The Fair Use Tango: A Dangerous Dance with [Re]Generative AI Models
By Neil Turkewitz
Every year around this time — designated by some organizations as “Fair Use Week,” the copyright world engages in a carefully orchestrated dance. Organizations & individuals that are skeptical about the role of copyright and worried about its implications for expressive freedoms celebrate the role of fair use in limiting proprietary interests while recognizing the value of what they see as an appropriately limited set of rights under copyright. For their part, copyright owners & folks that see the incentives of copyright as being fundamental to the interests of the public by creating incentives for the creation & distribution of creative works use the occasion to highlight the dangers of expanding fair use beyond its carefully circumscribed purposes, while recognizing the importance of what they see as an appropriately constrained fair use doctrine as an essential feature of the legal & cultural landscape.
Over the years, I have been a willing participant in this dance, although I have increasingly wondered whether there is any value in repeating the well-known choreography. Is anyone learning from this annual exercise? Is anyone speaking to folks that don’t already agree with them? I have tried on several occasions to do so, but I fear that my role as a known champion of creatives & copyright dooms such an effort to find common ground. Five years ago, I wrote:
”Lost in talk of “copyright wars” and an over-simplified copyright versus tech narrative, is the fact that there is a great deal of common interest and agreement. Indeed, the age of information would be arid (and uninformative) in the absence of either information or modes for distributing and accessing it. Tech and content need each other, are part of each other, and in fact bleed into one another. While there are certainly some outliers and differences related to theories of utilitarianism versus natural rights, at core there is broad agreement on the foundational issues and purposes of copyright law, including the need for a copyright regime that ensures support for working creators and promotes innovation. Most tech companies would agree that copyright protection serves the public interest, and that copyright systems should foster new forms of expression. The creative sector for its part supports the idea that the establishment of modern, robust and effective protection should include appropriate limitations or exceptions to rights, consistent with international treaties and the needs of different cultures (both legal and creative).”
I wish I could say that we have progressed a more common vision in the five years since that appeared, but that would be incorrect, and I will leave hallucinations to ChatGPT and other large language models. Instead, I will offer what I see as a fundamental truth — that regardless of where one is on the copyright-fair use scale, it is now time for us all to be joined in ensuring that the training of publicly available (i.e. non-research) AI generative models (StableDiffusion, Midjourney, DALL-E, ChatGPT, etc) takes place only on the basis of consent, not through nonconsensual text & data mining. This is way bigger than copyright, but copyright will, for better or worse, play an important role in the construction of rules for a digital world (which is to say, the modern world) and the extent to which it is anchored in human dignity and self-determination. My humble suggestion is that we strive for the “better” option.
The issues presented by generative AI are unlike any that society has faced before. It forces us to consider what makes us human, and the extent to which we will allow the constitutive elements of our personhood to be captured by machine. From our personal photos, to our stories, to our medical data, to…well, everything that is known about us and captured in some manner on the internet, is this all grist for a mill of incomparable appetite? Or do we want to take steps to ensure that our emerging and ever-changing digital world is built on the basis of clear, freely granted, consent? Not as part of the non-negotiated & non-negotiable terms of living in a digital world, but as an expression of free will in which the relevant actor has agency?
In many respects, copyright offers the cleanest, most immediate way of answering at least part of this question. It exists upon creation of an original expression, not dependent on any formalities, and making a work available to the public doesn’t divest the property interest — indeed, copyright is largely, albeit not completely, rooted in a desire to incentivize the sharing of one’s works with the public. Unlike more complicated privacy issues, there is no need to define the protected interest — it is already well-defined and protected globally through a variety of international instruments. There is really but a single issue standing between self-determination and machine supremacy — ensuring that training AI without the consent of the creator is not excused by exceptions to copyright, including of course, fair use.
Coming back to the premise of this piece, there are many complicated cases in which competing equities place the doctrine of fair use front & center, and in which parties may reasonably disagree (or agree) about where & how to draw appropriate boundaries in pursuit of achieving the Constitutional purpose of copyright. To my mind, the unauthorized use of copyright works to train AI is not such a case. And by this, I don’t mean that there are no legal arguments that may be made to justify such use. I think there are, although I think they are weak. But I do think that there are no moral arguments to be made to justify this unprecedented appropriation. That new transformative works may be made from a database developed on the back of conscripted labor may have some resonance in law. However, it fundamentally fails the humanity test.
Fair use has traditionally been seen as advancing redistributive justice — preventing powerful actors from restricting competition in the marketplace of the expression of ideas. As I wrote last year during Fair Use Week:
“Fair use is generally portrayed as a lever against monopolies, exercised by those without market power against those who do. But as noted by Ben Sobel in his 2017 article “Artificial Intelligence’s Fair Use Crisis,” “Today’s digital economy upends this narrative. … This pivot in market dynamics should prompt a corresponding shift in attitudes towards fair use. The doctrine no longer redistributes wealth from incumbents to the public; it shifts wealth in the other direction, from the public to powerful companies.”
This shifting of wealth from the public to powerful companies has now come into crystal clear vision with the emergence of generative AI models. To allow fair use to power this theft from the public and individual creators should be unthinkable. I know that fair use and fairness aren’t the same thing, and that fair use can sometimes be unfair to a particular creator in light of the broader perceived value of allowing a particular use. This, I submit, is not such a case. Fairness and a careful & balanced approach to fair use are wholly aligned. This Fair Use Week, let’s all commit to a human-centered future, rooted in an appreciation of the role of consent in defining our universe. The alternative is chilling — strip-mining of the soul, leaving only inhospitable conditions for the flowering of humanity.
