Delhi HC Refuses Interim Injunction to ANI Against OpenAI, Finds LLM Training Covered by Fair Dealing Under Section 52
Justice Amit Bansal held that OpenAI's storage of ANI's news articles to train ChatGPT qualifies as fair dealing under Section 52(1)(a) and that ChatGPT outputs are not substantial reproductions of ANI's works.
The High Court of Delhi on 24 July 2026 dismissed the application for interim injunction filed by news agency ANI Media Pvt. Ltd. against OpenAI OpCo LLC, declining to restrain OpenAI from using ANI's copyrighted news articles and interviews to train the large language models underlying ChatGPT. Justice Amit Bansal, sitting singly, held on a prima facie view that the storage of ANI's literary works for LLM training falls within the fair dealing exception under Section 52(1)(a) of the Copyright Act, 1957, and that the outputs generated by ChatGPT do not amount to substantial reproduction of ANI's original literary works. The judgment — spanning 135 pages and decided after submissions across more than twenty hearing dates between February 2025 and March 2026 — is among the first Indian judicial pronouncements directly addressing AI training and copyright in the context of news content.
ANI's Two-Pronged Copyright Claim
ANI filed the commercial suit accompanied by an interim injunction application in November 2024, alleging that OpenAI had infringed its copyright on two separate grounds. The first, described throughout the proceedings as the “training claim”, alleged that OpenAI scraped and stored ANI's news articles and interviews on servers located in the United States to train the LLMs underlying ChatGPT, without obtaining any licence. The second, described as the “reproduction claim” or “output claim”, alleged that ChatGPT's responses to user prompts substantially reproduced ANI's original literary works and communicated them to the public, thereby infringing ANI's exclusive rights under Section 14(a)(i) and Section 14(a)(iii) of the Copyright Act.
ANI placed on record nine illustrative instances from its plaint in which it claimed ChatGPT had reproduced its works. It also argued that OpenAI had entered into licensing arrangements with other news organisations, including the Financial Times, Associated Press, and Condé Nast, while consciously refusing to compensate ANI for the same use. ANI had itself written to OpenAI on 3 October 2024 offering a licence to its digital content for USD 7.5 million.
Six intervention applications were permitted. The Federation of Indian Publishers, the Digital News Publishers Association, and the Indian Music Industry broadly supported ANI. Flux Labs AI Private Ltd., IGAP Project LLP, and the Broadband India Forum broadly supported OpenAI. The court had also appointed two amici curiae: Mr. Adarsh Ramanujan, Advocate, and Professor Arul George Scaria.
Jurisdiction: Server Location Does Not Sever the Cause of Action
The court framed four issues on 19 November 2024. Because jurisdiction was foundational, it was taken up first.
OpenAI argued that the training claim involved acts occurring entirely outside India, on servers in the United States, and that the Indian Copyright Act could not apply to those acts. It relied on the terms of use of ChatGPT, which confer jurisdiction on courts in San Francisco, and on the Division Bench judgment in Blueberry Books v. Google India to contend that the infringing act must occur within this court's jurisdiction.
Justice Bansal rejected the submission. He noted that it was an admitted position that ANI's principal place of business and registered office are within the jurisdiction of this court, which independently invokes Section 62(2) of the Copyright Act. He also held that OpenAI specifically targets and offers its services to users and subscribers across India, including within Delhi, and that the alleged infringing responses were generated in India in response to prompts put by ANI.
On the territorial application of the Copyright Act to the training claim, the court held that “the storing of ANI's works on the US servers is a terminal step in the chain of events which begin from access of copyrighted works from India and transmission of the same abroad.” Accepting OpenAI's argument, the court observed, would allow infringers to evade Indian copyright law simply by shifting the terminal link of the chain to servers abroad. The court drew support from the coordinate bench judgment in Neetu Singh v. Telegram, where it was held that Telegram's choice to locate servers in Singapore could not deprive Indian copyright owners of their remedies. The training and output claims were also held to be inseparable: because the output is reproduced within the jurisdiction of this court, the training that underlies it cannot be examined in isolation.
Output Claim: Illustrative Articles Pre-Dated by Training Cut-Off
On the reproduction or output claim, a threshold factual finding disposed of ANI's nine illustrative instances. The training of the LLMs underlying ChatGPT ended in April 2022 for GPT-4 and April 2024 for GPT-4o. Every one of the news articles and interviews cited in the plaint was published in August or September 2024 — after both training cut-off dates. The court set out a table comparing each article's date of publication against the applicable training cut-off and held that the illustrative outputs could not have been produced through memorisation of ANI's works, because those works were not part of the training data.
The court reasoned that where a ChatGPT response refers to content published after the training cut-off, the most plausible technical explanation is that ChatGPT was using its Retrieval-Augmented Generation (RAG) feature — accessing external sources, including ANI's website, in real time rather than drawing on pre-trained knowledge. Whether outputs produced through RAG amount to copyright infringement had not been pleaded in the plaint, and the court declined to address it on the interim application.
On the question of substantial reproduction, the court examined the most detailed comparison put forward by ANI: an interview with Neeraj Chopra's mother, originally given in Hindi and translated into English by ANI. ANI's counsel highlighted this instance during oral submissions. The first prompt used by ANI was a standard question about what Neeraj Chopra's mother had told the media about the Pakistani gold medallist. The court found that ChatGPT's response to that prompt was not a substantial reproduction: it added its own commentary and differed sufficiently from ANI's expression. ANI then gave a second, adversarial prompt using the word “exactly” to try to extract the precise words. The court treated this as a deliberately engineered extraction attempt. Even so, the only similarity was a partial quote from the interviewee herself. The court went further and held, on a prima facie view, that the copyright in a public speech or interview vests with the person who delivered it under Section 17(cc) of the Copyright Act, and ANI had not produced anything to show an assignment of that copyright in its favour.
The court applied the test from the Supreme Court's judgment in R.G. Anand v. Deluxe Films — that works must be compared as a whole and not in dissected parts, and that infringement requires a transparent rephrasing or copy of a substantial and material part — and the Bombay High Court's judgment in Star India v. Leo Burnett, which prohibits compartmentalising a work into parts for comparison. Applying those tests to the nine instances collectively, the court concluded that ChatGPT's responses were not substantially similar to ANI's articles.
The court distinguished the Munich Regional Court judgment in GEMA v. Open AI on two grounds: in that case, the copyrighted song lyrics had been part of the training data, and the Munich Court had found verbatim reproduction in response to non-adversarial prompts. Neither condition was present here. It distinguished Associated Press v. Meltwater because that case involved verbatim scraping and delivery of excerpts by a news-monitoring service, not LLM outputs.
Training Claim and Fair Dealing Under Section 52(1)(a)
On the training claim, the court first addressed the scope of Section 14(a)(i) of the Copyright Act, which grants the copyright owner the exclusive right to reproduce a work “in any material form including the storing of it in any medium by electronic means.” OpenAI contended that storage during the training process is an intermediate step that does not persist once training is complete, and that a purposive interpretation of Section 14(a)(i) should be applied. Amicus curiae Mr. Ramanujan, by contrast, submitted that even one-time storage constitutes reproduction and therefore infringement. The court found, consistent with Mr. Ramanujan's view, that on a prima facie reading, collection and storage of ANI's works during LLM training does engage Section 14(a)(i).
The central question was therefore whether the training use qualified as fair dealing under Section 52(1)(a) of the Copyright Act. The court noted that Section 52(1)(a) uses the expression “private or personal use, including research” and that the question labelled Issue No. 3 — though framed using the US term “fair use” at the first hearing — had always to be read as the Indian concept of “fair dealing.”
The court conducted a two-part analysis: a purpose test and a fairness test, drawing on a body of Indian precedent as well as foreign decisions.
On the purpose test, the court examined whether OpenAI's use qualified as “private or personal use, including research.” OpenAI argued that the words “private use” and “research” must be read broadly, that commercial activity is not excluded from Section 52(1)(a) — unlike other provisions in Section 52 where commerciality is expressly excluded — and that the use of training data is non-expressive because what is extracted during training are statistical patterns, not the expressive content of the works. ANI countered that the exception is available only to the person conducting the research or private use, not to a commercial intermediary like OpenAI, and relied on Rupendra Kashyap v. Jiwan Publishing House. The court considered the legislative history of Section 52(1)(a), tracing the shift in language from “research or private study” in the original 1957 enactment, through the 1994 amendment expanding it to “private use including research”, to the 2012 addition of “personal use.” It also examined the scope of the Explanation added in 2012, which contains the phrase “not itself being an infringing copy”, and agreed with OpenAI that this limiting language applies only to computer programmes, not to other stored works.
The court held that the purpose test was satisfied on a prima facie view. The storage of ANI's works for LLM training constitutes use for research and does not lose that character merely because OpenAI is a commercial entity. The requirement of a “non-infringing copy” under Section 52(1)(a) does not apply to stored literary works of the type in issue.
On the fairness test, the court applied a multi-factor analysis adapted from the US four-factor fair use framework, which Indian courts have previously imported into the Section 52(1)(a) inquiry. It considered: the purpose and character of the use, including whether it is transformative; the nature of the copyrighted work; the amount and substantiality of the portion used; and the effect on the potential market for the original.
On transformativeness, the court found that training an LLM on text corpora is fundamentally different in purpose from reading or publishing news articles. ChatGPT is designed to assist users in analysing and generating text across a wide range of tasks; it is not designed to replicate or substitute for ANI's news syndication service. The court drew support from the United States District Court's findings in Bartz v. Anthropic and Kadrey v. Meta Platforms, while acknowledging that these decisions arose under the US Copyright Act, which differs from the Indian statute.
On market substitution, the court found that ANI had placed no material on record to show a loss of subscribers or reduction in syndication revenues attributable to OpenAI's operations. ANI's own offer to license its content for USD 7.5 million demonstrated that any harm was quantifiable in monetary terms. The court also noted that OpenAI had itself internally blocked ANI's website from its web crawlers and from the ChatGPT search/RAG function by the time of the hearing, and that ANI retained the ability to deploy opt-out mechanisms such as robots.txt or paywall crawlers, which it had not used.
On public interest — treated as a relevant factor particularly in intellectual property matters — the court referred to the wide societal benefits of LLM technology in education, healthcare, scientific research, accessibility for persons with disabilities, and information dissemination. It noted the broader policy interest in India developing its own LLM ecosystem, referencing Sarvam AI's launch of a domestic LLM in February 2026.
Balance of Convenience and Irreparable Injury
Having found that ANI had not established a prima facie case of infringement on either the training or the output claim, the court went on to address balance of convenience and irreparable injury.
OpenAI submitted that the interim relief ANI sought — deletion of all of ANI's works stored by OpenAI in whatever form — would amount to granting the final relief in the suit, and would also conflict with OpenAI's obligations under US law to preserve training data. ANI had offered to license its content for USD 7.5 million, which in the court's view confirmed that any loss ANI might suffer is quantifiable and capable of being compensated in money. An injunction against OpenAI, by contrast, could not be adequately compensated in monetary terms if OpenAI succeeded in the suit.
The court held that granting an interim injunction would be detrimental to the growth of AI in India and to the millions of users of ChatGPT in India, many of whom are not paid subscribers. Public interest weighed against the grant. On balance, both the balance of convenience and the irreparable injury test operated against ANI.
Order
Justice Amit Bansal dismissed I.A. 45300/2024, the application for interim injunction filed under Order XXXIX Rules 1 and 2 of the Code of Civil Procedure. He held on a prima facie view that OpenAI's act of storing ANI's original literary works for training LLMs falls under Section 52(1)(a) of the Copyright Act and therefore does not amount to infringement under Section 51. He also held that the outputs generated by ChatGPT were not substantially similar to ANI's original literary works and that ANI had not established memorisation and regurgitation of its works through ChatGPT's responses. ANI failed to make out a prima facie case, balance of convenience was against the grant, and irreparable injury would be caused to both OpenAI and the public at large if an injunction were granted at this stage. The court expressly noted that all findings are for the limited purpose of the interim application and will have no bearing on the final outcome of the suit, which remains pending.