The Three Judgments That Matter for AI and Copyright

The debate on generative artificial intelligence and copyright in India assumes that using copyrighted works to train machine learning models constitutes infringement, requiring justification through Section 52. Copyright owners proceed on this basis when they demand compensation or prohibition. Generative AI developers proceed on this basis when they invoke defenses. Courts in the United States have been asked to determine whether AI training qualifies for fair use protection under Section 107 of the US Copyright Act, with outcomes that remain uncertain. The premise underlying all of this – that AI training falls within the copyright holder’s exclusive rights and requires justification – has gone largely unexamined.

Under the Indian Copyright Act, 1957, using copyrighted works to train a large language model is not copyright infringement. The rights conferred under Section 14, both the reproduction right and the adaptation right, do not extend to uses that produce something of fundamentally different character from the original. Training is one such use. The question of fair dealing under Section 52(1)(a) does not arise because there is no infringement to excuse. This is not a novel proposition. Rather, it follows from doctrine that Indian courts have already established in a line of cases concerning guidebooks, adaptations, and derivative works. These cases simply have to be applied to the AI context.

The proposition rests on three judgments: Barbara Taylor Bradford v. Sahara Media Entertainment Ltd, Syndicate of the Press of the University of Cambridge v. B.D. Bhandari, and E.M. Forster v. A.N. Parasuraman. Together, these decisions establish that transformative uses fall outside the reproduction right and the adaptation right as defined under Indian law. The limitation is embedded within Section 14 itself, not located in Section 52 as an exception or defence.

Section 14 of the Copyright Act defines copyright as the exclusive right to do or authorize the doing of certain acts “in respect of a work or any substantial part thereof.” There is a limitation built into the right itself. First, the reproduction right, including storage, does not extend to every act that makes use of the work or draws upon it in some way. It extends to acts that result in a copy of the work itself or a substantial part of it, reproduced in material form. An act that draws upon a work but does not result in a copy of the work or any substantial part thereof does not engage the reproduction right. Storage of a work, for the purposes that do not amount to reproduction of the work would also, in a similar vein not amount to infringement, especially when the legislature did not deem storage to be an independent right but merely a subset of reproduction. Second, the adaptation right is defined with reference to specific categories set out in Section 2(a). Adaptation means, in relation to a literary work, the conversion of the work into a dramatic work; in relation to a dramatic work, the conversion into a non-dramatic work; the conversion of a literary or dramatic work into a cinematograph film; any abridgement of the work or any version in which the story or action is conveyed wholly or mainly by means of pictures; and – importantly – any use of such work involving its rearrangement or alteration, which was introduced via the 1995 amendment, which although intended to expand the meaning of adaptation, however, cautiously has been read down subsequently by judgments to only extend to trivial alterations. The adaptation right does not extend to every transformation of a work. It extends only to the specific categories enumerated.

This statutory structure differs materially from United States law, and the difference is critical for understanding why the Generative AI-copyright question has a simpler answer in India. Under Section 101 of the US Copyright Act, a “derivative work” is defined as “a work based upon one or more preexisting works, such as a translation, musical arrangement, dramatization, fictionalization, motion picture version, sound recording, art reproduction, abridgment, condensation, or any other form in which a work may be recast, transformed, or adapted.” The critical words are the catch-all: “any other form in which a work may be recast, transformed, or adapted.” The copyright owner’s exclusive right to “prepare derivative works based upon the copyrighted work” under Section 106(2) thus includes transformed versions of the work. India complied in letter and spirit with the mandates of TRIPS, whereas the American definition goes beyond TRIPS.

This creates what appears to be an oxymoron at the heart of American copyright law. The derivative work right includes “transformed” works within its scope, meaning transformation falls within the copyright holder’s exclusive rights. Yet the first fair use factor under Section 107 asks about the purpose and character of the use, and courts since Campbell v. Acuff-Rose Music have held that transformative uses – uses that add something new, with a further purpose or different character – are more likely to qualify for fair use protection. How can transformation simultaneously fall within the copyright holder’s exclusive rights and favour the defendant in fair use analysis? American courts navigate this tension case by case, asking whether a particular transformative use is sufficiently justified to qualify for the fair use defence despite falling within the scope of the derivative work right.

The Indian Copyright Act has no such tension because its statutory structure is fundamentally different. Section 2(a) defines “adaptation” through specific enumerated categories with no catch-all provision extending the adaptation right to “any form in which a work may be transformed.” The scope of the final category – “rearrangement or alteration” under Section 2(a)(v) – is critical. One might argue that “alteration” is broad enough to capture any transformation, including radical ones. But as Barbara Taylor Bradford holds, this interpretation would “make nonsense of the Indian Copyright Law.” The Court examined whether “alteration” could bear the meaning of extreme or radical transformation and concluded that it could not. The word “alter” in Section 2(a)(v) means “minor change, slight change, not making the original something beyond recognizable possibilities.” Transformation – use of the work to create a work of different character or purpose – falls outside the scope of the adaptation right. This is the opposite of the American position, where catch-all language explicitly brings transformed works within the copyright holder’s exclusive rights.

The consequence of this statutory divergence is that transformation and infringement are co-extensive in American law but not in Indian law. In the United States, if a use transforms the original work, that use falls within the derivative work right because derivative works include transformed works by definition. The question then becomes whether fair use excuses what would otherwise be infringement. In India, if a use transforms the original work into something of fundamentally different character, that use falls outside the adaptation right altogether because the adaptation right does not extend to transformations. The question is not whether fair dealing under Section 52 excuses infringement but whether there is infringement at all.

The Barbara Taylor Bradford judgment arose from an allegation that the television serial “Karishma” infringed the copyright in the novel “A Woman of Substance.” The creative director of the serial had stated in an interview that it was “based on” the book and had been “Indianised.” The plaintiffs argued that this admission established infringement. They relied particularly on Section 2(a)(v), which includes within the definition of adaptation “any use of such work involving its rearrangement or alteration.” The plaintiffs contended that even if the serial was substantially altered from the novel, this alteration itself constituted adaptation within the meaning of the statute.

The Division Bench of the Calcutta High Court rejected this interpretation. Justice Ajoy Nath Ray, writing for the Court, held that such an interpretation “would make nonsense of the Indian Copyright Law. A totally changed thing can never be termed a copy of the original thing. How can copyright affect the right in something, which is not related to the protected work’s copying or reproduction at all?” This passage addresses both rights. With respect to the reproduction right, a totally changed thing cannot be termed a copy. With respect to the adaptation right, copyright cannot affect something which is not related to copying or reproduction. The Court concluded that “the large change meaning cannot be ascribed to the word ‘alter’ in Section 2(a)(v) of the Copyright Act, 1957, because it renders the interpretation absurd.” This interpretation confines the adaptation right to its proper scope: it covers rearrangements and alterations that maintain a recognizable connection to the original, not transformations that produce something of fundamentally different character.

The B.D. Bhandari judgment is the most instructive of the three because of its analytical structure. The case concerned guidebooks published by the defendants based on textbooks published by Cambridge University Press. The textbooks were prescribed for students in certain courses, and the guidebooks provided answers, explanations, and step-by-step solutions to the questions and exercises contained in the textbooks. The copyright owners sued for infringement. The Single Judge dismissed the suit on multiple grounds: that no copyright existed in questions and grammar exercises, and that even if copyright existed, the guidebooks fell within Section 52 as “review” under Section 52(1)(a) and under Section 52(1)(h), which provides that reproduction by a teacher or pupil in the course of instruction does not constitute infringement.

The Division Bench reversed the Single Judge on the first point: questions and grammar exercises do merit copyright protection. The framing of questions involves creativity and skill, and works do not lose copyright protection merely because they deal with subjects that admit of only limited modes of expression. This was a significant finding for the copyright owners – it meant the guidebook publishers could not escape liability on the ground that there was nothing to infringe. The Division Bench then addressed Section 52(1)(h) and rejected it decisively: “We find force in the aforesaid submissions of the appellant and hold that Section 52(i)(h) of the Act would not apply.” This was as the said section pertains to in course of instruction for pupils/students. What happens next in the judgment is critical to understanding Indian copyright doctrine. Having rejected Section 52(1)(h), one might expect the Division Bench to proceed to Section 52(1)(a) – the “review” exception that the Single Judge had relied upon. But the Division Bench did not take that step. Instead, it moved to an entirely different question: whether guidebooks, being transformative works of different character, infringe the copyright at all.

The Division Bench defined a guidebook as a work meant to help, assist, and support students to understand problems given in a prescribed publication – providing explanation, step-by-step processes, fine points, and detailed analysis. It is a kind of commentary based on the original work to enable students to give effective answers to questions set in examinations. It then articulated the governing principle: when a published work is a textbook, a guide can be published provided it is materially different from the textbook. If a guidebook reproduces the original to a great measure without its own contribution, it may amount to infringement. But if a guidebook is written in a different format, dealing with subject matter contained in the original but with a purpose to help students, then it is a work different from the original. The critical holding: “If the guidebook has assumed different character, it would not amount to infringement of the original work, though it would have been drawn, to certain extent, from the original work.”

The analytical sequence must be understood precisely. The Division Bench upheld copyright in the textbooks – there was something to protect. It rejected Section 52(1)(h) decisively – commercial publishers are not teachers or pupils. At this point, one Section 52 defence has been eliminated. If transformative use operated as a defence under Section 52, the Court would have proceeded to Section 52(1)(a) to determine whether another defence was available. If no Section 52 defence applied, the conclusion would be infringement. But that is not what happened. The Division Bench rejected Section 52(1)(h) and then found no infringement without relying on Section 52(1)(a) at all. This establishes that transformative use analysis operates at Section 14, determining whether infringement exists in the first place, not as a defence under Section 52 that excuses infringement that would otherwise exist.

The E.M. Forster judgment confirms this understanding in the clearest possible terms. The case concerned a guidebook for university students based on “A Passage to India.” The author of the guidebook “was concerned to aid students taking the B.A. degree, in answering the several questions that would no doubt be set for them, upon the novel, in the examination papers for that Degree.” The guidebook consisted of chapter summaries, character analyses, critical essays, and probable questions with suggested answers. The copyright owners sued for infringement.

The Madras High Court’s finding on infringement was unequivocal: “It is abundantly clear that, either upon the principle of substantial reproduction, or even qualitative reproduction, as far as the verbal expression of the original work goes, there has been no infringement.” The guidebook dealt with plot, themes, and characters, but in its own words and analysis. The Court emphasized the distinction between protected form and unprotected content: “The idea could well be taken up by others and treated in their own way. But even after the work has found its verbal form, it is the form, which is protected, and not the theme, plot and characters.” The guidebook’s author “cannot be convicted” of reproducing the original, “and his guide-book remains distinctly his creation, even if it be regarded as an abridgment of the novel, in part, and, in part, a running commentary upon it.”

Having found no infringement, the Court addressed fair dealing in a single sentence that encapsulates the doctrinal point: “The question of ‘fair dealing’ does not arise, because of the failure of the appellants to make out a case of infringement prima facie.” Section 52 is a defence to infringement. It presupposes that infringement exists and provides a justification for what would otherwise be unlawful. Where there is no infringement – where the use falls outside the scope of Section 14 – the question of fair dealing does not arise. The guidebook was a transformative work of different character. Creating it did not engage Section 14, and Section 52 was irrelevant.

These three judgments, read together, establish a coherent doctrine that resolves the AI question. The reproduction right extends to reproduction of the work or any substantial part thereof – not to uses that draw upon the work but result in something different. The adaptation right extends to specific enumerated transformations that maintain recognizable connection to the original – not to transformations producing something of fundamentally different character. Transformative uses fall outside both rights. They are non-infringing as a matter of Section 14’s scope. The limitation is embedded within the rights, not located in Section 52 as an exception.

The application to Generative AI training follows directly from this doctrine. Training a large language model is transformative in the clearest sense. The process does not produce a copy of the works used. It produces something entirely different: a statistical model consisting of mathematical parameters – numerical weights and biases – that encode patterns learned from training data. The trained model does not contain the works. It does not store them. It does not reproduce the expression of any work or any substantial part of any work. What the model contains is a mathematical function that can generate text based on patterns it has learned, not copies of the texts from which it learned those patterns. What is extracted and encoded are unprotectable elements (statistical regularities) not protected expression, reinforcing the idea/expression distinction.

The guidebook cases provide the closest analogy to Generative AI training and illuminate why the doctrine applies. A guidebook draws upon a textbook to create something serving a different purpose. It uses questions, structure, and subject matter from the original, but adds explanation, analysis, step-by-step reasoning, and commentary. The result is a work of different character – not a substitute for the textbook but a complement to it, serving a different function. Under B.D. Bhandari and E.M. Forster, such guidebooks do not infringe – even when no Section 52 exception applies – because they fall outside the scope of Section 14.

Generative AI training is more transformative than guidebook preparation by any measure. A guidebook shares the medium of textual expression with the original work; a trained model does not – it is a mathematical function, not a text. A guidebook addresses subject matter related to the original work; a trained model does not address any particular subject matter – it encodes statistical patterns across the entire training corpus. A guidebook retains recognizable connections to the original: it refers to specific questions, discusses specific chapters, analyses specific themes. A trained model retains no such connections to any individual work. If guidebook preparation – which retains recognizable connections to the original – does not infringe, then Generative AI training – which retains none – cannot infringe a fortiori.

An objection may be raised concerning outputs. Even if training is transformative, outputs generated by AI models may compete in the same markets as works in the training corpus. A model trained on novels may generate text that competes with novels. Does this market competition engage the copyright holder’s rights and render the training process infringing?

This objection conflates two distinct questions. Whether the training process infringes must be analysed separately from whether particular outputs infringe? Training produces a model – not a copy of any work. Whether training infringes must be determined by examining what training produces, not by speculating about what outputs the model might later generate. Whether particular outputs infringe must be determined output by output: if a specific output reproduces substantial expression from a specific work, that output may infringe. But the possibility that some outputs might infringe does not make training itself infringing.

Moreover, market competition through works of different character is not copyright infringement. The B.D. Bhandari Court found no infringement even though guidebooks might reduce demand for textbooks. Copyright protects expression, not market share. It grants exclusive rights over copying and adaptation, not exclusive rights over markets or subject matter. Competition through transformation – through the creation of works of different character—is the ordinary and lawful operation of a market in creative works.

This analysis coheres with the foundational principle that copyright protects expression, not ideas. AI training extracts patterns and statistical relationships – features closer to ideas than expression. The trained model does not contain the expression of underlying works; it contains parameters encoding regularities. The Barbara Taylor Bradford Court observed that there is no monopoly in themes and patterns, because such a monopoly would prevent the free exercise of artistic skill. If there is no monopoly in themes and patterns, then extracting such patterns computationally cannot infringe copyright. The output may compete with the original work, but fair competition of different expressions has never been a copyright concern – copyright promotes such production. Labour displacement through creative disruption cannot be addressed through copyright law, which concerns a specific unit of production and cannot in its doctrinal form address broader economic concerns.

The contrast with United States law illuminates what is distinctive about the Indian position. Under US law, catch-all language means that transformed content falls within the copyright holder’s exclusive rights. Litigation over AI training has centered on whether training qualifies for fair use, with courts examining the purpose and character of the use, the nature of the copyrighted works, the amount used, and the effect on the market. Defendants must establish that fair use excuses what would otherwise be infringement. The outcomes remain uncertain because fair use is a fact-intensive determination, and because the derivative work right and the transformative use doctrine pull in opposite directions. India need not follow this path because the doctrinal architecture is different, and we must acknowledge that.

* Akshat Agrawal is the founder of AASA Chambers and is currently a PhD candidate at the University of Cambridge after being awarded the Squire Postgraduate Studentship in Law. The author would like to thank Sneha Jain for innumerable discussions and guidance.