Training Generative AI on Copyrighted Works: Fair Dealing or Infringement under Indian Copyright Law

Authored By: Roshni.R

Amity University Dubai

Introduction

In July 2026, the Delhi High Court refused to grant an interim injunction against OpenAI. ANI Media, one of the country’s largest news agencies, had asked the court to stop the company from storing or using its news articles to train its large language models. The court declined. On a prima facie basis, it held that scraping and storing copyrighted content to train artificial intelligence could fall within the fair-dealing exception under Section 52(1)(a) of the Copyright Act, 1957¹. The ruling was India’s first real judicial word on how generative AI and copyright fit together. It also left the hardest question, whether commercial, large-scale training of AI on protected works can really sit inside an exception meant for criticism and review, open for trial.

The stakes are not small. Generative AI systems learn from enormous datasets, and those datasets often contain copyrighted works scraped from the open internet. Writers, publishers, and news organisations see this as unauthorised reproduction carried out at industrial scale, with no payment and no permission. Developers push back. Training, they say, is not expressive; the model does not reproduce the work for readers, it learns statistical patterns from it. Indian copyright law does not presently have a clear answer to who is right. There is no dedicated text-and-data-mining exception to point to.

This article argues that the Delhi High Court’s prima facie reading of Section 52(1)(a) is understandable but fragile. Fair dealing was built around human uses, such as criticism, review, and private study, not around the commercial training of models later licensed to third parties. Stretching it to cover AI training tests the provision beyond what it was meant to bear. The better course is for Parliament to step in with a dedicated exception that permits training while preserving a workable opt-out and some form of remuneration for rightsholders. What follows sets out the existing framework, traces the case law, weighs the gap, compares India with the European Union, the United Kingdom, and Japan, and offers proposals for reform.

The Existing Legal Framework

The Bundle of Exclusive Rights

The Copyright Act, 1957 gives copyright owners a bundle of exclusive rights, including reproduction, issuing copies, public performance, and communication to the public.² Section 14 defines these with reference to each class of work. The reproduction right is the one most directly in play when AI training is discussed, because feeding a work into a dataset involves copying it, even where the copy is temporary and never shown to anyone. Copy without permission and you infringe, unless an exception covers you.

Infringement and the Fair-Dealing Exception

Section 51 is the infringement provision. Copyright is infringed where a person, without a licence, does an act that only the owner has the right to do.³ The exception that matters most here is Section 52, which lists acts that do not count as infringement. Section 52(1)(a) allows fair dealing with a literary, dramatic, musical, or artistic work for certain purposes, including private or personal use, research, criticism, and review.⁴ The crucial feature is that fair dealing in India is a closed list. Unlike the fair-use doctrine in the United States, which lets courts weigh factors and reach a judgment, an Indian court cannot simply say a use is fair because it seems fair. The use has to fit one of the listed purposes.

What the Act does not have is an exception for text and data mining. There is no provision saying, in terms, that copying a lawfully accessed work to train a computational model is permissible. That absence is the gap the whole dispute turns on. Whether Section 52(1)(a) is broad enough to be stretched over machine learning is, in effect, the question the Delhi High Court has been asked to decide.

Authorship of AI-Generated Output

A second question sits alongside the first: authorship. Section 2(d) defines the author of a work by reference to what kind of work it is, but it says nothing about works produced with substantial machine involvement.⁵ The provision assumes a human author. On the authority available, Indian law has not accepted that an AI system can be an author. What that means for output generated largely by AI is unsettled, and the case law below shows why.

III. Case Law Analysis

ANI Media v OpenAI: The Training Question

ANI Media sued OpenAI in November 2024. Its claim was that OpenAI scraped, stored, and used ANI’s copyrighted news articles, without permission, to train the models behind ChatGPT, and that ChatGPT’s outputs reproduced ANI’s works.⁶ ANI asked for an interim injunction to stop OpenAI from continuing to store or use the articles.

In July 2026, the court refused interim relief. It accepted, prima facie, that scraping and storing copyrighted content to train large language models could be fair dealing under Section 52(1)(a). The reasoning was that the storage served the purpose of training a model rather than directly exploiting the underlying works.⁷ The court also found that Indian courts had territorial jurisdiction, despite OpenAI’s servers sitting abroad. The order matters less for what it finally decides than for what it signals, a willingness, at the interim stage, to read the existing exception as wide enough to cover AI training. The substance is left for trial.

The Raghav Dispute: Authorship of AI Output

The authorship question has been tested through the Raghav dispute. An application was filed to register an artwork titled Suryast, naming the RAGHAV artificial intelligence painting application as co-author alongside its human developer. The Copyright Office initially accepted the registration, the first time an AI had been listed as a co-author in India. It then moved to withdraw it.⁸ In December 2023, the Copyright Review Board affirmed the rejection, confirming that a work needs a human author to qualify for protection.⁹ The episode is a useful window into the ambiguity around AI-generated works under the current statute, and it confirms that, for now, Indian law wants a human author.

Critical Evaluation

The ANI ruling is a pragmatic answer to a genuinely hard problem. The court could not push the question off until Parliament legislated, because the technology is already in wide use. Reading Section 52(1)(a) as capable of covering training avoids the alternative, which would be holding that nearly all AI training in India is infringing until Parliament catches up. That would be disproportionate, and it would push development to other jurisdictions.

The reading is fragile all the same, for three reasons. First, the purposes listed in Section 52(1)(a) are tied to human uses, such as private study, research, criticism, and review. Commercial training of a model later licensed to third parties fits awkwardly with any of these. The court’s reasoning treats storage as the operative act and frames the purpose as training rather than exploitation, but that reframing pushes the provision close to its limit.

Second, India’s fair-dealing framework is closed. Where the United States lets courts weigh the purpose of the use, its character, and its effect on the market, Indian courts cannot simply balance equities and call a use fair. If it does not fit a listed purpose, it is not fair dealing. Stretching Section 52(1)(a) over AI training risks blurring that line and turning a closed exception into a flexible one, one case at a time.

Third, the approach leaves rightsholders with little to work with. If training is fair dealing, then the copying that happens during training is not infringement, and rightsholders have no statutory footing on which to demand licensing fees or to opt out. The absence of an opt-out is what most separates India from the European Union. It removes a tool that could have balanced innovation against the interests of creators. The law, as it stands, answers the infringement question by effectively removing the infringing character of the act, and offers nothing in its place to the people whose works were used.

Comparative Perspectives

The European Union

The European Union has dealt with text and data mining head-on. Article 4 of Directive (EU) 2019/790 allows reproductions and extractions of lawfully accessed works for text and data mining, but it gives rightsholders an opt-out, one they can express through machine-readable means or, where appropriate, through terms of service.¹⁰ The opt-out was a deliberate choice. It lets broad text-and-data-mining activity support innovation while giving rightsholders a workable way to reserve their works. It is not above criticism; the technical and legal uncertainty around how to opt out has blunted its practical effect. But it is a more balanced settlement than the all-or-nothing choice Indian law currently offers.

The United Kingdom

The United Kingdom has a narrower exception. Section 29A of the Copyright, Designs and Patents Act 1988 permits copying for text and data analysis, but only for non-commercial research, and only where the person already has lawful access.¹¹ The UK government at one point floated broadening this to cover commercial uses. It reversed course in 2023, leaving the narrow exception in place. The episode shows how hard it is politically to widen such exceptions against creator opposition, and it shows that a research-only exception leaves commercial AI development without clear cover.

Japan

Japan has gone furthest. Article 30-4 of its Copyright Act, amended in 2018, allows the use of works for information analysis, including AI development, without the rightsholder’s permission, with limited exceptions for works enjoyed for the value of the work itself.¹² Japan’s framework is often cited as the model for jurisdictions that want to promote AI. It is permissive in a way India’s fair-dealing provision is not, because it is a purpose-built exception rather than a creative reading of a provision designed for other ends.

Conclusion

The Delhi High Court’s refusal to restrain OpenAI on an interim basis reflects a real difficulty: fitting AI training within a copyright framework that was written before the technology existed. The court reached a defensible result. But it rests on a reading of Section 52(1)(a) that strains the provision past what it was meant to cover. The Raghav dispute, for its part, confirms that Indian law requires human authorship, leaving the status of AI-generated output unsettled.

This article has argued that the existing framework is not up to the task. Fair dealing was not designed to govern commercial, large-scale machine learning, and reading it as if it were risks producing arbitrary outcomes and stripping rightsholders of any leverage.

Parliament should enact a dedicated text-and-data-mining exception, modelled on Article 4 of the EU Directive, that permits copying of lawfully accessed works for AI training, recognises a workable opt-out for rightsholders, and addresses remuneration for uses that have not been opted out. It should also put the authorship question beyond doubt by providing that copyright subsists only in works of human authorship, so that the issue is settled by statute rather than by administrative reversal. The court has shown the need. It is for Parliament to supply the answer.

Footnote(S):

ANI Media Pvt Ltd v OpenAI OpCo LLC, Delhi High Court, 24 July 2026. https://www.livelaw.in/high-court/delhi-high-court/delhi-high-court-openai-ani-copyright-infringement-interim-relief-chatgpt-542746

ANI Media v OpenAI (n 1).

Ibid.

‘India recognises AI as co-author of copyrighted work’ (2020); see also Copyright Office proceedings concerning the artwork Suryast. https://www.ipwatchdog.com/2020/11/21/exclusive-india-recognizes-ai-co-author-copyrighted-work/id=127646/

Copyright Review Board decision on Suryast, December 2023. https://www.livelaw.in/copyright-office-rejects-bid-register-artwork-co-authored-ai-240224

Directive (EU) 2019/790, art 4. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32019L0790

Bibliography

Cases

ANI Media Pvt Ltd v OpenAI OpCo LLC, Delhi High Court, 24 July 2026.

Re Suryast (Raghav Artificial Intelligence Painting App), Copyright Office of India (2020; withdrawn 2023).

Legislation

Directive (EU) 2019/790 of the European Parliament and of the Council on copyright and related rights in the Digital Single Market, art 4.

Secondary Sources

P Narayanan, Copyright and Industrial Designs (3rd edn, Eastern Law House 2020).

L Bently and B Sherman, Intellectual Property Law (5th edn, Oxford University Press 2014).

World Intellectual Property Organization, WIPO Conversation on Intellectual Property and Generative AI (2023).

Authored By: Roshni.R

Amity University Dubai

Related Posts

Leave a Comment Cancel Reply