The U.S. Copyright Office last week extended the deadline to submit comments in its inquiry into artificial intelligence, “To ensure that members of the public have sufficient time to prepare fulsome responses to the Office’s questions.” Respondents will now have until the stroke of Halloween to get their comments in, 12 days later than the original deadline. But while U.S. regulators and Congress deliberate, a group of deputies in France’s National Assembly (parliament) this month introduced an AI copyright proposal that could serve as an object lesson in just how difficult it will be to design a coherent and workable copyright regime for the new technology.
I don’t speak French, so I am unable to read the original bill. But according to a very helpful and seemingly thorough English-language summary posted on LinkedIn by Irish attorney and former head of CISAC’s Copyright and Artificial Intelligence Working Group, Barry Scannell (well worth the click to read the very lengthy discussion in the comments), it is much more specific as to licensing and rights management than the more general risk-based framework in the European Union AI Act.
The bill includes four titled sections. Article 1 would require a license or express authorization by authors or rightsholders for the inclusion and exploitation of copyrighted works in and by AI software, which presumably would include their use in training generative AI models. That tracks closely with what authors and rights owners in the U.S. and Europe and been demanding since the introduction of ChapGPT last year (although Scannell points out it could potentially conflict with the EU’s existing copyright exemptions for text and data mining).
Article 2 would establish that the only persons or entities that can claim the copyright in an AI-generated work are the authors or rightsholders whose works were “instrumental” in producing the generated work. That would diverge from current U.S. law, which does not recognize any copyright in works produced by a machine.
The French bill further proposes that existing author’s societies and other collective management organizations be relied on to manage those rights and remunerate the appropriate parties. But as discussed here in previous posts, identifying which works were “instrumental,” and in what measure, to producing an AI-generated output is effectively impossible in the case of transformer-based foundation models like GPT, which ingest billions of individual works in constructing their models.
It also misconstrues how generative AI models actually work. As Scannell notes, such systems “don’t typically draw from particular copyright works to create outputs – but rather a latent space mathematical representation of concepts of that work.”
Article 3 would require explicit disclosure of any AI involvement in producing a work, including the names of the authors whose work “inspired” the new creation. How those sources of inspiration would be identified apparently is not specified.
The final article would establish a system of taxation on works derived from untraceable sources that would be used to benefit the relevant rights organization — which sounds like a recipe to create the biggest “black box” in the history of collective management.
My grasp of French politics is even less sure than my grasp of the language, so I won’t hazard a guess as to whether the proposed law will go anywhere. But its introduction certainly reflects the greater urgency with which governments in Europe are moving to regulate AI than the U.S. government has thus far managed to muster. It also reflects the strong droit du auteur orientation of French copyright law, which sets it apart from the more instrumental approach in Anglo-American law.
If something like the French proposal were to be adopted by the broader European Union it could establish a regulatory framework that U.S. based AI companies would be bound to follow regardless of what the U.S. government ultimately does, much as the EU’s General Data Protection Regulation (GDPR) did for how social media platforms must handle users’ personal data.
What the French proposal shares with much of the debate around AI and copyright in the U.S. is the dubious premise that generative AI is simply another, albeit more sophisticated technology for the reproduction and distribution of copyrighted works, and that the same lattice of exclusive rights of authors can be applied. It also would seem to share the same blithe assumptions often heard here that existing collective management systems and organizations are up to the task, or could be brought up to the task, of collecting and equitably distributing royalties or license fees on the scale that would be required.
None of those assumptions really answer the challenges posed by generative AI. Nor are those answers likely to be found within existing copyright law, no matter how rigorously applied.
At bottom, copyright law operates as a form of technology regulation. It regulates the means and market mechanisms by which the fruits of human creativity are rendered perceptible, either directly or at a distance, and reserves the right to authorize the use of those mechanisms exclusively to authors.
Generative AI models clearly derive value from the training material they ingest, which is manifest in their output. But they do not derive that value by directly performing any of the operations regulated by copyright law. In that sense, generative AI models represent a kind of market failure. They extract value from the work of others, but copyright offers no ready model by which markets could put a rational price on that value.
Another model is needed. C’est la vie.