Many have argued, on a variety of bases, both economic and legal, that GenAI-works simply should not hold copyright while non-GenAI works should. If only it were so simple. Unfortunately, differentiating copyright protection between GenAI-works and “traditional” works relies on a false dichotomy, entirely at odds with a creative landscape where GenAI tools are fast being integrated into all stages of the creative process.
In this short essay, we consider the challenges in meaningfully differentiating copyright over otherwise identical works based on a level or nature of GenAI use. In turn, we speculate on copyright’s sustainability in prospective digital economies with foundationally shifted modes of GenAI-supported creative production. With a focus on music industries, we consider the legal and economic ramifications of allowing copyright over gluts of content that require minimal investment to produce. We explore how this would affect copyright’s fundamental incentives and may lead to low-cost “copy-mining”. We then speculate on the effects of the opposite extreme – removing copyright entirely from these works – and the dangers of driving production towards close-to-zero marginal cost works.
Although much is still unknown, we conclude that failure to craft appropriate adaptive responses to the foundational challenges to legacy copyright regimes afforded by new modes of Gen-AI enabled production will likely result in significant disruptions to creative industries and efforts that copyrights are intended to protect.
Introduction – Beware of False Dichotomies
While Generative AI (GenAI) has captured the global imagination and the prospect of a tidal wave of GenAI-curated and generated content is widely anticipated, there is something of a hubbub over whether GenAI outputs should be able to hold copyright.
Many have argued, on a variety of bases, both economic and legal, that GenAI-works simply should not hold copyright while non-GenAI works should. If only it were so simple. Unfortunately, differentiating copyright protection between GenAI-works and “traditional” works relies on a false dichotomy, entirely at odds with a creative landscape where GenAI tools are fast being integrated into all stages of the creative process. This means that we need not ascertain whether a work developed using a GenAI tool can hold copyright, but rather when. What level and nature of interaction with GenAI tools is (in)sufficient for an author to receive copyright to their work? While courts worldwide have grappled with establishing and interpreting thresholds for human involvement as the requisite dividing lines that underlie the dichotomy, there are a few complicating factors to determining this question.
Generative AI Use in Music and the (In)Distinguishability Conundrum
The first is the incredible and evolving spectrum of uses for GenAI tools, from those that negligibly impact the final work to those that generate the entire final work with negligible human effort. Let us just consider the world of music. There are GenAI tools that only affect the production of an already written and recorded song. There are other GenAI tools that only generate (or re-generate) one instrument – such as the bass or the drums. There are still others that generate entire songs from scratch. The products of each of these can be iterated and reiterated upon, both with and without GenAI tools, until they may bear no resemblance at all to their initial AI-generated forms. These tools can be easily integrated into professional creative workstations – mere buttons to be pushed, toys to be played with. They can be relied upon in both trivial and substantial amounts to form the final product. Given this infinite spectrum of creative processes, which works are uncopyrightable GenAI-works and which are not?
Here’s the next problem – how do we tell the difference? We have no means of granularly tracking an artist’s creative workflow such that we could appreciate every single instance when a GenAI tool was used, how it was used, and to what extent. Watermarking tools remain unreliable at determining whether GenAI was used at all, let alone at providing a detailed understanding of exactly how it has been used. In turn, even the artist may not know exactly where she has used GenAI if she has not been recording her entire creative process, and even if she did, she might be disinclined to disclose its use in order to receive full rights to her work.[1]
Thus, it would be patently absurd to argue that even negligible GenAI tool use on a minor element of a work should instantly disqualify it from holding copyright. Rather, as we mentioned briefly above, regulators have tried to ascertain not whether, but what threshold of GenAI use will allow copyright over the output. Naturally, different national copyright regimes applying different authorship tests have reached different conclusions. The Beijing Internet Court in China granted copyright to an image which was created entirely through prompting on a text-to-image GenAI -tool, on the grounds that the final work sufficiently reflected the plaintiff’s “intellectual investment” and “personalized expression” to meet the requisite thresholds of “originality” and “intellectual achievement.”[2] The US Copyright Office (USCO) has stated that authors could hold copyright in GenAI-works, while exclusively publishing decisions rejecting copyright to AI-generated elements.[3] Most ambitiously, the USCO seeks to granularly grant copyright to just the non-GenAI-elements within works partly built using GenAI tools. The only court in Europe to tackle the issue so far was in the Czech Republic, which avoided the threshold question entirely by rejecting copyright over the GenAI-output outright on the misguided grounds that “[t]he plaintiff did not personally create the work; it was created by artificial intelligence.”[4]
Thus, instead of a “non/GenAI-work” dichotomy,[5] we are faced with the unfortunate circumstance of a litany of disparate disharmonised uncertain national authorship thresholds. Each of these seeks to maintain a separate dichotomy by drawing (somewhat arbitrary) lines in the sand somewhere in the endless sea of prospective GenAI use cases, with different jurisdictions drawing different lines.
Problematically, a lack of international harmonization of regulatory approaches creates regulatory uncertainty and arbitrage opportunities. This regulatory uncertainty undermines one of the key economic rationales behind the entire copyright regime: the creation of a property right to facilitate the existence of markets for original content. Uncertainty as to whether artistic works may have copyright protection or not diminishes and distorts incentives by artists (and investors) in the creation of new content. Arguably, this might have been less of a problem with physical works than for digital works that may be inherently infeasible to control within a single sovereign nation’s regulatory rules. Multiple economic challenges may arise in GenAI-fueled creative industries as those seeking copyright protection or seeking to bypass such protections engage in venue shopping, strategically choosing where to create works, or how to use AI in the creation process based on private preferences for protection. Fundamentally, such venue shopping based on non-harmonized copyright regimes would distort market competition and induce regulatory competition, favoring a lowest-common-denominator regulatory regime.
Beware of the Measurement & Auditing Challenges
Even were the legal authorship threshold somehow harmonised in every country, we confront a complex measurement challenge that fundamentally questions the enforceability of threshold-based regimes that imply regime separation and regulatory bifurcation.Co-existing regimes based on a binary non/GenAI-work distinction require a robust definition of a dividing line between the two categories. Where such a dividing line is or should be is hard to determine, especially given not only the evolving nature and variety of GenAI tools and uses but also the heterogeneous skills users have in building and using GenAI tools.[6] A singular focus on quantitative thresholds is certainly insufficient.[7] Additional qualitative thresholds assessing the intellectual contribution (e.g., based on similarity considerations that are inherently multidimensional and become particularly delicate given the wide range of GenAI uses and the capability of GenAI tools to create customised content at scale) raise concerns about the scalability of case-by-case assessments.[8]
The situation is further complicated by regimes that seek to separate the copyrightability of individual elements within works, predicated on which elements have been overly reliant on GenAI and which have not. Each of these dichotomous threshold difficulties then applies to every individual element of a work which might have used GenAI, meaning parts of each work could fall into the public domain in some jurisdictions, without it being clear which of these do without detailed understanding of the work’s production process. A truly chaotic outcome.
Even if we find a suitable threshold, a fundamental information problem remains. As of today, we have no means of auditing each work to a granular enough detail to understand exactly and reliably which elements have used GenAI, which have not, and which have used GenAI but then been meaningfully refined. As mentioned before, technological auditing solutions proposed (and in some cases, mandated), such as watermarking, are not internationally harmonised and are reportedly not robust.[9] Yet, even if they were, merely informing that GenAI has been used in the development of a work tells us nothing of the author’s relationship to it (any more than a “software was used here” sticker).[10]
As one of us authors has written:
“In order for a coherent international market to form predicated on a dichotomy of rights over identical works, there must be (1) a level of detailed understanding of how each work is produced (which may be impossible to achieve with many works), in order for (2) the works to be evaluated against a certain specific enough standard (which is currently neither certain nor harmonised). There are significant challenges in reconciling both of these elements which may be insurmountable.”[11]
Absent the ability to distinguish between different categories, seeking to differentiate between non/GenAI works might be well-intentioned but seems infeasible and unworkable.
Thoughts & Speculations About Possible Futures
If the dichotomous approach is not workable, then it is worthwhile considering what might be done. How might copyright be reimagined or reformed to accommodate a future where a growing share of new artistic works are created using GenAI tools? In the limit, we might imagine GenAI tools being able to generate content without any human artist involvement of such quality that based solely on its final product the works are indistinguishable from artist created content of all types and qualities.
Two extreme options immediately present themselves. First, we might simply dispense with the effort to limit copyright protection to works that are created with less than some threshold level of human effort and grant copyright protection uniformly to non/GenAI-works of a certain quality. Second, we might adopt the other extreme position and dispense with copyright protection uniformly for non/GenAI-works. Economic challenges arise with either alternative, even setting aside the obviously difficult transition and enforcement issues either such regime change would confront.[12]
The first option might have significant disruptive effects on creative industries in the short and long term. In the limit, a surge of supply in near-zero-marginal-cost-to-create GenAI content might near-fully displace artists’ content in the market for new works; and even in the short term, it might tilt the balance of competitive advantage in favor of major rightsholders with more GenAI resources and skills.
Still, our relationship with art seems fundamentally linked to human expression (for now), rendering a “world without human artists” unlikely. Even if GenAI can create works of aesthetically equal quality, their value may be diminished if this aesthetic is not itself linked to the expression of a human author. A perfectly convincing Taylor Swift knock-off is unlikely to hold the same value as Taylor Swift’s new song, even if the former had a catchier melody and more evocative lyrics. Thus, a market flooded with aesthetically high-quality but company-authored works may rather increase the value of artist-made creative works that are more expressive of individual personality or narrative, regardless of their aesthetic quality. Some superstar economics may be at play here.
Yet not all creative products require that level of human expression nor indeed even aesthetic quality of particular merit. We might consider, then, a spectrum of creative products, with the left side serving wholly functional creative works.[13] These built-for-purpose works would require minimal aesthetic differentiation and consumer appreciation. On the right side of the spectrum, we would have artistic creative works whose value is predicated on consumer appreciation of differentiation of human expression. We might stipulate that artistic labour in works furthest to the left side of the spectrum would be most readily substituted for by GenAI tool capabilities. The other end of the spectrum is more complicated. As the barrier to entry to create works is lowered for the public at large, this will devalue works that are easily created with low skills. In turn, the perceived quality of artistic works will be in a constant state of evaluation against the immediately available creative standard of GenAI tools, creating a “man vs machine” dynamic as artists try to differentiate themselves from easily generated works while the machines ingest their innovations, rendering future imitation easier. This likely feeds the aforementioned value of the individual artist’s persona and brand, rather than the pure aesthetics of her work, as it is less easily imitated by lower-skilled content creators (at least, authentically).
We might consider this prospective heightened value of persona in line with contemporary creative platform ecosystems, where for many creators, value may be derived less from each individual work they create and more from the entirety of the artist’s brand or oeuvre. On many of the most popular and lucrative creative platforms, creators are incentivised to create large amounts of content regularly to boost their algorithmic visibility and engagement, meaning gluts of new content can increase the value of the artist’s portfolio of work. It is exactly this demand for continuous creation of new content that GenAI tools can assist creators with supplying.
These circumstances might be especially problematic if we are to grant each of these works copyright. Firstly, the public domain of ideas might quickly be mined, as GenAI affords creators the ability to develop seemingly endless streams of content. If creators are then granted a monopoly to all of this content and the marginal cost of generation is close to zero, then “copy-mining” of content is inherently incentivised, allowing those with the most computational power to own the most copyright, granting them rent-seeking power over other artists.[14] Yet more foundationally, it would be increasingly difficult to ascertain whether another artist has infringed upon the first at all. Currently, if two similar enough works are created and the author of the latter had access to the former, it is often assumed that they have copied from the earlier work. This is no longer the case, as GenAI tools allow different authors to create highly similar works without ever having engaged with one another.[15] Thus, artists may be able to rely on the GenAI defense to copyright infringement. Clearly, maintaining copyright as it stands would be dysfunctional in these new circumstances.
The alternative extreme is to effectively give up on copyright for new works. This might be justified by a view that the stated purpose of copyright regimes is to incentivise investment in new art in order to ensure adequate supply is rendered moot in a world where the marginal cost of generating unlimited quantities of art is close to zero. As it stands, we are yet to see an especially valuable artistic work that required zero human creative labour, let alone markets forming around such works. Yet, as individuals further tailor models towards consumer preferences, it is far from inconceivable. The removal of copyright without replacement might therefore drive production towards close-to-zero marginal cost works, even if they are not as accomplished as works that require greater creative investment. It is clear then that creative investment must still be protected, lest works that require time and effort are diverted from entirely towards a market of automatic art. Still, the form this protection takes must itself foundationally adapt to the new realities of creative production. Thus, it appears that copyright must strike a new balance – one that incentivises and protects the creation of new innovative works (rather than solely the cheapest to fulfil a function) without incentivising “copy-mining” and rent-seeking from potential future rightsholders with monopolies over massive swathes of content despite minimal investment.
As this brief essay should make clear, we believe that GenAI poses an existential threat to international (legacy) copyright regimes and that disjointed efforts to impose a dichotomous non/GenAI-works protection regime are doomed to failure. Any disruption to the existing regime threatens the complex industry structure, value chains, and fabric of contracts, business models and practices that continue to evolve with the evolution of digital technologies and the market opportunities they create. The forces coming for the copyright regime are unstoppable and failure to craft appropriate responses will likely result in significant disruptions to creative industries and efforts that copyrights are intended to protect. Further work on the likely economic implications of alternative approaches is clearly needed.
[1] For a detailed analysis of the challenges we discuss in this contribution, particularly here regarding auditing / enforcing GenAI use, see: Zachary Cooper, “The AI Authorship Distraction: Why Copyright Should Not Be Dichotomised Based on Generative AI Use” (July, 2024).
[2] Li v Liu, Jing 0491 Min Chu No. 11279, (Beijing Internet Court), 15.
[3] Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 88, FED. REG. 16,190 (Mar. 16, 2023) (to be codified at 37 C.F.R. Part 202), 16192 ; Edward Lee, The Code Red for Copyright Law, 76 Florida Law Review (2024, 9
[4] Czech Republic Judgment In The Name Of The Republic (Case No. 10 C 13/2023-16) (Oct. 11, 2023)
[5] That is, a classification dichotomy that treats works that are authored without genAI (“non”) differently than works that are authored with genAI (“GenAI”).
[6] Even when a tool may support GenAI, artists may not elect to use those options, and even when they do, may elect to use those options with significantly different implications for the role GenAI may have played in its creation. An electric guitar may have lots of capabilities for producing music that differ significantly depending on who is playing the guitar and what the musician is trying to do. Furthermore, it is worth noting that not all digital tools are AI and not all AI is GenAI and that differentiating between these technologies and tools is a gray area.
[7] Think about a creator who has built and trained their own GenAI tool and uses it with great skill. Arguably, the creator could come up with something original and innovative with a single prompt. What we should keep in mind also is that this example shows that the sunk investment (e.g., related to model building and training, data, and skill building) is much greater than the incremental or marginal cost of content production, which is often considered to be the cost to be recovered from copyrights. The latter might be more accurate if a user uses an existing GenAI tool via a platform, but it might not be accurate enough to refer to sunk investment if the creators do more. Note here also that GenAI tools are technologically complex, and building and training them is not necessarily easy. Despite this, open source access to creative models, such as generative audio models like RAVE and Dance Diffusion, are increasingly allowing artists to train models on content of their choosing. This has led to a rich underground of bespoke artist-created GenAI tools and AI art communities. These communities are mostly rooted in Discord servers, such as the DADABOTS KVLT server, among many others.
[8] Think about content moderation on social media platforms. Humans are in the loop here. As content production is ramped up due to lower cost and entry barriers with GenAI, the inability of algorithmic/AI tools on the auditing side to counter the supply-side glut would inevitably lead to worsened signal-to-noise ratios. It is unlikely that human labor can be scaled to match the AI-enabled production increase and content glut. See also William H. Lehr & Volker Stocker (2024). Competition Policy over the Generative AI Waterfall. In: A. Abbott & T. Schrepel (eds.) AI and Competition Policy, Concurrences (pp. 335-358). While this emphasizes human labor that is sometimes invisible and perhaps unknown to many, invisible work may also be similarly relevant for auditing in the GenAI assessments we focus on in this piece.
[10] Cooper, 5-16. In other words, transparency with respect to the use of GenAI in preparing work does not remedy the information problem informative about the role of humans vs. the role of AI.
[12] Significant transition issues always arise when long-lived regulatory frameworks need to be substantially overhauled. GenAI is a wrecking ball to existing copyright protection regimes. The enforcement issues arise because it will be substantively harder to determine infringement violations which might occur in either direction – new GenAI infringing new non-GenAI and vice versa.
[13] For example, elevator music.
[14] Cooper, 43-44. One could argue that there is an arms race to be the first mover and first scaler with regard to content production.
[15] Mark Lemley, How Generative AI Turns Copyright Law Upside Down, 25 Science and Technology Law Review (2024), 37-38. Note that what we expect to see is similar but also different than what we know about designing around patents or patent trolling.
Zachary Cooper
Zachary Cooper is a legal scholar, lecturer and doctoral researcher for the Amsterdam Law & Technology Institute (VU Amsterdam), where he developed the course “The Law and Ethics of Robots and AI”
William Lehr
Telecommunications and Internet industry economist and consultant with over 30 years of experience. Research Associate at MIT.
Volker Stocker
Volker Stocker is an economist who specializes in the digital economy and the Internet ecosystem. Since May 2019, he has been leading the multidisciplinary research group “Digital Economy, Internet Ecosystem, and Internet Policy” at the Weizenbaum Institute (prior to September 2022, the group was named “Work and Cooperation in the Sharing Economy”).
