ANI Media Pvt. Ltd. v. Open AI Inc. & Anr. (now Open AI OpCo LLC)

Asian News International (ANI), run by ANI Media Pvt. Ltd., is one of India's largest news agencies. It licenses text, video and photographs to subscribing publications. In November 2024 it sued OpenAI in the Delhi High Court, making it the first Indian media organisation to sue a generative-AI developer over the use of copyrighted content.⁷¹⁰ ANI alleged that OpenAI had scraped and stored its news reports to train the large language models (LLMs) behind ChatGPT, and that ChatGPT reproduced its content verbatim or nearly verbatim in answers to users. The plaint also complained that ChatGPT attributed fabricated stories to ANI. ANI sought injunctions, deletion of stored copies of its works, and damages of ₹2 crore.⁷

Two facts from before the suit later weighed heavily. By a letter dated 3 October 2024, ANI had offered OpenAI a licence to its digital content for a fee of USD 7.5 million. OpenAI also said it had internally blocked ANI's website from its web crawlers, and in oral argument it added that it had blocked the site from the ChatGPT search (retrieval) function as well.¹

The suit was registered as CS(COMM) 1028/2024 in the court's Commercial Division. Alongside it were an application for an interim injunction (I.A. 45300/2024, under Order XXXIX Rules 1 and 2 of the Code of Civil Procedure) and an application for discovery (I.A. 45301/2024, under Order XI). It was first listed before Justice Amit Bansal on 19 November 2024. That day the court issued summons, appointed two amici curiae and framed four issues for deciding the interim application:¹²

1. Whether OpenAI's storage of ANI's data (news content claimed to be protected under the Copyright Act, 1957) to train ChatGPT amounts to infringement.
2. Whether using ANI's copyrighted data to generate responses for users amounts to infringement.
3. Whether the use qualifies as 'fair use' under Section 52 of the Copyright Act. In its judgment the court said this should be read as 'fair dealing', the term the statute uses.
4. Whether Indian courts have jurisdiction, given that OpenAI's servers are in the United States.

ANI's case had two parts, which the court called the training claim (copying and storing ANI's works to train the LLMs) and the output claim (reproduction of ANI's works in ChatGPT's responses).¹

At the hearing on 28 January 2025, Senior Advocate Amit Sibal for OpenAI objected that the court lacked territorial jurisdiction. The court said it would consider this while hearing the injunction application rather than as a preliminary question. It listed hearings from 21 February 2025, with the amici to be heard first.² OpenAI's position was that training took place entirely outside India on data stored on servers outside India, and that the Copyright Act does not apply beyond India's territory.¹

The two amici, advocate Adarsh Ramanujan and Prof. Arul George Scaria, filed written submissions that were taken on record on 28 January 2025. The court recorded that both were 'broadly in agreement' that the four issues framed on 19 November 2024 captured the dispute.²

Six organisations applied to intervene under Order I Rule 8A of the CPC:¹³

The music industry's application followed reports in February 2025 that labels including T-Series, Saregama and Sony Music wanted to join the case, alleging unauthorised use of sound recordings and compositions to train OpenAI's models.⁸ OpenAI opposed the publishers' application on the ground that the Federation had not filed a resolution authorising it to intervene, and the court directed that one be filed.² As early as January 2025 the court warned that, if interventions were allowed, 'the scope of the suit shall not be enlarged'.² On 4 July 2025 it allowed all six applications but limited the intervenors to the legal issues framed on 19 November 2024. They could not raise factual matters or new pleas.³

The court heard the interim application on 21 dates between 21 February 2025 and 27 March 2026, when it reserved judgment.¹

ANI argued that when an LLM converts tokens back into text it reconstructs its training data, so the whole process amounts to reproduction. It said some of ChatGPT's responses were exact or nearly exact copies caused by memorisation, and that public availability of news does not cancel copyright. It also pointed out that OpenAI had licensed content from the Financial Times, the Associated Press and Condé Nast.¹ DNPA argued that LLMs inherently memorise content, and that the Copyright Act draws no distinction between 'expressive' and 'non-expressive' uses. IMI argued that any output that substantially reproduces a work infringes it.¹

OpenAI replied that the training cut-off was April 2022 for GPT-4 and April 2024 for GPT-4o. All 14 works identified in the plaint were published in August or September 2024, so none could have been in the training data. It argued that Section 14 gives no exclusive right to 'use' a work, that its models do not store or reproduce training data, and that verbatim regurgitation is a rare failure triggered by adversarial prompting.¹ The Broadband India Forum argued that the freedom of speech under Article 19 includes a right to receive information. It said an injunction would force LLM developers to take licences from every news outlet, even though ANI had chosen not to use paywalls or crawler blockers.⁹¹

The amici focused on the wider consequences. Prof. Scaria warned that requiring OpenAI to take a licence would harm the growth of LLMs, especially those being developed in India. Mr. Ramanujan noted that ANI was demanding a licence fee, which suggested money could compensate it, and that ANI could itself block OpenAI's crawlers.¹

Justice Bansal delivered a 135-page judgment on 24 July 2026 and dismissed I.A. 45300/2024.¹¹⁰

Issue 4 (jurisdiction). Taking jurisdiction first, the court followed the Division Bench decision in Blueberry Books and the coordinate-bench decision in Neetu Singh v. Telegram. It held that storing ANI's works on US servers is 'a terminal step in the chain of events' that begins with access to copyrighted works from India. Accepting OpenAI's argument would let infringers escape Indian copyright law simply by moving that last step abroad. Because the output claim depends on training and outputs are produced within the court's jurisdiction, the two claims could not be separated. Prima facie, the court had jurisdiction and applying the Copyright Act to training was not extraterritorial.¹ Commentators noted that the court relied on both Section 62(2) of the Copyright Act and Section 20(c) of the CPC.¹²

Issue 2 (outputs). Because the illustrative articles were published after the models' training cut-offs, ANI could not show memorisation at this stage. The court called memorisation at best a disputed question for trial. Comparing ChatGPT's responses with ANI's articles, it found no substantial reproduction. News reports facts that can be expressed in limited ways, so the threshold for substantial similarity is higher than for creative works such as song lyrics. The court distinguished GEMA v. OpenAI, where lyrics were reproduced in response to non-adversarial prompts, noting that ANI had used 'fairly detailed, repeated prompts', including the word 'exactly', and still had not obtained a substantial reproduction. It also distinguished Associated Press v. Meltwater, Infopaq and NLA v. Meltwater as cases of verbatim extracts.¹¹²

Issues 1 and 3 (storage and fair dealing). Taking these together, the court accepted that OpenAI stores ANI's works at least temporarily, and held that deleting them after training does not change that. It then asked whether the storage falls within Section 52(1)(a)(i), which permits fair dealing for 'private or personal use, including research'. The court concluded that it does, and that the storage is therefore not infringement under Section 51.¹

ANI argued that Section 52, as an exception to infringement, should be read narrowly. The court rejected this. Relying on Narendera Publishing, the Division Bench decision in University of Cambridge v. B.D. Bhandari, and the Supreme Court of Canada's decision in CCH Canadian, it held that Section 52 is not a mere proviso to Section 51. Instead, it 'independently defines the rights/privileges of a user' and must be read broadly.¹¹¹ It then applied a two-step test.

Purpose test. (a) Commercial use. DNPA and IMI argued that the 'non-commercial' limit in Section 52(1)(ad) should be read into clause (a). The court noted that Parliament had expressly limited clauses (ad), (k)(ii), (l), (n) and (o) to non-commercial use but had not limited clause (a). Commercial purpose therefore does not by itself defeat the defence. (b) Non-infringing copy. The court held that the Explanation's requirement that the copy not be infringing applies only to computer programmes. In any event, OpenAI took ANI's works from its freely accessible website, not from pirated sources or by breaking a paywall, which distinguished Bartz v. Anthropic. (c) Private use and research. 'Private' is not limited to individuals. Because the data is held in a closed environment accessible only to the models, 'the use amounts to being purely private'. Applying an updating construction, the court said research 'is no longer confined to humans' and covers machine learning.¹¹¹

Fairness test. Prima facie, the use was limited to training, and ANI had not shown lost subscribers, lost syndication revenue or market substitution. The court also found that LLMs serve the public interest in information, education, research and accessibility.¹

Balance of convenience. ANI could block crawlers, and OpenAI said it had already blocked ANI's site for both training and RAG. ANI's USD 7.5 million licence offer showed its loss could be measured in money, while an injunction would harm OpenAI irreparably. An injunction would also be 'detrimental to the growth of AI', particularly LLMs being developed in India, and to millions of ChatGPT users. The court stressed that its observations were limited to the interim application.¹ A corrigendum of 27 July 2026 replaced a citation to Bartz v. Anthropic in paragraph 80 with Kadrey v. Meta Platforms.⁴

The judgment is India's first detailed judicial treatment of whether training generative AI on copyrighted works infringes copyright. At the interim stage it favours AI developers. It reads Section 52(1)(a) to cover commercial machine-learning 'research', extends Indian courts' user-rights reading of fair dealing to AI training, and treats a rights-holder's failure to opt out of crawling as relevant to interim relief. Its jurisdiction holding cuts the other way: foreign AI providers that serve Indian users cannot avoid the Copyright Act by training abroad.¹¹¹¹²

Unlike the Munich court in GEMA v. OpenAI, the ruling put decisive weight on training cut-off dates and on how prompts were phrased. SpicyIP commentary welcomed the user-rights framing but criticised the merger-doctrine discussion as unnecessary. It questioned whether word choice alone, such as using 'exactly', makes a prompt adversarial, and warned that the broad view of territorial jurisdiction could encourage forum shopping.¹²

For the news publishers and music industry who supported ANI, the ruling makes interim relief against AI training hard to obtain without evidence of memorisation, verbatim outputs or market harm. That shifts attention to the trial, the appeal and possible reform of the Copyright Act.

As of 14 October 2026, the case is pending at two levels.

Appeal. ANI challenged the refusal of interim relief in FAO(OS)(COMM) 237/2026. On 15 September 2026 a Division Bench of Justice Avneesh Jhingan and Justice Manmeet Pritam Singh Arora issued notice. OpenAI's counsel accepted notice and was given four weeks to reply, and the appeal is listed on 8 December 2026. Senior Advocate Kapil Sibal also appeared for an intervenor.⁵

Suit. After the judgment, the suit was sent to the roster bench. On 14 September 2026 Justice Jyoti Singh listed it, together with the pending discovery application (I.A. 45301/2024) and I.A. 26192/2025, on 27 January 2027 at the plaintiff's request.⁶ Whether storage for training and ChatGPT's outputs infringe ANI's copyright will be decided finally at trial, on evidence. That includes whether OpenAI's models memorise and regurgitate ANI's works.