ANI Media v. OpenAI and the Fair-Dealing Frontier of Generative AI Copyright Law in India
Author: Abdul Rahim J, B.A. LL.B., Government Law College, Salem
To the Point
On 24 July 2026, a single judge of the Delhi High Court, Justice Amit Bansal, refused to grant Asian News International (ANI) an interim injunction against OpenAI in what is being described as India’s first substantive judicial engagement with the copyright implications of training large language models (LLMs). The order held, at the prima facie stage, that OpenAI’s storage of ANI’s news content for training ChatGPT fell within the fair dealing exception under Section 52(1)(a) of the Copyright Act, 1957, treating AI training as a form of “research” and reading “private use” to include use by a closed corporate system rather than only by an individual. The ruling does not finally decide the underlying suit, but it supplies India’s first working framework for how machine-learning training may be tested against a copyright statute drafted decades before generative AI existed.
Use of Legal Jargon
A short glossary of the technical and legal terms central to this dispute is set out below, to keep the analysis that follows precise:
● Prima facie case — a case that appears, on the face of the pleaded material, to disclose a triable issue; the threshold a plaintiff must cross to obtain interim relief, short of final proof.
● Interim injunction — a temporary, pendente lite order restraining a party’s conduct until the suit is finally decided, granted only where prima facie case, balance of convenience, and irreparable injury are jointly satisfied.
● Fair dealing — the statutory defence under Section 52 of the Copyright Act, 1957, carving out specified uses (private use, research, criticism, review, reporting of current events) that do not amount to infringement even though they use a copyrighted work without licence.
● Idea–expression dichotomy — the doctrine that copyright subsists only in the original expression of an idea, not in the underlying facts, information, or ideas themselves.
● Retrieval-Augmented Generation (RAG) — a technique by which a language model retrieves relevant external text (e.g., a live news article) at the moment of answering a query, rather than reciting content memorised during training.
● Verbatim reproduction / memorisation — the risk, alleged but not established in this case, that a trained model reproduces a protected work’s exact expression rather than generating independently phrased output.
● Territorial jurisdiction — a court’s competence to hear a dispute; here, the Court held Indian courts may adjudicate claims against a foreign AI company where the effects of the impugned conduct are felt within India.
● Transformative use — a use that adds new purpose, meaning, or character to the original work rather than merely substituting for it in the market; borrowed conceptually from comparative fair use jurisprudence and invoked to describe AI outputs generated in a different form from the source text.
The Proof
The Court’s reasoning proceeded in two structured steps rather than a single, undifferentiated ‘fairness’ inquiry, and it is this structure that gives the order its evidentiary and doctrinal weight.
First, on the reproduction question: the Court accepted that electronic storage of a literary work, even temporary storage during training, falls within the reproduction right under Section 14(a)(i) of the Copyright Act. This finding mattered because it meant OpenAI could not argue that ingesting text for training was outside the statute altogether; the defence had to be located within Section 52, not outside Section 14.
Second, on the fair dealing defence: the Court applied what commentators have called a two-step purpose-and-fairness test under Section 52(1)(a)(i). The purpose test asked whether the storage was for private or personal use, including research; the Court held that AI training is a closed, internal process that qualifies as research once the provision is read in light of technological change, and that ‘private’ use is not confined to natural persons but can extend to a closed corporate system whose training corpus is never released to the public. The fairness test then asked, separately, whether the use as a whole was fair — examined, on the evidence before the Court, through the extent of the use, its commercial character, and its effect on the market for ANI’s own content.
On the evidentiary record, ANI’s central factual difficulty was that the specific articles it produced as examples had been published after OpenAI’s training data cut-off, so they could not have formed part of the training corpus; the outputs ANI relied on were, on the Court’s assessment, generated through Retrieval-Augmented Generation rather than reproduced from memorised training data. ANI also could not demonstrate any loss of subscribers, licensing revenue, or market substitution, and the Court noted that ANI retained the technical ability to block OpenAI’s web crawlers from accessing its site going forward. On commercial use, the Court held that the fair dealing clause in issue does not carry an express non-commercial limitation — unlike other sub-clauses of Section 52 where Parliament did impose one — so OpenAI’s commercial character did not, by itself, defeat the defence. Balancing these findings, the Court concluded that ANI had not shown a strong enough prima facie case, nor a balance of convenience in its favour, to justify restraining a functioning global product mid-litigation. The order is emphatically interim: the main suit, including trial on full evidence of whether training on copyrighted news content is lawful under Indian law, remains open.
Abstract
The Copyright Act, 1957 was not drafted with machine learning in mind, yet Indian courts are now being asked to fit generative AI training within a fair dealing provision built for photocopiers, libraries, and news reporters. This article examines the Delhi High Court’s interim order in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, the first Indian ruling to engage substantively with whether training a large language model on copyrighted news content infringes copyright, and whether the resulting outputs constitute actionable reproduction. It traces the Court’s two-stage reasoning on the reproduction right under Section 14 and the fair dealing exception under Section 52(1)(a), situates the order within the idea–expression dichotomy and originality jurisprudence developed in earlier Indian precedent, and places it alongside comparable transformative-use litigation abroad. The article argues that while the order offers a workable interim framework, its reliance on an open-ended ‘public interest’ consideration to widen the fair dealing exception is doctrinally unsettled and likely to invite appellate and, eventually, Supreme Court scrutiny as the underlying suit proceeds to trial.
Case Laws
ANI Media Pvt. Ltd. v. OpenAI OpCo LLC
I.A. 45300/2024 in CS(COMM) 1028/2024, Delhi High Court, order dated 24 July 2026 (Bansal, J.)
The primary authority discussed in this article. ANI, alleging unlicensed use of its news content to train ChatGPT and reputational harm from fabricated attributions, sought an interim injunction restraining OpenAI’s training and output use. The Court rejected OpenAI’s objection to territorial jurisdiction, held storage of the works to be reproduction under Section 14, but found the storage protected as fair dealing under Section 52(1)(a)(i) as ‘private use, including research,’ and declined interim relief for want of a strong prima facie case and demonstrated market harm. The main suit remains pending.
R.G. Anand v. M/s Delux Films
(1978) 4 SCC 118
The foundational Indian statement of the idea–expression dichotomy: copyright protects only the particular form in which an idea is expressed, not the idea, theme, or underlying facts themselves. The ANI order leans on this principle in holding that copyright protects ANI’s expression of the news, not the news events reported, raising the threshold for infringement where the same facts can only be expressed in limited ways.
Eastern Book Company v. D.B. Modak
(2008) 1 SCC 1
The Supreme Court’s ‘modicum of creativity’ standard for originality, rejecting both the ‘sweat of the brow’ test and a very high creativity bar. Relevant to any future trial-stage inquiry into which specific elements of ANI’s reporting are original enough to attract copyright at all, as against the un-copyrightable factual substratum of a news event.
Tips Industries Ltd. v. Wynk Music Ltd.
2019 SCC OnLine Bom 1362
Held that a party could not invoke the Section 52 fair dealing exception where the copyrighted works were being made available to the public via a subscription streaming service. The Delhi High Court distinguished this precedent in the ANI matter on the basis that OpenAI’s training corpus is never released to third parties in natural-language or tokenised form, unlike Wynk’s public-facing streaming use, and is therefore genuinely ‘private’ internal use.
Authors Guild v. Google, Inc.
804 F.3d 202 (2d Cir. 2015) (United States)
A comparative transformative-use authority, cited in the wider commentary on the ANI order. The Second Circuit held Google’s mass digitisation and indexing of books for search purposes to be a transformative fair use that did not substitute for the original works in the market — a reasoning structure echoed in the Delhi High Court’s own emphasis on transformation and absence of market substitution.
The New York Times Company v. Microsoft Corporation & OpenAI, Inc.
Case No. 1:23-cv-11195 (S.D.N.Y., filed 2023) (United States)
The leading comparative litigation abroad, still pending, alleging that training ChatGPT on Times articles and reproducing near-verbatim excerpts in outputs infringes copyright and falls outside fair use. It frames the same doctrinal question the Delhi High Court confronted — training-data reproduction versus market substitution and transformative purpose — and its eventual outcome will be a natural point of comparison for the trial of the ANI suit.
Conclusion
The ANI order is important less for what it finally decides — nothing, strictly, since it is only an interim ruling — than for the analytical architecture it supplies: separating the reproduction question from the fair dealing question, reading ‘research’ and ‘private use’ to accommodate closed, internal machine-learning processes, and refusing to treat commercial character as fatal to the defence where Parliament did not say so. That architecture will now be tested at trial on a fuller evidentiary record, including expert evidence on tokenisation and memorisation that the interim stage did not require. The order’s more contestable move is its invocation of an open-ended ‘public interest’ in AI development to widen fair dealing beyond its enumerated categories — a expansion that sits uneasily with the closed, exhaustively-listed structure Section 52 has traditionally been understood to have, and one that intervening bodies such as the Federation of Indian Publishers and the Digital News Publishers Association are likely to contest as the litigation proceeds. For a jurisdiction with no AI-specific text-and-data-mining exception, unlike the European Union’s Digital Single Market Directive, the Delhi High Court has, for now, done through interpretation what the legislature has not yet done through amendment. Whether that interpretation survives appellate review, and whether Parliament eventually intervenes with a purpose-built exception or licensing regime for AI training, will determine whether ANI v. OpenAI becomes India’s settled law of machine learning and copyright, or merely its opening argument.
