The arrival of OpenAI’s Sora is forcing Hollywood to confront generative AI more urgently.
While lawsuits are being considered, studios also need to figure out the future of licensing with AI companies, even beyond OpenAI, the creator of the controversial video-generation app. “If I were a major media company, I would certainly be asking my lawyers about the when, where and how to litigate, in parallel with meaningful licensing discussions,” said Michael Jury, cofounder and CBO at AiPhelion, an AI and IP consultancy.
Licensing could happen on the input (training data) or output sides of generative AI models and services:
Input: Nascent Licensing Market Taking Shape
None of the major studios have engaged in training data licensing, at least not publicly. (The Lionsgate-Runway partnership is often misconstrued as a licensing deal when instead it focused on training a custom model for Lionsgate’s use.)
Some film and TV producers, online creators and even streaming video services have secured deals, for instance through AI licensing intermediaries such as Protege Media or directly from services including Curiosity Stream (as discussed in the Luminate podcast In the Lab).
So far, AI licensing deals for audiovisual content have emerged for a few reasons, a source familiar with AI licensing tells Luminate Intelligence. Ethical AI developers such as Moonvalley and Adobe have purposefully chosen to license to serve creative enterprise need for clean models. Developers have also licensed out of a need for higher quality (e.g., HD video) or more specialized data (to fill training gaps) they can’t otherwise access by scraping.
Interestingly, some AI companies, including those that have relied on a fair use defense, are beginning to license audiovisual content. This suggests developers that have resisted licensing may be increasingly willing as they hedge against the potentially existential possibility of fair use weakening or collapsing in court.
That possibility isn’t foregone, judging by the Bartz v. Anthropic and Kadrey v. Meta cases with book authors. Though Meta won its case, the judge’s decision in fact argued forcefully against fair use because of the market harm and dilution of building a tool with material that enables it to generate countless competing, substitutional works.
Output: Sora Hints at Future of Licensing
Even as Sora created chaos in Hollywood, it did more to clarify what it might look like to license specific character IP or likenesses for duplication in a generative AI service, as long as rights holders and talent have given their consent, are fairly compensated and can set controls managing the output (e.g., ways, actions or scenarios the character is or is not allowed to appear or perform).
Even if it’s never with OpenAI as a partner, forms of licensing that monetize the outputs of AI services seem destined to emerge. For example, a studio might license a specific character IP to an AI company under an agreement that authorizes the developer’s service to let users generate outputs containing that character.
This could entail an upfront payment, in addition to royalties, whenever a user prompt generates that character. Likewise, talent could choose to license their real-world likeness, analogous to celebrities who lent their voices to Meta’s problematic but still lucrative chatbots or who have granted their images with Sora’s Cameo feature.
Licensing Complexity Tough to Navigate
For Hollywood, though, AI licensing is fraught with limits and complexity, whether input or output. Foreseeable issues would likely arise from attempts at dealmaking that put studios and talent agencies and guilds at odds.
For example, any character IP is bound together with talent likeness (image and/or voice), meaning a studio rights holder wouldn’t have total say over if and how any character it owns gets licensed into a generative setting.
“In order for studios to unlock the gold in the hills, they need talent to buy in,” said Dan Neely, CEO of Vermillio. “The best path to licensing comes with talent at the negotiating table.”
Licensing inputs and outputs aren’t mutually exclusive, but licensing IP to monetize AI outputs while ignoring the initial training on studio content could seem like tacit acceptance, leaving AI companies free to keep studio data in the model without a license as long as they reliably control and compensate for the outputs.
Yet some now believe the fight to monetize training rights is too precarious and therefore the focus should be on reaping value from the outputs. “Make money when AI companies try to create your materials,” said Luke Arrigoni, CEO at Loti. “If someone trains a model, it’s hard to say how that damages your company. But as soon as that model gets used, that’s when damages happen — so charge at that point.”
For studios, part of the catch-22-style challenge of figuring out a material strategy toward generative AI is that every option currently comes with constraints.
