From Prompt to Video: The IP Puzzle Behind Google’s Veo 3

Reproduction Right in AI Video Generation

In the initial stage, which entails data collection and curation, developers inevitably make multiple copies of copyrighted works, thereby engaging the right of reproduction. ^{This occurs as they download, transfer, reformat, or modify content, often starting with material sourced from publicly available locations. According to Professors Pamela Samuelson, Christopher Jon Sprigman, and Matthew Sag, training generative AI models requires extensive web scraping, which creates local copies of millions, or even billions, of copyrighted works. This view is widely accepted, as most commentators do not dispute that copying during the acquisition and curation process implicates the reproduction right. Although portions of the data may be discarded after training, reports indicate that some developers retain the training datasets for potential future use.}

Additionally, the training process also invokes the reproduction right. To facilitate training, developers download and store vast datasets on high-performance systems, creating further copies of copyrighted material. ^{As the model processes these datasets, works or substantial portions are repeatedly and temporarily copied, and these copies may persist long enough to constitute infringement depending on the model and technical setup. Further, if refining model weights embeds copyrighted works, distributing the model could also lead to infringement, even by third parties.}

Further, RAG reproduces copyrighted works by copying material into internal databases or pulling it from external sources, which involves unauthorized use of protected content. ^{When this retrieved material is combined with a user’s prompt to generate output, it may infringe the reproduction right if the result closely mirrors the original, and depending on the nature of the content, could also implicate derivative works as well as public display or performance rights.}

Copyright Ownership in AI generated Videos

Copyright ownership of AI-generated works hinges on whether the creation was produced entirely by an AI system. ^{If the work lacks any human authorship, it’s not eligible for copyright protection and remains in the public domain. This is consistent with most current legal frameworks, such as Kenya’s Copyright Act, 2001, which limits protection to natural and legal persons. In practice, however, identifying the rightful author is more complex due to the involvement of multiple actors, including the software developer, those who trained the system, and the users who provided the prompts that led to the creation of the content. From a user’s perspective, authorship depends on their level of involvement in guiding the AI. If a user provides specific and detailed prompts to shape the final output and realise their creative vision, they could be considered the author.}

Moreover, if an AI tool uses pre-existing copyrighted material, whether provided by the user or sourced from publicly available content, the original rights holder may assert a claim over the generated output and retain all beneficial rights. ^{If the instruction text incorporates copyrighted literary works, the resulting videos may be considered derivative works. This is because they draw upon the narrative and themes of the original material and thus require permission from the copyright owner. Conversely, if the instruction text is original and the AI-generated video demonstrates some creativity, the text itself may also be protected by copyright. From the standpoint of developers, the outputs of AI tools are not typically attributed to them as authors, because they may own the software but do not have creative control over what users generate.}

The Dilemma of Publicity Rights in AI Video Generation

The International Trademark Association (INTA) defines the right of publicity as an intellectual property right that protects individuals against the unauthorized commercial use of their name, likeness or other identifying elements such as a nickname, pseudonym, voice, signature or photograph. In other jurisdictions, rights akin to the right of publicity are often recognized as “personality rights,” “rights of persona,” or similar terms, with their source and scope varying accordingly. This can be illustrated by the Kenyan case of Jessicar Clarise Wanjiru v Davinci Aesthetics & Reconstruction Centre & 2 Others where the court noted that personality rights consist of two components: the right of publicity, which guards against the commercial exploitation of a person’s likeness and image without consent or compensation, and the right to privacy, which protects against unwanted public representation of one’s personality. Generative AI impacts the right of publicity both in the datasets it uses for training and in the outputs it creates, with different platforms affecting different aspects of this right.

Misappropriation cases generally arise in two ways: claims against third parties who use AI to commercially replicate another person’s likeness, and claims against the AI platforms themselves for profiting from technology that enables such replication.^{The first misappropriation (claims against third parties) which is the primary focus of this article, occurs when a third party uses a generative AI system to produce a person’s likeness and exploits that output for commercial gain without the individual’s consent. For instance, the company Lisa AI used an AI-generated imitation of actress Scarlett Johansson’s voice in advertisements to promote its generative AI app without her consent. The creation of AI voices relies on voice-cloning technology, where AI speech models are trained on human recordings to replicate natural tones and produce lifelike speech patterns. Likewise, models like Veo 3 expand this concern to the realm of video, as their capacity to generate lifelike visual content could be used for the commercial exploitation of a person’s likeness without their consent. Regarding the second claim, a platform may misappropriate publicity rights by using a person’s likeness in its training data without consent, for instance, if Google Veo 3 were trained on images or voices without permission.}

In conclusion, Veo 3 and similar AI video generators are transforming how content is created, offering remarkable creative possibilities. However, this innovation presents a range of IP challenges. Copyright issues arise at every stage, from dataset collection to final output, while questions of authorship and derivative works remain unsettled. Meanwhile, the ability to replicate voices and likenesses raises real concerns about the right of publicity. Navigating this intersection of technology and law will be essential to protect creators and individuals alike, ensuring that AI-driven creativity unfolds responsibly.

Image used was generated via gemini.google.com