unexpected new move in the tense confrontation between AI companies and copyright giants.
It marks that the leading players in the industry are starting to try to shift the conflict from court battles to business cooperation, attempting to find a new way for ecological co - construction for the elephant in the room of AI copyright. However, can this new model fundamentally solve the controversial AI video copyright disputes?
This article will conduct a legal analysis of Sora's new copyright revenue - sharing policy from the perspective of an intellectual property lawyer to explore the underlying pattern of interest distribution.
Image: Pikachu and Mario images generated by AI
The legal dilemmas behind Sora 2
To understand the significance of Sora's new policy, we must first clarify the current legal dilemmas faced by AI companies.
Behind the Sora copyright storm is a core legal issue that has remained unresolved since the birth of generative AI: Does using a large number of copyrighted works to train commercial AI models without authorization constitute copyright infringement? This question can be further broken down into the legality of training data at the "input end" and the risk of infringement of generating similar content at the "output end."
First, the original sin of the training data at the "input end." The powerful capabilities of AI models are built on learning from a large amount of data, and these data inevitably contain a large number of copyrighted pictures, videos, and texts. AI companies generally claim that their scraping and use of these data for model training fall under fair use, but this defense faces great challenges in judicial practice.
In recent years, major copyright holders have launched multiple rounds of lawsuits. From Getty Images, a photo library, suing Stable Diffusion, to the six major Hollywood studios jointly suing Midjourney, and then to Warner Bros. filing a copyright infringement lawsuit over the AI - generated Superman image, the focus of all these cases points to the act of using copyrighted content for model training without authorization. As of now, all parties are still waiting for the court to make an authoritative judgment on the application of the fair use principle in the AI era.
Meanwhile, some enterprises have begun to seek solutions outside of litigation.
Not long ago, in a class - action lawsuit brought by several authors against Anthropic, Anthropic, the company behind the AI model Claude, agreed to pay a settlement of up to $1.5 billion to the authors and publishers, which also reflects from the side that AI companies still lack a solid legal footing on the fundamental issue of training data.
Second, there is the risk of infringement of the generated content at the "output end." Even if the threshold of training data is crossed, the specific content generated by AI may still constitute an infringement of existing works. In the early days, Sora 2 was able to generate highly realistic videos of Pikachu or Tanjiro (the protagonist of the manga "Demon Slayer"). This can easily be regarded as an unauthorized reproduction of the character image or the creation of an infringing derivative work, which is a typical copyright infringement act.
Image: Naruto, One Piece, and Demon Slayer generated by Sora 2
For example, in 2024, the Hangzhou Internet Court ruled that an AI platform was liable for contributory infringement for allowing users to upload Ultraman pictures to train LoRA models and generate similar infringing content. In the same year, the Guangzhou Internet Court ruled that the AI painting function of another AI platform could stably generate pictures substantially similar to the Ultraman image, constituting direct infringement.
The Opt - out policy shifts the responsibility of identifying and removing infringing content to the copyright holder. This approach is quite passive legally. Once copyright giants (such as Nintendo and Disney) initiate lawsuits, the platform will bear huge legal and compensation risks.
Therefore, Sora 2's Opt - out policy exposes it to both of the above - mentioned legal risks. OpenAI's rapid shift is less of a business model innovation and more of a practical consideration to avoid legal risks. Facing the large number of second - creation videos of "Pokémon" or "Demon Slayer" generated by thousands of users, continuing with a lenient policy is tantamount to acquiescence or even encouragement of infringement, which is unsustainable in the current judicial environment.
Image: A screenshot of a video starring Sam Altman generated by Sora 2
The copyright revenue - sharing model of Sora 2 and its pros and cons
Facing legal uncertainties and huge commercial pressures, OpenAI did not choose to engage in a long - drawn - out battle in court but played a smart business card. The core of Sora's new plan is to shift from blocking to guiding, trying to turn copyright holders from litigation opponents into partners.
The plan is roughly as follows: Define user - generated content as interactive fan - made creations; Promise to provide copyright holders with more refined IP control. Copyright holders can independently decide whether and how their characters are used by Sora (for example, allowing the generation of daily scenarios but prohibiting violent content), and can even completely prohibit the use of certain core IPs; Propose an IP revenue - sharing mechanism. Copyright holders who allow users to use their characters can receive a share of the platform's relevant revenue.
This policy can create new incremental revenue for copyright holders and also assist them in IP promotion and value exploration.
The IP revenue - sharing blueprint depicted by OpenAI is similar to YouTube's copyright revenue - sharing model, and the prospects are very attractive. Imagine a creator making a video of "Sun Wukong fighting Iron Man." The platform can automatically identify the IP ownership and distribute the generated revenue to the corresponding copyright holders in proportion. This will greatly stimulate the enthusiasm for secondary creation and open up new monetization channels for copyright holders, creating a win - win situation for multiple parties. However, there are still problems and risks that cannot be ignored in the design of this business model:
1. The business model is not yet clear. OpenAI admits that the specific revenue - sharing model still needs to be determined through repeated trials. How to negotiate authorizations with thousands of IP holders around the world, how to design a fair and reasonable pricing system for different IPs, and how to ensure the sharing ratio, billing method, and transparency will all be the focus and difficulty of future negotiations. If a one - by - one authorization negotiation model is adopted, it may be more beneficial to leading IP giants with strong bargaining power. How to effectively protect the rights and interests of a large number of small and medium - sized creators remains an unknown.
2. The complexity of technological implementation. The AI generation process has a certain black - box characteristic. How to accurately and transparently track and identify each copyright element (characters, scenarios, music, etc.) in the generated content and conduct accurate value attribution and revenue distribution is a huge technological challenge. YouTube's Content ID system took more than a decade to gradually improve, and it will only be more complicated for Sora to build a similar copyright identification module. If Sora fails to fulfill its management obligation when the copyright holder clearly requests the removal of relevant IP characters, it may still need to bear legal responsibility.
3. It treats the symptoms rather than the root cause. Sora's IP revenue - sharing mechanism mainly solves the problem of infringement at the output end. It still does not give a direct answer to the fundamental legal question of "whether it is legal to use copyrighted data for initial model training." This is the core crux faced by the industry, that is, how to deal with the original sin of training data. Currently, legislators and scholars are still exploring various solutions such as collective licensing models, market - based trading models, and legislative exemption models, and no effective consensus has been reached. Therefore, at present, Sora's plan is essentially a business detour. It is very rational and practical in decision - making, but once unfavorable legislative regulations or judicial precedents appear in the future, the compliance foundation of the entire business model will still be shaken.
The essence and future of AI copyright disputes
From the widespread lawsuits filed by IP giants such as Disney and Warner Bros. to Sora's copyright revenue - sharing policy, the essence of AI copyright disputes is not so much a matter of legal right or wrong as an issue of interest distribution caused by technological changes.
Generative AI creates new and huge value by using existing human knowledge achievements, and the core of the dispute lies in: Who should cut this new cake, and how should it be distributed?
The reason why Sora's plan is of reference significance is that it returns to the essence of the problem and the original intention of intellectual property, encouraging the birth of more creative IP content by motivating creators to make profits. This idea of shifting from one - time authorization to revenue - sharing based on usage may be more suitable for the characteristics of low marginal costs of content generation and personalized user needs in the AI era.
However, to make this system really work, in addition to perfecting the details of the above - mentioned business model and technological supervision, deeper - level problems such as the legality of data sources, standardization of value assessment, and updating of the legal framework also need to be solved to provide clearer and more predictable guidance for the links of AI training, generation, and use.
In summary, the new copyright rules designed by OpenAI for Sora are not a "silver bullet" to solve all problems, but they indicate that the AI industry will enter the paid - licensing stage from the stage of wild growth. It indicates that in the face of the subversive challenges brought by technology, perhaps what can solve the problems is not just stricter laws but also more ingenious institutional designs that can achieve a win - win situation for multiple parties. This may be the next - level thinking that we should focus on in the AI wave.
This article is from the WeChat official account "Tencent Technology." The author is Li Yunkai, and the editor is Guo Xiaojing. It is published by 36Kr with authorization.