Diluting the Market Dilution Theory: Why Kadrey v. Meta Gets Market Harm Wrong
Introduction
The AI training and copyright debate has witnessed pivotal moments with summary rulings in Richard Kadrey v. Meta Platforms, Inc and Bartz v. Anthropic (now settled). In Meta, the court introduced a novel ‘market dilution’ theory, which suggests that AI-generated works, even if non-infringing, could flood the market and indirectly harm original authors. This concept has also made its way to India in ANI v. OpenAI, where it has been contended that digital news will be replaced with AI-generated content. These developments make it crucial to critically examine the theory of market dilution.
In this piece, I argue that market dilution is conceptually incoherent (not copyright-cognisable), practically self-defeating, and penalising of efficient technology. It conflates legitimate technological competition with infringement, and risks transforming copyright into a tool of protectionism rather than a mechanism to balance incentives and progress.
What happened in Kadrey v. Meta?
Facts
In Kadrey v. Meta, a group of authors alleged that Meta unlawfully copied their books to train its large language model (“LLM”) called LLaMA. They claimed this wholesale copying infringed their copyrights (pp. 1–2). Meta countered that training an LLM was transformative and therefore protected as fair use.
Court’s Analysis and Judgment
The Court upheld Meta’s use of copyrighted-works to train its LLM Llama as fair use under Section 107 of the Copyright Act. However, the judgment’s finding upholding fair use is fact-specific and not a general precedent validating such use. The Court observed that the “in many circumstances it will be illegal to copy copyright-protected works to train generative AI models without permission” thus requiring such works to be obtained after paying licensing fees (Page 4).
In reaching this conclusion, the Northern District of California applied the four fair use factors under 17 USC § 107: (1) purpose and character of the use, (2) nature of the copyrighted work, (3) amount and substantiality of the portion used, and (4) market effect. On the first factor, it found Meta’s use highly transformative: the books were copied not to reproduce them, but to extract patterns of language for a new and innovative purpose (p. 16). On the second factor, while the works were undeniably creative, the Court held that this carried less weight in light of the transformative and functional nature of the use (pp. 23–24). On the third factor, it noted that Meta’s copying of entire works was reasonable, given its transformative purpose (p. 25).
The fourth factor (market effect) received the most attention. The Court dismissed arguments that LLaMA could regurgitate books, finding no evidence of substantial reproduction (pp. 26–27). It also rejected claims of harm to a licensing market for AI training, warning against circularity in lost-licensing arguments (pp. 27–28).
At the same time, it advanced a new theory of market harm called market dilution, suggesting that AI-generated works, while not infringing, could flood the market with competing content in the same genres, disproportionately harming lesser-known authors and undermining incentives to create (pp. 28–32). While plaintiffs failed to present evidence of such dilution here, the Court suggested that this theory may often be decisive against fair use in future AI training cases (pp. 32–33).
Ultimately, Meta prevailed, but only because the plaintiffs had not substantiated market harm (p. 33).
What is Market Dilution?
The ‘market dilution’ theory focuses on copyright harm via indirect substitution rather than direct reproduction. It suggests that by enabling the mass generation of new, non-infringing works in the same genre/subject, LLMs could flood the market with competing works. It is argued that it could dilute demand for the originals, especially harming lesser-known authors/works, constituting a new kind of market harm. For example, an LLM trained on numerous romance novels could generate new novels in the same theme, competing with and potentially reducing demand for the originals.
The Court illustrated this concern by noting that while established authors may remain relatively unaffected, “lesser-known works or works by up-and-coming authors” would face greater harm. It termed this “harm of indirect substitution” citing Matthew Sag, who nevertheless acknowledges that such harm is not copyright-cognisable as it exceeds the scope of copyright law and extends protection against competition from indirect substitution.
The theory, in sum, posits that LLMs would rapidly generate works which “are similar enough (in subject matter or genre) that they will compete with the originals and thereby indirectly substitute for them”. These AI generated works are said to compete with the original works “even if those works aren’t themselves infringing”.
While the market dilution theory raises concerns about AI-generated content flooding creative markets, its application in copyright law is fundamentally flawed for several critical reasons.
Why Market Dilution Fails in Copyright Law
[1] Market Harm is Work Specific
The analysis of market harm under copyright law is to be assessed for that particular work as the rights are specifically bestowed upon each work (US Const, art I, § 8, cl 8, 17 USC § 106, India: Section 14 of the Copyright Act 1957). While the intention of copyright law might be to incentivise creation, such framework is work-specific and copyright law does not contemplate calculating market harm considering authors as a class. The distinction between market dilution and market harm is that the former considers authors as a class whose interest is sought to be protected whereas the latter analyses work-specific market harm caused by infringement.
Moreover, copyright law does not contemplate protecting authors against competition in themes/genres/styles/subject matters as a whole (17 USC § 102(b), Nichols v Universal Pictures (1930) and India: RG Anand v Delux Films (1978)). The ex-ante structure of copyright lies in the fact that the legal guarantee of certain rights incentivises the creation of works, even before they come into being. However, copyright remains a work-specific right, vesting only when the work is expressed in a tangible form (17 USC § 102(a) and India: Section 13 of the Copyright Act 1957).
The market dilution theory which calculates harm via indirect substitution and generation of competitive works by LLMs is not work-specific, but considers legitimate competition in similar genres/styles/subject matters including future works, and is thus beyond the scope of copyright law.
[2] Technology v. Infringement
The Court has stressed on the generative powers of LLMs noting that the “technology … can generate literally millions of secondary works, with a miniscule fraction of the time” with the “potential to flood the market with competing works”. However, the Court here does not seem to analyse the direct effect of alleged infringement but is instead focusing on the effect of the technology. The scope of copyright is to protect the interests of the authors, only against infringement of work-specific rights bestowed by the copyright law (17 USC § 106 and § 501, India: Sections 14 and 51 of the Copyright Act 1957), it cannot be used to target legitimate competition via disruptive/efficient technology. The Court’s emphasis on the impact of technology rather than the direct effect of alleged infringement in the training stage of LLMs, yet again stretches the scope of copyright law to target legitimate competition. Advancement of technology and efficiency leading to dilution in the market is akin to damnum sine injuria (damage without injury): there is no legal right flowing from copyright law being violated via such legitimate competition. Thus, copyright law cannot be used to accord authors absolute protection against technological advancement.
As Justice Alsup rightly points out in Anthropic, the authors’ contention that training LLMs will lead to an explosion of works competing with their original works is simply “no different than it would be if they complained that training schoolchildren to write well would result in an explosion of competing works”. He rightly notes that such competitive/creative displacement does not concern copyright law and it does not protect authors against competition. Justice Chhabria, in Meta, dismissed the applicability of this observation to market effect analysis while stating that LLMs can “generate countless competing works with a miniscule fraction of the time and creativity” which is distinct from an individual’s limited capacity. Justice Chhabria is, yet again, penalizing disruptive technology, rather than focusing on whether there is any actual copyright-cognisable substitution taking place.
Moreover, copyright is neither a guarantee against competition, nor a tool to freeze technological progress. The court’s rejection of Justice Alsup’s analogy conflates legitimate downstream competition with infringement-based market substitution. The Copyright Act protects authors from having their own works copied and sold as substitutes, not from having to compete with new works, whether those are written by future authors or generated with new tools. To treat technological efficiency as a copyright-cognisable market harm risks warping fair use into a doctrine that penalises efficiency rather than policing infringement.
[3] Market Dilution is Self-Defeating
Let us assume the AI companies seek to comply with such a framework and train their LLMs based only on licensed works, and are also able to afford the same. How would the same solve the problem which the Court puts forward? Flooding of the market with competitive AI-generated content would still continue, especially harming new/lesser-known works/authors, disincentivising them to produce works while licensing fees are paid only to authors of existing works. The exact problem which the Court puts forth via this theory would still continue even if AI companies comply with such a framework and pay licensing fees to train LLMs. The disincentives caused by legitimate competition are beyond the scope of copyright law, which cannot be used to penalise technological advancement.
Furthermore, if all authors embraced the rationale behind this theory, many would refuse to license their works to AI companies that would compete with their existing or future works. The alternative, compulsory licensing, would still lead, inevitably, to market dilution, albeit with added transaction costs. Thus, market dilution will take place sooner or later, even if AI companies choose to train models only on licensed works:
- Authors willing to license, AI companies willing to pay → market dilution
- Authors unwilling to license, AI companies willing to pay → compulsory licensing → market dilution
The only likely effect of this theory would be passing the costs onto the ultimate users of LLMs or delaying the development of efficient AI models, which might slow down, but cannot prevent, the creation of competitive works.
[4] Commercial Nature Should Not Influence Market Harm Analysis
The Court repeatedly emphasised the commercial scale of LLM training, noting it is “expected to generate billions, even trillions, of dollars for the companies that are developing them” and suggesting that these companies “will figure out a way to compensate copyright holders for it”. But how is the financial status of an entity relevant in objective analysis of infringement and fair use? Commerciality is relevant only under the first fair use factor where both Meta (page 16) and Anthropic (page 11) have acknowledged the highly transformative aspect of LLM training.
The US Supreme Court in Campbell observed that “the more transformative the new work, the less will be the significance of other factors, like commercialism”. Thus, the commercial nature of training LLMs should not matter much. The aforementioned observations of the Court seem to be a “fifth fair use factor” where the judge’s personal sense of right or wrong seep into fair use analysis (Rich Stim).
[5] Legitimate Competition vs. Infringement
Meta argued that the generation of competitive works by LLMs would constitute legitimate competition while relying on Sega v. Accolade (1992) and Sony v. Connectix [203 F.3d 596 (9th Cir. 2000)]. However, the Court, distinguished the cases and dismissed the argument noting that LLMs’ ability to generate text is directly related to the creative expression in copyrighted books, hence it is directly benefiting from such usage and thus, it cannot be considered “legitimate”.
The Court has misread Sega and Connectix, as those cases do not exclude competition that gains from expression. Rather, they exclude competition that copies expression itself i.e., expressive use.
In Sega v. Accolade, Accolade reproduced the code of Sega’s gaming console only to unearth its functional elements to develop games compatible with such console. Sega sued Accolade for infringing the copyright over its software code, however, the Court upheld Accolades copying as fair use and also observed that there exists a “distinction between the copying of works in order to make independent creative expression possible and the simple exploitation of another’s creative efforts” (Para 15). In its analysis of the fourth factor, the Court, observed that consequences are not attached “to a use which simply enables the copier to enter the market for works of the same type as the copied work … Accolade sought only to become a legitimate competitor in the field” (Para 15). The Court also notes indirect harm caused to Sega but does not take cognizance of the same (Para 16). Thus, the Court clearly does not observe against intermediate copying benefitting from copyrighted work, it merely observes against copying which exploits the creative efforts of an author (expressive use).
In Sony v. Connectix, Sony sued Connectix for copying its code in the process of reverse engineering Sony PlayStation’s built-in operating software to create an emulator that let Sony’s game CDs run on PCs. The Court while upholding fair use also observed that Connectix is a “legitimate competitor in the market” and the economic loss (market harm) suffered by Sony as a result of such competition does not compel a finding against fair use as copyright law “does not confer such a monopoly” (Para 43). Connectix is clearly benefitting from intermediate copying; however, the Court does not rule against such gain from copyrighted work.
Therefore, the Court has clearly misread and misapplied Sega and Connectix.
Furthermore, training LLMs leverages creative expressions in a non-expressive manner, making the resulting competition legitimate under fair use. While training involves “reading” expressive works, the purpose is not to reproduce or enjoy the creativity but to extract structural/functional patterns of language (grammar, syntax, frequency, semantic relations) [Shivam Kaushik, Part 1]. Copyright protection cannot extend to patterns and statistical relationships within the text, the protection is limited to that of original expression [17 USC § 102(b) and Shivam Kaushik, Part 2].
Thus, the Court has not only misread aspects relating to legitimate competition in Sega and Connectix but also sought to protect functional elements which lack copyright protection, thereby mischaracterising legitimate downstream competition as infringement-based market substitution.
Conclusion
The market dilution theory in Kadrey v. Meta is simply beyond the scope of copyright law. Copyright protects individual works, not entire genres, styles, or future authors. It goes beyond the work-specific protection (against direct substitution) accorded in copyright law and considers authors as a class while seeking to protect non-copyrightable elements like genres/subject matters and functional/statistical elements within the works. Application of this theory in the context of LLMs would translate to penalizing efficient/disruptive technology and cracking down on legitimate competition. Moreover, even if AI companies were required to license copyrighted works, non-infringing AI-generated content would continue to flood the market, leaving the purported “harm” unaffected and making the theory practically self-defeating. The Court’s emphasis on commercial scale, efficiency, and indirect substitution risks warping fair use into a mechanism for protectionism rather than a balanced framework that incentivizes creativity while fostering technological progress.
Lastly, it must be recognised that creative destruction (innovations and technologies replace outdated ones) has always been integral to competition in capitalist markets (Joseph Schumpeter p.83 and Raymond Ku). Each disruptive technology whether the printing press, radio, broadcast media, VCRs and DVDs, the internet, or now generative AI has reshaped the contours of copyright law. At its core, copyright law accepts static inefficiencies to secure long-term dynamic efficiencies by incentivising creativity. Generative AI has the potential to enhance both static and dynamic efficiencies, though it inevitably increases legitimate competition for authors which is not copyright-cognisable harm. In this sense, embracing AI and allowing such creative destruction to recalibrate copyright norms and market structures better aligns with copyright’s fundamental objectives of promoting such dynamic efficiency and increasing the propagation of creative works.
[*The Author is a fourth year B.A. LLB. (Hons.) student at the National Law School of India University (NLSIU), Bangalore.]