Licensing Your Movie & TV Content for AI Training: Can You? Should You?
Constructing a realistic audiovisual scene from a text description signifies an ongoing and exponential advancement in generative AI capabilities. In search of the video to train this AI text-to-video generation software, developers have approached studios for access to film and television content.
With a thus-far tepid response from major studios to these initial licensing requests, AI companies will likely expand their search to both U.S.-based and foreign owners of smaller but still significant film and television inventories.
As the most obvious advantage for library owners, granting training data licenses creates an additional revenue stream for existing programming. However, without suitable implementation, the potential cons for the licensing library owner include harming relationships with creative collaborators and partners whose persona, work or assets are ingested into the AI model along with the licensed library programming.
Guilds and Creative Collaborators
Any productions produced under the auspices of SAG-AFTRA, WGA, DGA or other entertainment guild must comply with the relevant union’s terms. Under the AI-related terms resulting from their respective 2023 strikes, SAG-AFTRA placed consent and payment requirements on each use of an actor’s digital replica, while WGA prohibits crediting AI as authoring scripts or other literary material.
Yet the prospect of licensing existing content for AI training remains unresolved. None of the new agreements negotiated by each of the three guilds explicitly prevent producer signatories from making past film and television productions available for use as AI training data.
Instead, all kicked the training data can down the road. SAG-AFTRA and WGA stop short of prohibiting AI ingestion but reserve the right to argue that making programming available for use as training data violates the guild agreements and/or other laws. The DGA mandates twice-yearly meetings with studios to discuss the intended uses for generative AI and appropriate remuneration for any material directed by its members that may be used as training data.
Meanwhile, owners of nonunion content aren’t completely cleared of needing to consider creative professionals in licensing decisions, but their responsibility will depend on individual talent contracts. Performers and other collaborators on nonunion productions normally do, and should, render their performances and services as a work-made-for-hire, which under U.S. law requires specific circumstances and/or contractual language per instance.
AI-related carve-outs are becoming more common in deals, but, without such a carve-out, a nonunion producer’s right to offer the work as training data relative to those collaborators seems solid.
However, both union- and nonunion-bound producers are still subject to laws beyond talent agreements. Notably, there is still the potential that output emerging from an AI model trained on licensed works would violate a performer’s publicity rights, which poses a liability risk for any producer contemplating licensing for AI datasets.
Producers still lack the right to integrate a performer’s face, voice and other identifying characteristics into AI model output unrelated to the production for which the producer hired the performer, thus the producer has no such rights to sub-license to an AI company.
Jointly Owned Content
Libraries with multiple owners may need unanimous owner approval. In the absence of a written agreement dictating control of licensing decisions, U.S. copyright law allows any single owner to grant a nonexclusive license as long as the licensing owner shares license income with the non-licensing owners.
However, this single-owner approach can complicate a licensee’s use of the content outside the U.S. That complication partially explains the entertainment industry custom of requiring all rights owners to participate in a license grant.
AI companies, infamous for gathering training data by scraping content without authorization, might risk going forward with single-owner license grants. That option is potentially workable for the granting single owner, as long as the licensing arrangement contains no commitments the single owner cannot fulfill.
Underlying Assets in Film & Television Content
Film and TV programming often contains assets the producer licensed from a third party, such as news footage, movie clips, images, music or artwork. Including those third-party assets in a program entails a multistep research-negotiation-documentation process, as discussed in detail in VIP+’s 2023 special report “Rights Clearance in Film & TV.”
Few, if any, producer-third party license agreements that predate generative AI’s first public products (November 2022, with ChatGPT) specifically address ingestion of the licensed assets into an AI training dataset. Even if an agreement includes the holy grail of licensing language permitting the producer to use the licensed assets “in perpetuity throughout the universe in any manner, language, and media, whether now known or hereafter devised,” one can still reasonably argue that language does not permit AI training.
The success of such an argument might be uncertain. Much more certain is the outrage from rights holders upon learning that their assets have been ingested into an AI model without their explicit consent and without compensation.
Before concluding any training data arrangement with an AI company, each library should conduct its own clearance analysis and might conclude it has the right to offer its programming with the underlying assets. If not, a technical workaround might be flagging all underlying licensed materials and excluding them from AI ingestion.
One business workaround entails resubmission of rights requests and renegotiation with the asset rights holders. Renegotiation can be onerous, as ownership of the assets may have changed and rights clearance records are frequently not retained in a format that facilitates historical research. Of course, the reckless approach means proceeding with AI ingestion without addressing the underlying assets and relying on the hope that asset rights holders remain unaware or indifferent.
Legislative Intervention
Current legislative momentum leans toward curbing any potential negative impact of AI by expanding publicity rights, banning harmful deepfakes, protecting creators and holding accountable all who enable the misuse of another’s persona. This momentum surges with each high-profile AI misdeed, such as the deepfake pornographic images of Taylor Swift and the perceived similarity of OpenAI’s GPT-4 “Sky" to the voice of actress Scarlett Johansson.
Laws such as Tennessee’s recently signed ELVIS law (the Ensuring Likeness Voice and Image Security Act) and drafts of the federal NO FAKES Act seek to strengthen, clarify and expand publicity laws that allow each individual to control and profit from the commercial value of his or her own identity — often with emphasis on insuring protections for voice.
In addition, recently enacted and pending U.S. federal and state laws clarify that AI developers can be liable for deepfakes and unauthorized digital replicas produced by their models.
If passed, the Generative AI Copyright Disclosure Act introduced by California Rep. Adam Schiff last April would require AI developers of consumer-available generative AI systems to file a notice with the U.S. Copyright Office listing all the copyrighted works used in the system’s training dataset. The notices would be publicly available and would inform creators and owners whether any of their projects were used as training material.
Many of these laws are pending and may not pass in their current iterations. Nevertheless, they serve as an alert of forthcoming patterns. While they primarily target AI developers, companies providing training data or otherwise partnering with generative AI companies should be cautious about becoming entangled in any upcoming laws and regulations.
Perspectives Guiding a Decision on Whether to License
Film and TV program owners contemplating a training data license may balance a desire for short-term financial gain with qualms about supporting a system that might eventually compete with and undercut the economic longevity of many industries, not just creative ones.
The integration of generative AI capabilities throughout the creative industries seems significant and inevitable. For stakeholders that want to stay relevant, active engagement may yield a better outcome than sitting on the sidelines.
Having a voice in AI’s development might open opportunities to influence how it is ultimately used. Some have opted to exercise that voice through one of the many pending lawsuits against AI developers or through submitting testimony and comments in relevant legislative and regulatory proceedings.
Controlled collaboration offers a different method of influencing outcomes, even if that influence is confined to one’s own library or body of work. Given the current dynamics and uncertainty, licensing one’s film and TV programs for AI training without conditions and without monitoring seems perilous.
A more participatory arrangement seems more prudent by providing opportunity to institute guardrails that discourage infringing and offensive AI model output and to minimize other potential licensor risks. This might be why major Hollywood studios are less eager to allow their content to be used indiscriminately and more open to discussing partnerships with AI companies.
Collaborations could extend beyond licensing to partnering with AI companies for development of consumer facing tools or company internal tools. Of course, even controlled collaborations need to navigate the rights, guild and legislative challenges outlined above.
Joy R. Butler is an East Coast-based attorney providing strategic counsel for the entertainment and digital technology sectors.