Main

Following the discussion on Judge Alsup and Judge Chhabria’s findings in Anthropic and Meta cases respectively on lawfulness of the acquired copy, Akshat Agrawal discusses their findings on market dilution and assesses these findings from the lens of the Indian Copyright law. Akshat is a practicing litigator working at Saikrishna and Associates. He did his LLM from Berkeley Law in 2023 specialising in IP and Tech law. His previous posts can be found here. He adds the following disclaimer: After some discussion around an earlier draft and an admitted history of verbosity, I would also like to acknowledge the usage of Claude.ai for helping me re-frame the draft more succinctly and in a reader friendly manner. Views expressed here are personal.

By Akshat Agrawal

Continuing from my previous post examining the first divergence between Judge Alsup (Anthropic’s case) and Judge Chhabria (Meta’s case) in the Gen-AI copyright debate, this post analyzes their second divergence: on the “effect on the market” fair use test – Whether Gen AI training using copyrighted works, which generates content that competes with (need not necessarily be substantially similar) human works constitutes a relevant copyright market harm? Judge Chhabria believes such “market dilution” undermines creative incentives; Judge Alsup dismisses it as an irrelevant  pro-competitive result of Gen-AI, beyond copyright’s scope. I discuss which view is more susceptible to be adopted under Indian law.

As discussed in Part 1, US fair use doctrine has a four factor test: (i) purpose and character of use, (ii) nature of the copyrighted work, (iii) amount and substantially of use and (iv) effect of the use on the potential market for or value of the work. The fourth fair use factor examines market substitution—whether the challenged use adversely affects a relevant copyright market. Judge Chhabria in Meta, citing Harper v. Row, calls this the single most important factor. Yet this ignores Campbell v. Acuff-Rose and Google v. Oracle,where the US Supreme Court held that both the amount and the “source” of the effect on the market matter: does the harm stem from copyright-cognizable injury or from permissible transformative use?

In any case, both Judge Alsup in Anthropic and Judge Chhabria in Meta rejected two Plaintiff’ theories of market harm:

The judges however have diverged on a third theory. Authors contend that training LLMs floods markets with competing high-quality works, undermining incentives for human creation. Note: this concerns not competing substantially similar works, but works in the same genre, style, or subject matter. The argument’s core: Copyright incentivizes human creativity which is being undermined by Gen-AI creation, which itself is dependent on human creativity for its capability, but does not compensate those humans. It is like learning without paying your teachers.

Is such harm copyright’s concern? Both Judges offer distinct views.

In Meta, Judge Chhabria finds fair use, while berating the Authors’ lawyers for not advancing an argument that he believes to be the best. He believes training on copyrighted works “dramatically undermines” human creative incentives, regardless of output similarity. Mechanically generated content at unprecedented scale competes with human works without compensation, dissuading traditional creation — and resulting in market dilution. Using copyrighted works enhances AI output quality, intensifying competition. He terms this “indirect market substitution.” He emphasizes that markets of human works become vulnerable due to increase in indirect competition from Gen-AI outputs. But is increased competition (without substantially similar copying) a copyright concern? In any case, as, the plaintiffs never advanced this argument, all this passionate obiter fills the judgment despite its irrelevance to the pleadings.

In Anthropic, authors did raise this argument. Judge Alsup, however, took a somersault. He dismissed Judge Chhabria’s argument as an irrelevant pro-competitive concern beyond copyright’s scope. Copyright provides incentives only against substantially similar copying, not against all kinds of competition. Competitive displacement without substantial similarity isn’t copyright’s domain.

Judge Chhabria assumes copyright solves all creative sector concerns, ignoring the Act’s specific limitation to imitative copying—not pro-competitive production. In any case, this rant by Judge Chhabria was unneeded, once it was irrelevant to the pleadings of the case before him. Similar to how it was irrelevant for Judge Alsup to go into training using works downloaded from shadow libraries as those facts were not ripe before him (see: Part 1).

How would this argument of Judge Chhabria fare under Indian law?

ANI argues Gen-AI trained using ANI’s news articles would result in market displacement without any compensation, echoing Judge Chhabria. Indian law, however, would likely follow Judge Alsup.

In S.K. Dutt v. The Law Company, the Allahabad High Court allowed a competing legal commentary to use extracts from plaintiff’s work, holding: Every person can take what is useful from the original work, improve, add, and give to the public the whole comprising the original work with additions and improvements. Infringement requires slavish copying or colorable imitation, not mere competitive use.

Similarly, in Akuate Internet v. Star India, the Delhi High Court rejected an exclusive broadcaster’s claim over cricket score updates. The Delhi High Court held that using ideas or facts from exclusive content doesn’t create quasi-property rights merely to prevent competition.

Indian law is clear: if defendant’s work (i) merely uses plaintiff’s work without expressive reproduction and (ii) competes with it, no infringement occurs. Using works to produce pro-competitive transformative outputs isn’t infringing, merely because it is a lot of indirect competition.

Judge Chhabria’s concern about human creative displacement isn’t copyright law’s province. It may be a concern better addressed, if needed, using other policy instruments – for instance a human author cess, or a tax on Generative AI development that would be redirected towards human authors. However, expanding the boundaries of copyright to include exclusion of non-substantially similar output (or training that leads to it) would have detrimental spillover in all other- human to human contexts as well. In other words, if you don’t pay your teacher who taught you certain themes, it doesn’t amount to copyright infringement unless you copy the specific expression taught by your teacher– in a competitive landscape and in a substantially similar manner. It may be unethical or a labour misappropriation issue, but it has nothing to do with copyright law.

Both decisions’ most problematic aspect lies in analyzing Factor 2: Nature of the Work and its Use.

Judge Alsup, in Anthropic, correctly identifies the inquiry of Factor 2—revealing differences between the original work and its secondary use, and the relationship between the taking and its use—but misapplies it without adequate consideration.

However, Judge Chhabria, in Meta, equates using/ collecting and reproducing statistical data about word patterns, frequencies, and relationships in a work, with reproducing its creative expression itself. He claims reproduction of these relationships equals reproduction of expression. This conflates learning grammar and writing mechanics with reproducing creative expression. Copyright protects skill and judgment in communicative content, not grammatical choices. These are entirely distinct.

Learning statistical patterns differs fundamentally from copying expressive content—the former constitutes unprotectable ideas or methods. Judge Alsup, separately, recognizes this: “Claude has outputted grammar, composition, and style that the underlying LLM distilled from thousands of works. But if someone were to read all the modern-day classics because of their exceptional expression, memorize them, and then emulate a blend of their best writing, would that violate the Copyright Act? Of course not. Copyright does not extend to “methods of operation, concept,  principles illustrated or embodied in a work” This core feature of the idea-expression distinction in Copyright doctrine escapes Judge Chhabria in his quest to shield humans from AI competition.

But Judge Alsup’s finding above lends another interesting question:

If Gen-AI tools use only methodical elements (like word patterns and grammar techniques or syntax synthesizing techniques) from works they are trained on, why jump to the question of fair use at all? Copyright (in its scope) “does not extend” to methods of operation. Isn’t this a threshold scope of rights question, that requires consideration prior to establishing infringement? I shall leave you with that!