Who Owns the Harvest? AI Training, Fair Dealing, and the Uncharted Limits of Section 52

This article is one of the winning entries from Academic IP Unveil 6.0 organised by Symbiosis Law School, Hyderabad.

A critical analysis of whether the ingestion of copyrighted content to train large language models constitutes “fair dealing” under the Copyright Act, 1957, through the lens of ANI Media (P) Ltd. v. OpenAI Inc.¹

A farmer has spent years developing a piece of farmland and harvests, prepares food from the wheat she produces, and then sells that food. One day a large factory shows up and takes her entire harvest without any authority to do so, then makes a different product with the wheat using their machinery, and sells it for profit and does not give any amount to the farmer. When she complains about her wheat being stolen, she is told “you have no more wheat, we are now selling our own product”.

This is the basis for the case of ANI Media², which is the first lawsuit against a generative artificial intelligence (AI) system for copyright violation by way of excessive use of a large body of text-based copyrighted works to train a large language model (LLM).

This case raises a completely novel question in that the Indian Courts have never ruled on whether or not the ingestion of a large body of copyright protected materials for the purpose of developing an LLM constitutes copyright infringement under the Copyright Act, 1957 (the Act); and if so, can OpenAI escape liability via “fair dealing” under Section 52?

In this blog, the author contends that Section 52 currently provides no space for AI commercial trainings. More importantly, they stress that when a court rules on this issue, a whole lot of other issues come into play too — namely, whether or not the existing Indian copyright law can deal with the emergence of generative AI, and if not, what role will Parliament play?

The dispute: What actually happened

The dispute over copyright between Asian News International (ANI) and OpenAI is built around the plaintiff’s claim that the defendant infringed on its copyright by using articles — many of which are subject to paywalls — in order to train ChatGPT. ANI has alleged that ChatGPT reproduced certain content from its proprietary journalism whenever it was queried, and has sought monetary damages as well as an injunction that would prevent OpenAI from using any additional content. ANI has filed suit with the Delhi High Court seeking Rs 2 crores in monetary damages and a permanent injunction against OpenAI’s continued use of its content.³

In its defence, OpenAI has argued that: 1) the Delhi High Court cannot assert jurisdiction over OpenAI, as both its servers and the training occurred outside of India;⁴ 2) LLM training does not involve “reproduction” in the traditional sense, but rather develops statistical probabilities (tokens) based on data it has access to, rather than using the actual copyright protected work itself;⁵ and 3) if reproduction were established, the use of ANI’s material would be excused under “fair dealing” as defined in Section 52 of the Act.

Interventions in this matter have been made by a number of stakeholders including the Digital News Publishers Association, the Indian Music Industry, the Federation of Indian Publishers, and technology-based startups — transforming this originally bilateral dispute into an overarching policy issue that will have profound implications for how we define intellectual property in an era dominated by AI.⁶

The outcome of this case will carry significance not just for the parties before the court, but for every creator, every publisher, and every AI company operating in India.

— Ranjan Narula & Parth Bajaj, IAM India IP Guide 2026

The legal foundation: Sections 14 and 51, and the exclusive rights at stake

It must first be established that there is a prima facie infringement before discussing fair dealing. Section 14 of the Act provides that a copyright owner has the exclusive right to control the reproduction of a literary work, which expressly includes the electronic storage of the work by any means.⁷ Section 51 states that copyright is infringed when a person has engaged in conduct prohibited by the Act and has reproduced or electronically stored a work without having obtained an appropriate licence from the copyright owner.⁸

The argument made by OpenAI — that its training does not use any “expressive tokens” — seems factually and legally questionable. In order to build such a model, it has to ingest a huge corpus of materials as a whole, copy them, tokenise them, and retain them for the long term. The LLM cannot function by reading and storing impressions of the material as a human would. The LLM relies upon the structural and semantic data from the entire body of data it has ingested in order to function.

Advocate Adarsh Ramanujan, the Amicus Curiae appointed by the Delhi High Court, appears to support this viewpoint by stating that OpenAI’s actions do engage the reproduction right under Section 14(a)(i), and that OpenAI’s storage of the materials was not “transitory or incidental”, as OpenAI suggests.⁹

The second Amicus, Prof. Arul George Scaria, takes a more permissive view, arguing that the storage of copyrighted material for the purposes of AI training may be allowed under the Act, and that the court should determine whether OpenAI used the material for purposes other than training.¹⁰ These opposing opinions illustrate a key point of contention: Indian copyright law was created before the internet existed and before generative AI was conceived, and its silence on machine learning has now left a chasm through which billions of rupees’ worth of creative output are disappearing.

The fair dealing fortress: What Section 52 actually permits

Under the Act, Section 52(1) sets out those acts that do not constitute infringement of copyright. It is the Indian equivalent of “fair use” under US copyright law — but whereas the fair use principle in the United States is an open-ended, factor-based analysis, the fair dealing provisions in India are a closed, exhaustive list.¹¹

The relevant head for purposes of OpenAI would be Section 52(1)(a), which allows for fair dealing with any work for the purposes of 1) private or personal use, including research; 2) criticism or review of any work; and 3) the reporting of current events.

None of these categories cover AI training for commercial purposes, and here is why:

Private use and research: The “private use” exception is aimed at individual users of works for non-commercial personal reasons — a student copying a piece of text for their own study, for instance. OpenAI’s training of ChatGPT represents large-scale, systematic, and unmistakably commercial use. ChatGPT is a commercial application which generates several billion US dollars of revenue worldwide. Claiming that training a commercial product constitutes “private use” stretches the provision beyond all understanding of its intent. In Super Cassettes Industries Ltd. v. Chintamani Rao, the Delhi High Court held that Section 52 is to be read exhaustively — its exceptions “carefully and exhaustively enlist” the acts permitted, leaving no room for judicial expansion.¹²

Criticism or review: In order to rely on this exception, the use must be for the purpose of commenting on or evaluating the work in question. OpenAI’s training of ChatGPT is clearly not a commentary or evaluation of ANI’s journalism — it is the wholesale copying of that work and the extraction of information from it to build a commercial product. The primary purpose is not to engage with ANI’s content expressively, but to extract patterns from it for commercial use.

Reporting current events: The exception at Section 52(1)(b) relates to the use of copyrighted material by newspapers, magazines, and broadcasters to report current events. OpenAI is none of these things, and its use of ANI’s content does not fit within the scope of news reporting.

56. …Section 52 carefully and exhaustively enlists various actions which would not constitute infringement of copyright in different classes of works and the limits on such use.¹³

The fundamental issue is structural. Unlike the United States, which uses an open-ended four-factor “fair use” test under 17 U.S.C. § 107 capable of accommodating unforeseen technologies, India’s fair dealing doctrine is a finite, closed list. When Parliament passed the Act and even when it amended it in 2012, AI training data ingestion was simply not a consideration. It is not there — and OpenAI cannot invoke a provision that does not exist for its particular use-case.¹⁴

The “transformative work” mirage

OpenAI may also claim that its outputs are fully “transformative” — that the responses of ChatGPT do not resemble the input articles at all, and therefore even if infringement is established at the ingestion stage, the final product does not carry the original work.

This argument has limited authority in India. There is no separately codified defence of transformative use in the Indian Copyright Act. While some Indian Courts have referenced the idea of “transformative use” — particularly in cases involving parody and educational usage — it has always been applied as a consideration within an existing Section 52 exception, not as a freestanding independent defence.¹⁵

If the AI training does not fall within any recognised Section 52 category, the fact that what it has produced may be “transformative” or “new” is legally irrelevant. The copying which is said to be infringing occurs at the point of ingestion — when ANI’s articles are copied and used to train the model. The final outputs of ChatGPT do not negate the act of infringement which has already taken place. A defence of “we no longer have your grain” does not apply if the act of consuming the grain without permission was itself the wrong.

The global picture: What other jurisdictions reveal

India is not alone in its current situation, but its statutory response is comparatively limited. The EU’s Digital Single Market Directive (2019) introduced an explicit text and data mining (TDM) exception — but crucially, it allows rights-holders to opt out and requires AI developers to disclose the datasets used for training purposes.¹⁶ ANI had already blocklisted its domain from AI training tools as of October 2024 — mirroring the European opt-out mechanism — yet OpenAI allegedly continued to use previously ingested content.¹⁷

The United States is presently navigating this question through dozens of Federal Court cases, with outcomes that remain deeply unsettled.¹⁸ Japan and Singapore have taken more permissive approaches, allowing broader data mining for AI development. India has no equivalent statutory provision — and this gap is now painfully exposed by the ANI litigation.

In November 2025, India’s Ministry of Electronics and Information Technology released AI Governance Guidelines, and the Department for Promotion of Industry and Internal Trade of India (DPIIT) has initiated consultations on whether the Act requires amendment to address AI training.¹⁹ These are policy signals, not law — but they acknowledge what the ANI Media case²⁰ has made unmistakable: The existing framework is inadequate.

The reform imperative: What India must do

The question before the Delhi High Court is narrow: Does Section 52, as it stands today, permit OpenAI’s conduct? This blog has argued that it does not. But even if the court rules against OpenAI, a deeper problem will remain — India lacks a coherent legal framework for AI and copyright, and resolving these disputes through litigation after the fact is not an adequate substitute for a clear and comprehensive legislative solution.

Three immediate reforms are urgently needed. First, Parliament should implement a clear TDM exception with opt-out provisions — permitting researchers and non-commercial users to develop AI models using publicly available materials while protecting copyright owners who have expressed opposition. Second, any commercial use of copyrighted data for AI training should require either a licence or the payment of a reasonable royalty to the copyright owner — consistent with the DPIIT’s ongoing consultations.²¹ Third, all AI developers operating in India — including those physically based abroad — should be required to publicly disclose the datasets they utilised, so that copyright owners can identify and assert their claims.

The Commerce Ministry has assembled a panel of eight experts in response to the ANI Media case²² to assess whether the Act adequately addresses AI technology.²³ This assessment is very much needed. India’s creative economy — journalism, music, literature, and cinema — cannot be sustained if AI systems train on its output without consent or compensation, and then compete commercially against it.

Conclusion

The farmer-and-factory metaphor does not merely illustrate a point — it illustrates a real structural injustice. AI companies claim they are creating new works from the materials they consume, and that their finished products do not contain the original content. But the question is not about what the AI company is producing; it is about whether they were entitled to take what they took.

Section 52 of the Act does not permit that taking. The exceptions set out in that provision do not — and cannot — accommodate the commercial mass-ingestion of copyrighted works for LLM training. ANI Media case is not just India’s first AI copyright case; it is the case that will determine whether Indian creators have rights that the law can actually protect in the age of AI.

The Delhi High Court’s eventual judgment will be scrutinised across the world. What it says about Section 52 will shape not only the outcome of this lawsuit, but the entire architecture of AI and copyright in India for the decade ahead. The court must say clearly: The harvest belongs to the farmer.