ANI, the news agency, sued Open AI alleging two distinct forms of copyright infringement. The training claim is that Open AI scraped and stored ANI’s news articles to train the LLMs that power ChatGPT, and that the output or reproduction claim that ChatGPT generates responses to users that reproduce ANI’s copyrighted content.
Background
Looking to understand Copyright Law better? Read our complete Copyright Law Guide.
ANI sought an interim injunction under Order XXXIX Rules 1 and 2 CPC. Six intervenors joined the case. The Federation of Indian Publishers, the Digital News Publishers Association, and the Indian Music Industry supported ANI. Flux AI Labs, IGAP Project LLP, and the Broadband India Forum supported Open AI. Two amici curiae, Adarsh Ramanujan and Professor Arul George Scaria, assisted the court. Hearings ran from February 2025 through March 2026, spread across roughly twenty dates.
Issues Framed by the Court
The court framed four issues on 19 November 2024.
- Does storage of ANI’s data for training ChatGPT infringe copyright?
- Does use of ANI’s data to generate user responses infringe copyright?
- Does Open AI’s use qualify as fair dealing under Section 52 of the Copyright Act, 1957?
- Do Indian courts have jurisdiction given that Open AI’s servers sit in the United States?
The court took jurisdiction first because it is foundational. It then took the output claim, because its outcome bears on the fair dealing analysis. It took the training claim and the fair dealing defence together, since both sides agreed the two were intertwined.
Issue 4 – The Jurisdiction
Open AI argued that training happens entirely on foreign servers using data already stored abroad, that Section 62 of the Copyright Act is territorial and confers no extraterritorial reach, and that the purposeful availment test from Banyan Tree Holding v A. Murali Krishna Reddy cannot manufacture jurisdiction over conduct occurring outside India. It further argued that ANI had bundled two separate causes of action, the training claim and the output claim, and that jurisdiction had to be independently established for each.
ANI relied on Section 62(2), which lets a copyright owner sue where it resides or carries on business, and pointed to its own Delhi office. It also invoked Section 20 CPC, arguing that Open AI runs an interactive commercial website reaching Indian subscribers.
The court’s central move was to refuse the split Open AI wanted. It reasoned that the training claim cannot be entirely separated from the output claim, since the output claim rests on the same training and the output itself is reproduced within the court’s jurisdiction. Training also necessarily involves accessing and transmitting ANI’s data from India before it is stored abroad, so the process is not purely extraterritorial. The court drew support from Neetu Singh v Telegram, where a coordinate bench held that a defendant cannot escape Indian jurisdiction merely by locating servers outside the country when infringing material still circulates inside it.
Finding
On a prima facie view, the court has territorial jurisdiction and there is no extraterritorial application problem at this stage.
Issue 2 – The output or reproduction claim
ANI’s case rested on the idea that ChatGPT memorises training data and regurgitates it. It argued that detokenisation, the process by which the model converts numerical tokens back into readable text, is itself a reproduction of the underlying raw data, and that several ChatGPT responses closely tracked ANI’s original wording.
Open AI’s response was structural. An LLM is not designed to output extracts of its training data. It predicts the next token from a probability distribution learned across billions of words, and the same prompt given to different users returns different answers. Professor Scaria’s submissions supported this, citing research suggesting memorisation is rare and shows up mainly during deliberate extraction attacks, occurring in a small fraction of cases and mostly where a piece of content was duplicated many times across the training corpus. The court also drew on the US decision Bartz v Anthropic, where even the plaintiffs’ own expert could coax only around fifty words of copyrighted text out of a model using adversarial prompting techniques.
The finding that did the most work here was a dating problem. Training for the relevant models closed in April 2022 (GPT-4) and April 2024 (GPT-4o). ANI’s nine illustrative examples of alleged reproduction were all drawn from articles published after those cut-off dates, several in September 2024. A model cannot memorise text that did not exist when its training ended. The court also noted that ANI’s prompts, which explicitly asked ChatGPT to reproduce content “exactly,” were themselves adversarial rather than neutral, and even so failed to extract anything approaching substantial reproduction.
The court distinguished the foreign authorities ANI relied on. GEMA v Open AI involved verbatim reproduction of song lyrics under non-adversarial prompts, a different fact pattern entirely. Associated Press v Meltwater concerned a news aggregator scraping and redistributing articles directly, not an LLM. Cohere was decided at the motion to dismiss stage on far more numerous alleged instances of verbatim copying. Positive Black Talk dealt with song lyrics, where the threshold for substantial similarity is lower than for news writing, since a news article’s purpose is to report facts rather than to showcase original expression. Infopaq and the UK Meltwater case both involved literal extract-taking, again unlike the pattern here.
Finding
ANI failed to establish, at the prima facie stage, that ChatGPT’s outputs are a substantial reproduction of its works, or that memorisation and regurgitation occurred. The output claim fails for interim relief purposes.
Issues 1 and 3 – Storage for training, and the fair dealing defence
Scope of the reproduction right
The court first confirmed that Section 14(a)(i), as amended in 1994, gives a literary work’s owner the exclusive right to reproduce it “in any material form including the storing of it in any medium by electronic means.” Storage counts as reproduction whether temporary or permanent, and the purpose of storage does not matter for establishing the right. But Section 14 operates “subject to the provisions of this Act,” which brings Section 52 into play. An act that falls within a Section 52 exception is not an infringing copy at all, so the question becomes whether Open AI’s storage fits within Section 52(1)(a)(i), the exception for “private or personal use, including research.”
The court applies a two part test here, a purpose test and a fairness test.
The purpose test
Is commercial use disqualified. ANI argued that a for-profit company using data to build a commercial product cannot claim “private or personal use.” The court rejected this. It observed that where the legislature meant to restrict a Section 52 exception to non-commercial use, it said so explicitly, as in Sections 52(1)(ad), (k)(ii), (l), (n) and (o). Section 52(1)(a) contains no such limitation. The court also reasoned by analogy. A paid book review or a paid news report can still claim the fair dealing defence under adjoining sub-clauses of the same section, so commercial motive alone cannot disqualify a claim under 52(1)(a)(i) either. It cited the Canadian Supreme Court’s CCH Canadian, which held that “research” must be given a large and liberal reading and is not confined to non-commercial contexts, extending even to lawyers researching for paying clients.
Does the copy have to be non-infringing to start with?
ANI argued Open AI first had to obtain a lawful, non-infringing copy of the work before any exception could apply. The court held that this “non-infringing copy” requirement, found in the Explanation to Section 52(1)(a), is textually confined to computer programmes and does not extend to literary works stored electronically. In any event it was undisputed that Open AI took ANI’s content from ANI’s own freely accessible website, not from behind a paywall or through unauthorised access, so the point did not arise on the facts.
Does training qualify as research?
This is the most significant doctrinal move in the judgment. Mr Ramanujan, for the amici, argued “private” and “personal” should be read as cognate terms confined to individual human beings. The court disagreed, applying what it called an updating construction, the idea that an old statute should be read to give effect to its purpose as technology changes. It reasoned that research and learning are no longer confined to humans, and that the ultimate purpose of the exception, benefiting people, is served whether the research is done by a human or by a machine acting at human direction and for human benefit. It drew an analogy to the teacher exception in Section 52(1)(i), asking rhetorically whether that protection should evaporate the moment a human teacher is replaced by an AI tutor. It concluded that would be a regressive reading that would limit societal progress.
Finding on the purpose test
Training the LLMs underlying ChatGPT on ANI’s stored articles falls within “private or personal use, including research.” The purpose test is satisfied.
The fairness test
The court then asked whether the dealing, having passed the purpose test, was also fair in substance. It looked at three things.
- Whether Open AI’s use of ANI’s works was limited to training purposes.
- Whether the use caused economic harm or acted as a market substitute for ANI’s work.
- Whether the use served the public interest.
On market harm, the court found nothing on record beyond bare assertions from ANI showing any loss of market share or subscription revenue. It leaned on the transformative use reasoning from the US Second Circuit’s Authors Guild v Google, the Google Books case, where digitising copyrighted books for search functionality was found highly transformative and not a substitute for the books themselves. The more transformative a use, the less its commercial character weighs against it. The court held ChatGPT’s use of ANI’s articles is similarly transformative and does not substitute for ANI’s own product.
On public interest, the court took a broad view of the societal value of LLMs, citing benefits to education, research, translation, software development, and accessibility for people with disabilities.
Finding on the fairness test
All factors favour Open AI. Combined with the purpose test, storage of ANI’s articles for LLM training falls within Section 52(1)(a) and is not infringement.
Balance of convenience and irreparable injury
Even a party with a strong prima facie case can be refused an interim injunction if the balance of convenience and irreparable injury point the other way, so the court addressed this independently.
ANI argued that Open AI’s use diverts readers and advertising revenue from ANI’s own site, and that Open AI has already licensed comparable data from the Financial Times, Associated Press and Condé Nast while declining to pay ANI, amounting to unjust enrichment.
Open AI argued that deleting all stored training data, the relief ANI actually sought, would effectively grant final relief at the interim stage and would also conflict with its US law obligations to preserve data. It pointed to a letter in which ANI itself had offered Open AI a licence for 7.5 million US dollars, treating this as proof that ANI’s claim is a quantifiable, monetary one that can be compensated later if ANI succeeds at trial, whereas the harm to Open AI from an injunction could not be undone or compensated in money. Open AI also noted it had already, without prejudice to its legal position, stopped accessing ANI’s website for both training and retrieval purposes.
The court accepted Open AI’s framing. Because ANI’s own licence offer put a number on the claim, any eventual loss is compensable in damages, while an injunction against a global AI product used by millions in India, many of them non-paying users, would cause harm that cannot be reversed if Open AI later succeeds at trial. It cited a Niti Aayog paper on AI for inclusive development and reasoned that requiring LLM developers to licence data from every source in advance would make model development economically unworkable, with knock-on damage to India’s own emerging AI sector. Public interest, the court noted, is an established fourth factor in intellectual property injunction cases in India, citing Zydus Lifesciences, F-Hoffmann-La Roche and Astrazeneca.
Finding
Both balance of convenience and irreparable injury favour Open AI. No interim injunction should be granted.
Also Read: Music Licensing in Restaurants: Bombay HC Injunction
Final order (Interim Injunction)
I.A. 45300/2024, the injunction application, is dismissed. The court’s prima facie conclusions, all confined to the interim stage are:
- Storage of ANI’s literary works for training the LLMs underlying ChatGPT falls under Section 52(1)(a) and does not infringe Section 51.
- ChatGPT’s outputs generated using retrieval augmented generation are not substantially similar to ANI’s works and do not infringe.
- ANI did not establish memorisation or regurgitation of its works by ChatGPT.
- Balance of convenience and irreparable injury both favour Open AI and the wider public.
The judgment expressly states these observations are made only for deciding the injunction application and have no bearing on the final outcome of the suit, which continues to trial on evidence.
