Lionsgate Reaches a Deal With Runway for AI Training Data
But the law of the conservation of information could have the last word
From the beginning, the movie and TV studios have had a different and more complex relationship with generative AI from their peers in other media sectors. News publishers, music labels and publishers, and photo agencies have been concerned primarily with litigating and lobbying against the unlicensed use of their archives to train AI models, and, more recently, pursuing paid licensing deals with developers. The movie and TV studios, in contrast, have been keen to make use of AI technology and have focused instead on ensuring any new regulation around it not unduly limit their ability to exploit its capabilities.
The Motion Picture Association, for instance, initially opposed the California measure recently signed into law by Gov. Gavin Newsom regulating the use of performances generated by AI, relenting only after changes were made to the bill to exempt AI’s use in standard post-production techniques from the prohibition.
The studios have also tread more cautiously than other rights owners when it comes to seeking out licensing deals with AI companies for their libraries, although not because AI companies haven’t come knocking. The studios must be careful in how any such deals would be structured due to the collective bargaining agreements reached last year with writer and actors following painful work stoppages that shut down production for months. Those strikes were triggered in large part by disputes over the use of AI in production and the subsequent use of production elements (scripts, performances, etc.) to train AI models.
Some of that reticence may be starting to ease, however. This week, Lionsgate reached a deal with Runway to provide the video-generator developer with access to the studio’s library of moves and TV shows to train a new AI video model. A key to the deal, according to reports, is that the new model will be trained exclusively on Lionsgate’s 20,000-title archive and the model will be proprietary to the studio for its own exclusive internal use. Lionsgate intends the use the model primarily in pre- and post-production, such as for storyboarding and VFX. As studio vice chair Michael Burns told the Wall Street Journal, “We do a lot of action movies, so we blow a lot of things up and that is one of the things Runway does.”
For Lionsgate, the deal is something of a workaround. It attempts to tap the value now sitting in its vaults without running afoul of the guild contracts, by limiting the model’s application to parts of the production process not covered by those agreements, and by keeping the fruits of the model training to itself.
That may work better on paper than in practice, however. In physics, the principle of conservation of information holds that information cannot be destroyed, or deleted from the universe, even within a black hole. It still exists inside the black hole, even if it is permanently locked away from observation (Hawking radiation notwithstanding).
Something similar applies to AI models. The information they extract from their training data remains in the model even if the dataset that trained it is deleted. What a model “learns” are essentially fundamental, generic principles about whatever type of data it’s fed, abstracted from the data’s observable qualities.
Once the Lionsgate/Runway model is trained, it will no longer matter that it was trained exclusively on Lionsgate content. It will have learned what it learned about motion pictures and it can’t be made to unlearn it. AI models do not forget what they’ve learned.
Insofar as the Lionsgate model will be built on Runway’s foundational video-generator model that information extracted from the Lionsgate material will permanently alter the weights (parameters) of the underlying Runway model. The Runway model will have learned what it could from the Lionsgate data, never to unlearn it.
Unless the Runway model itself were to disappear, the informational value within the Lionsgate data will no longer be proprietary to Lionsgate, regardless of any customized model that’s created from it.
It is a fundamental dilemma facing all rights owners as they contemplate licensing their content to AI developers for training models. Once the informational value of that content is extracted it stays extracted even after the expiration of any business deal between the parties.
Absent any means to monitor exactly how any individual work is being used to generate a particular output and to put a price on its ongoing use, the AI company would have no technical need to renew the deal to continue to benefit from the value that has already transferred.
More on that in a future post.