Human-Content-to-Machine-Data_Final - Flipbook - Page 22
Licenses Have Limitations
We9re aware of efforts to develop new licenses and contracts, including to shape more
responsible AI development (e.g., Responsible AI Licenses85), curtail extractive practices (e.g.,
Post-Open Zero Cost Licenses86), and practice digital sovereignty (e.g., Nwulite Obodo Open
Data License87). Others have proposed ways to adapt CC licenses to address AI training.88 We
are currently treading cautiously when it comes to using licenses or contracts as a path to a
new social contract.
To have bite, licenses need an underlying intellectual property right. This is a key aspect of
the effectiveness of CC licensing: by design, they apply only when copyright applies, and they
do not impose contractual obligations on activity otherwise permitted under law (e.g., via
exceptions and limitations to copyright).
Given that AI training currently falls outside the scope of copyright in many scenarios,
compliance with the license conditions may not be required when using CC-licensed works
for AI training.89 For example, in a jurisdiction like Japan, where the act of reproducing a work
for purposes of AI training is permitted under an exception to copyright law, the CC license
would have no effect on that use. Given the wide and varied scope of the exceptions and
limitations to copyright law that apply to AI training, CC licenses are not well-designed for
imposing reciprocal terms on AI developers.
We are wary of using contract law to ûll that gap, at least on the public web. From a functional
perspective, contracts are difûcult to enforce when access to the information is technically
unrestricted. Without the control of copyright or another underlying intellectual property
right, a contract requires afûrmative agreement between the parties in order to impose
enforceable obligations. And, even if agreement can be secured, it wouldn9t bind other
parties, who may gain access to the data from a different source or further down AI9s
complex value chain.
85
Responsible AI Licenses. (n.d.). Responsible AI Licenses (RAIL). https://www.licenses.ai/
86
Post Open. (n.d.). What is Post Open?. https://postopen.org/about-post-open
87
C. Okorie & M. Omino. (n.d.). Licensing African Datasets. https://licensingafricandatasets.com/
88
Szkalej, K., & Senftleben, M. (2024, June 12). Mapping the Impact of Share Alike/Copyleft
Licensing on Machine Learning and Generative AI. Open Future.
https://openfuture.eu/wp-content/uploads/2024/06/Share-Alike-and-ML-Report-FINAL.pdf
89
Creative Commons. (2025, May 14). Using CC-Licensed Works for AI Training.
https://creativecommons.org/using-cc-licensed-works-for-ai-training-2/
22
