Responsible AI Starts with Licensing

Abstract

Responsible AI starts with licensing. AI outcomes are strengthened by reliance on responsibly sourced, high-quality copyrighted works. Consent and human centricity are hallmarks of human advancement, and also will enable better AI. The default position of our copyright system is that the person who wishes to use copyright protected works must seek out and obtain a license before engaging in conduct that would implicate any rights protected by copyright law to avoid an infringement claim. This will only seem fair to most observers, given the intellectual and economic labor involved in creating original works, and also in light of the natural and economic justice of granting the creator the right to determine how her works will be used.

Of course, the copyright system contains exceptions in certain cases where authorization may not be needed at all (e.g., fair use in the U.S.) or where rights are limited to non-negotiated licenses (e.g., non-voluntary licenses and/or levies). But these exceptions or limitations to rights, in order to serve the public interest of promoting the creation and distribution of original creative works (as well as to comply with international law), must be carefully circumscribed to avoid unfairly prejudicing the legitimate interests of the author. To be acceptable under the Berne 3-Step test, they exist in areas of market failure.

AI training generally requires the express and voluntarily granted consent of the author. Any other approach threatens fundamental values underlying our copyright system. This observation is grounded in a number of core truths: that AI training requires the reproduction of protected works; that copies of such works are not (only) ephemeral or transitory and are stored in a manner that permits their retrieval for the purpose of producing expressive output that derives from the training data; and that such AI output do, and are likely to, directly compete in the marketplace for expressive works with the works on which the AI was trained, as well as to unfairly displace licensing opportunities that would otherwise exist for authors of the original works.

As we approach how to balance competing interests around AI technologies, we are faced with a litany of arguments that copyright is somehow not fit-for-purpose for this AI. These are the same arguments that were raised with respect to the sound recordings, cable, the photocopier, the internet, and every new technology where one party wanted to make money through uncompensated and unconsented to use of another’s creative works. AI is not as different as the advocates for unfettered, uncompensated reuse pretend.