Generative AI and Copyright: The Seedance 2.0 Controversy

Written By: Brennan Paterson, Lex Tecnica Law Clerk

Though still in its infancy, the use of generative artificial intelligence, or AI, has begun to reshape creative industries at a remarkable pace. Tools capable of generating images, music, text, and even full-length video from simple prompts are being used every day to create both new and sometimes strikingly derivative content.  As these technologies evolve, they raise interesting legal questions, particularly around copyright law.

The recent dispute involving ByteDance’s generative AI video model, Seedance 2.0, provides a useful lens for understanding how copyright law applies to AI training, AI output, and how creators can protect their work.

Seedance 2.0 is a video production AI model developed by ByteDance, the parent company of TikTok. Seedance 2.0 can generate realistic videos based on simple text prompts, enabling users to create entire cinematic scenes involving your favorite real-world actors, fictional characters, and even the unique visual style of your favorite cinematographer. Recently, a video generated by Seedance 2.0 made the rounds showing Tom Cruise and Brad Pitt fighting on a rooftop. Hollywood studios and industry groups were displeased, to say the least, with the Motion Picture Association (MPA), Paramount, Disney, and others sending cease-and-desist letters accusing ByteDance of large-scale copyright infringement.

Citing the Brad Pitt/Tom Cruise video among other examples, the MPA and the other studios argued that such outputs strongly suggested that copyrighted content had been used to train Seedance 2.0 without authorization. The MPA characterized the model as engaging in systematic copyright infringement, asserting that the problem was not merely user misuse but potentially inherent in the technology itself. While ByteDance responded that it respects intellectual property rights and is working to strengthen safeguards, the controversy brings to light two novel legal issues: the use of copyrighted works in AI training datasets, and the legal status of AI outputs.

AI Training

To understand these issues, it helps to briefly examine how generative AI systems are built. Most generative models are trained using vast datasets scraped from the internet. These datasets may include images, books, movies, music, and so on, many of which are protected by copyright. During training, the model analyzes the datasets, “learning” the patterns and relationships that allow it to produce content based on user prompts.  The legal question is whether this process itself infringes copyright.

Typically, copyright owners have exclusive rights to reproduce their works, distribute copies, prepare derivative works, and publicly perform or display the works. Training an AI model arguably involves copying copyrighted material, at least temporarily, in order to analyze it. If that copying occurs without permission, a copyright owner may argue that it violates their reproduction rights. Some AI developers have argued that this activity qualifies as fair use, particularly when the training model does not precisely reproduce the original works.

However, this question has not been settled in the courts, and several lawsuits are currently testing the boundaries of copyright law in this AI context.

AI Output

Separate from the issue of whether inclusion of copyrighted materials in AI training datasets infringes on those copyrights is the issue of whether AI infringes on copyright if the output is substantially similar to protected works. For example, if an AI system generates a scene that clearly reproduces distinctive characters, story elements, or visual designs from a film franchise (such as Brad Pitt and Tom Cruise fighting on a rooftop), that output could be considered an unauthorized derivative work.

While this is a threatening situation for copyright holders and creators, it’s also a potential risk for AI developers.  However, developers can reduce their exposure in several ways. First, and perhaps most obviously, developers can license training datasets from copyright owners. Some AI developers are already pursuing agreements with publishers, music labels, and other media companies to obtain legal access to large datasets. Since the copyright owners are compensated under this model, licensing is more expensive than simply gathering materials for free from the internet.  On the other hand, such cooperation protects developers from lawsuits and provides high quality training data, including data that remains unavailable on the internet.

Another approach is to rely on public-domain materials, user-submitted data with clear permissions, or synthetically generated datasets.  While limiting datasets to such available materials may slow development and hinder the scope of AI learning, it significantly reduces the risk of expensive infringement claims. 

AI developers can also implement safeguards designed to prevent users from generating infringing content in the first place. ByteDance and other developers have indicated that they are beginning to employ such safeguards, such as blocking user prompts referencing copyrighted characters, filtering recognizable visual styles, and limiting or prohibiting generation involving real people or actors. 

How Creators and Copyright Owners Can Protect Themselves

Creators and copyright owners concerned about AI appropriation can also take steps to protect their work. Perhaps the most effective tool is to register their copyrights.  Formal copyright registration strengthens the owner’s ability to enforce their rights and pursue damages in court. A formal copyright can only be protected, of course, if the copyright owner finds out that their work has been used without permission. In order to stay informed, owners can utilize tools that track image, video, and music re-use across the internet.  These tools have been used for many years to discover still images and music being used without permission, and they become more effective every day in discovering content which may not be the copyrighted work itself, but only derivative of it.  Or, as noted above, copyright owners can take a “if you can’t beat em, join em” approach by licensing their content to developers, thereby profiting from the use of the content that may end up used without permission anyway.

The Future of AI and Copyright

The Seedance 2.0 dispute won’t be the last confrontation between AI developers and copyright owners, and as various lawsuits move through the court system new legal doctrines may be developed, or even new legislation.  However, two things seem certain: AI isn’t going anywhere, and compliance with copyright law isn’t optional. Thus, the pace of AI innovation may depend on whether developers and copyright owners can find workable ways to cooperate rather than litigate. Understanding these issues today may help avoid costly legal battles tomorrow.

This article has been reviewed and approved for legal accuracy by Ashlyn Hauber, Esq. It is intended for informational purposes only and does not constitute legal advice.