Technical Needs to Establish Licensing for AI Training
Note: This article relates to the March 2025 VIP+ special report “AI Training — Consent & Content,” available only to subscribers.
Despite a still unsettled legal environment, 2024 saw the first stirrings of a commercial licensing market between AI companies and rights owners for the use of copyrighted works to train generative AI models. Nearly three dozen content licensing deals were announced or reported by the end of 2024 involving rights owners in the U.S., Canada and Europe, according to VIP+ tracking.
SEE ALSO: VIP+’s Special Report “Generative AI & Licensing”
While encouraging, at least from rights owners’ perspective, the deals so far have been limited in scope and scale, raising questions as to how widely the model those deals have followed can be applied beyond their immediate reach.
Most of the content licensing deals announced so far have had a few characteristics in common:
For a more robust, sustainable and broad-based licensing system to evolve — one that could enable participation by rights owners beyond the tier of large publishing, music and image archives capable of engaging directly with technology companies — at least three key technical capabilities will need to be in place:
All of those systems will also need to operate at a machine-to-machine level, so the processes they manage can be automated sufficiently to handle the volume of works involved. These systems will further need to be standardized under one or a limited number of technical methodologies to ensure broad adoption.
SEE ALSO: VIP+ Unearths AI Insights From All Angles — Pick a Story
As lawmakers take positions on AI training, they have begun to encourage developments on the technical side that some expect will be needed. Several jurisdictions — notably the EU and the U.K. — have begun to coalesce around an “opt-out” standard as part of a larger permissive stance on online data scraping and the use of copyrighted works to train AI models.
Most notably, the EU AI Act stipulates that such a reservation of rights must be in a machine-readable form to require compliance by AI systems. Laws modeled on the AI Act include a similar requirement.
Rights owners generally object to the opt-out legal standard on grounds that it inverts the traditional copyright model by placing the burden to comply on creators rather than on those who wish to use or exploit their copyrighted works.
Apart from the impracticality of opting out of all AI systems individually, the machine-readable opt-out requirement also leaves the industry with a significant technical challenge.
There are yet still no “widely used industry [technical] standards to identify and comply with” machine-readable expressions of rights reservations, as stipulated in a draft version of the EU AI Office’s code of practices for AI developers.
The lack of any such standards and systems has spurred technology developers, entrepreneurs and rights owners themselves to develop machine-readable tools and systems to enable opt out of AI training.
Now dig into the VIP+ subscriber report ...
Despite a still unsettled legal environment, 2024 saw the first stirrings of a commercial licensing market between AI companies and rights owners for the use of copyrighted works to train generative AI models. Nearly three dozen content licensing deals were announced or reported by the end of 2024 involving rights owners in the U.S., Canada and Europe, according to VIP+ tracking.
SEE ALSO: VIP+’s Special Report “Generative AI & Licensing”
While encouraging, at least from rights owners’ perspective, the deals so far have been limited in scope and scale, raising questions as to how widely the model those deals have followed can be applied beyond their immediate reach.
Most of the content licensing deals announced so far have had a few characteristics in common:
For a more robust, sustainable and broad-based licensing system to evolve — one that could enable participation by rights owners beyond the tier of large publishing, music and image archives capable of engaging directly with technology companies — at least three key technical capabilities will need to be in place:
All of those systems will also need to operate at a machine-to-machine level, so the processes they manage can be automated sufficiently to handle the volume of works involved. These systems will further need to be standardized under one or a limited number of technical methodologies to ensure broad adoption.
SEE ALSO: VIP+ Unearths AI Insights From All Angles — Pick a Story
As lawmakers take positions on AI training, they have begun to encourage developments on the technical side that some expect will be needed. Several jurisdictions — notably the EU and the U.K. — have begun to coalesce around an “opt-out” standard as part of a larger permissive stance on online data scraping and the use of copyrighted works to train AI models.
Most notably, the EU AI Act stipulates that such a reservation of rights must be in a machine-readable form to require compliance by AI systems. Laws modeled on the AI Act include a similar requirement.
Rights owners generally object to the opt-out legal standard on grounds that it inverts the traditional copyright model by placing the burden to comply on creators rather than on those who wish to use or exploit their copyrighted works.
Apart from the impracticality of opting out of all AI systems individually, the machine-readable opt-out requirement also leaves the industry with a significant technical challenge.
There are yet still no “widely used industry [technical] standards to identify and comply with” machine-readable expressions of rights reservations, as stipulated in a draft version of the EU AI Office’s code of practices for AI developers.
The lack of any such standards and systems has spurred technology developers, entrepreneurs and rights owners themselves to develop machine-readable tools and systems to enable opt out of AI training.
Now dig into the VIP+ subscriber report ...