ANI v OpenAI is the first Indian ruling on whether training an AI model on copyrighted work infringes, and the Delhi High Court refused ANI an injunction. It held, prima facie, that storing news articles to train ChatGPT is fair dealing under Section 52(1)(a) of the Copyright Act, 1957.
In a 2026 judgment, Justice Amit Bansal of the Delhi High Court dismissed ANI Media’s application for an interim injunction against OpenAI, in a suit the news agency had filed in November 2024. The Court took up four issues: whether it had territorial jurisdiction over a company whose servers sit in the United States, whether ChatGPT’s outputs reproduced ANI’s articles, whether storing those articles to train the models infringed, and whether fair dealing covered that storage. ANI won the first and lost the other three. Two amici curiae assisted the Court, and intervenors appeared on both sides, including the Federation of Indian Publishers for ANI and the Broadband India Forum for OpenAI.
This article sets out what ANI v OpenAI leaves a publisher holding, and what it gives an AI developer.
Where the ANI v OpenAI ruling leaves a publisher whose content was scraped without a licence
A publisher in ANI’s position is left with a live suit and no interim protection. The Court dismissed the injunction application outright, holding that ANI had made out no prima facie case, that the balance of convenience ran against it, and that an injunction would cause irreparable injury to OpenAI and to the public at large. Every one of those findings is interim, and the Court said in terms that none of them bears on the final outcome of the suit.
One issue went ANI’s way. OpenAI argued that its servers sit outside India, so training on Indian material could not engage the Indian Copyright Act without giving the statute extra-territorial effect, and the Court declined to accept that at the prima facie stage.
The reasoning runs through where the copying begins rather than where the copy lands. Storing ANI’s works on servers abroad is, in the Court’s words, a terminal step in a chain of events that begins with accessing copyrighted works from India. The Court followed Neetu Singh v. Telegram, which held that a server located abroad cannot leave a rights holder completely remediless. It also refused to treat the training claim and the output claim as separate for jurisdictional purposes, because whatever ChatGPT produces is reproduced in front of users inside its jurisdiction.
An Indian news agency can sue a United States AI developer in Delhi over material taken from an Indian website, and the forum objection will not dispose of the case at the threshold. The Court also accepted, prima facie, that ANI owns copyright in its articles even though they sit free and open on its own website, so publishing without a paywall does not push work outside the Act.
Why the output claim against ChatGPT failed on substantial reproduction
The output claim failed on dates before it failed on anything else. ANI put illustrative articles on record and showed ChatGPT reproducing material from them, but GPT-4’s training data stops at April 2022 and GPT-4o’s at April 2024, and every one of those articles was published after the relevant cut-off. A model cannot recite from memory something it was never trained on.
The Court drew the obvious inference. If the articles postdate training and their text still surfaces, the system is fetching them rather than recalling them. The judgment describes the instances as being in the nature of live links, perhaps reflecting the retrieval-augmented generation technique, where the model searches the live web and summarises what it finds.
That inference opens a hole the judgment does not fill. Whether retrieval-augmented outputs infringe was not pleaded in the plaint, so the Court expressly declined to decide it. A publisher whose real complaint is about live retrieval rather than about training should read this ruling as a pleading lesson and not as an answer.
On the material that was pleaded, the Court applied the test from R.G. Anand v. Delux Films. Infringement turns on the form, manner and arrangement of the expression rather than on the idea or the information, and where two works report the same events, the question is whether an ordinary reader would see the second as a copy of the first. Comparing the responses against ANI’s articles, the Court found short factual overlaps and shared quotations alongside different sentence structure, ordering and framing.
News reporting carries thin copyright, and the Court said so. Facts and events belong to nobody, so an agency’s protection extends to the way it writes them up and no further, a proposition the judgment traces through Feist Publications v. Rural Telephone Service. A direct quotation from a public figure is the speaker’s words, not the words of the agency that transcribed them.
ANI’s fallback was memorisation, the contention that the models had internalised its articles and could regurgitate them on the right prompt. The Court held that this could not be accepted at the prima facie stage, because what OpenAI’s models do or do not store is a disputed question that can only be determined at trial on evidence. That is a deferral rather than a rejection, and the distinction matters to anyone reading the ruling as a licence. A claim over a person’s name, voice or likeness under personality rights never meets this problem, because there the protected subject matter is the identity itself rather than the expression inside a news report.
The conclusion on this issue is that ChatGPT’s responses are not a substantial reproduction of ANI’s works. Anyone bringing a similar claim now needs sampled outputs that predate the model’s training cut-off, and a comparison run on expression rather than on facts.
What ANI never placed on record about its own opt-out and licence offer
What ANI never placed on record decided the injunction. It did not show that it had used the opt-out available to it, it did not show that it had lost a single subscriber, and it had itself put a price on the claim months before the judgment. Each of those gaps went to the balance of convenience, and together they cost ANI the application.
The opt-out came first. Any website operator can block a crawler, and the same control blocks the live retrieval that the Court thought was actually producing the outputs ANI complained about. Despite having that option, the judgment records, ANI evidently did not exercise it, while OpenAI had by then blocked ANI’s site from its own crawlers and from the search and retrieval function in ChatGPT.
The second gap ran wider. ANI’s paying subscribers republish its copy on their own sites, and they too could have blocked OpenAI and chose not to, possibly, the Court observed, for their own commercial reasons. Content the agency licensed onward therefore stayed reachable through channels ANI controlled by contract if not by robots.txt.
The third gap was the most expensive. There was no averment and no material showing that ANI had lost subscribers or suffered any loss in its syndication business, and on market share and subscription revenue the Court found nothing beyond bare averments. Against that emptiness sat a communication of October 2024 in which ANI had offered OpenAI a licence for USD 7.5 million. The Court read the offer as ANI’s own valuation: a claim whose holder has already priced it is compensable in damages, and an injunction shutting a model out of Indian news content is not.
For a publisher the operative lesson is not about copyright at all. The technical control you did not use, the loss you did not evidence and the number you already named become the answers to your own injunction application. Blocking a crawler creates no right, but declining to block one while asking a court for urgent relief against the crawling is a contradiction the other side will read back to you.
What an AI developer training on Indian content can rely on after ANI v OpenAI
A developer can rely on Section 52(1)(a) of the Copyright Act, and on very little else. The Court held that storing ANI’s works to train the models satisfies both limbs of that defence and so does not infringe under Section 51. It held so prima facie, at an interim hearing, in terms that expressly do not bind the trial.
Start with what the ruling does not say. It does not say that ingesting a work for training falls outside the owner’s rights. Section 14(a)(i) covers reproducing a literary work in any material form, including storing it in any medium by electronic means, and the Court read that as reaching the copies OpenAI made. The developer’s position rests on a defence, and a defence can be lost on facts the developer controls.
ANI’s main answer to that defence was the Explanation to Section 52(1)(a), which says the sub-clause does not apply to a copy of a computer programme unless it is a non-infringing copy. Read ANI’s way, anything stored electronically would have to be a lawfully obtained copy before fair dealing could apply at all. The Court rejected the reading, because the Explanation speaks of a computer programme and a literary work stored electronically is not one.
The illustration the judgment uses is worth carrying into any advice on this. A researcher who photocopies a page in a library is within Section 52(1)(a), and on ANI’s construction the same researcher scanning the same page would fall outside it, because the scan is an electronic copy. That distinction, the Court held, cannot be the intent of the provision.
Provenance still matters, and this is where the ruling is narrower than its headline. The Court took the point from Bartz v. Anthropic, where a United States District Court separated lawfully acquired copies used for training from books downloaded from pirate libraries and treated the second as a different question. ANI never alleged that OpenAI had broken a paywall or taken its material from an unauthorised source, and the articles were freely available on ANI’s own site.
The defence as reasoned therefore attaches to lawful collection from openly published material. It says nothing for a developer that scraped subscription content, defeated an access control, or built a corpus from a shadow library, and the approving citation of Bartz signals where such a case would go. It also says nothing about what the statute will look like in two years, with India’s first dedicated AI legislation still moving through the policy pipeline.
How the Court read Section 52(1)(a)(i) to cover training an AI model
The Court read Section 52(1)(a)(i) to cover model training by separating two words that are usually treated as one. The sub-clause permits fair dealing with a work for private or personal use, including research, and ANI, supported by one of the amici, argued that a private company training a commercial model cannot be making private use. The Court held that private and personal are not the same word doing the same job: personal may be confined to an individual, but private ordinarily covers a group, a body or a company. Reading the two interchangeably would leave private with no work to do, which cannot be right when Parliament put both into the clause in 2012.
Support came from B. Malini Mallya, where the Supreme Court treated a dance performance at an educational institution as private use within Section 52(1)(a)(i). A use inside an institution rather than by a single individual was still private use.
Research then took an updating construction. The word has to be read against modern technical advancement, so research is no longer something only a human does, though the judgment keeps the anchor in place: the research is ultimately at the behest of humans and for the benefit of humans. The Court’s own illustration is a human teacher replaced by an AI bot, where confining the provision to human researchers would be a regressive view. On that reading, training a large language model on stored works satisfies the purpose test.
The three fairness factors the Court built for AI training in place of the US four-factor test
The three factors are whether the use is limited to training the models, whether it creates economic competition that prejudices the owner’s legitimate interests, and whether the resulting functions serve the public interest. The Court wrote them for this case rather than taking them from anywhere, and the reason it had to matters more than the list.
India has no four-factor fair use test, and OpenAI could not import one. A Division Bench of the same court held in the Rameshwari Photocopy Services case that the American test does not apply here, because Section 52 enumerates permitted purposes rather than opening a general defence. Indian courts have instead used different tests case by case, with no single uniform standard, and counsel on both sides accepted as much.
That leaves fairness as a question of fact, degree and the overall impression the court forms. The three factors are this judgment’s answer to that question on these facts, not a test a later bench is bound to apply.
The second factor decided the issue, and it decided it on evidence. The Court needed to know whether ChatGPT competes with ANI in ANI’s own market, and found that except for bare averments, nothing had been placed on record to show lost market share or reduced subscription revenue. Without proof that readers were substituting the chatbot for the wire service, there was no actual or potential damage to weigh. The factor is not about whether the use is commercial, and OpenAI’s commercial motive was never in doubt.
The comparative material pointed the same way. Bartz v. Anthropic and Kadrey v. Meta Platforms both treated training as transformative, because a model learns statistical relationships rather than serving the work back, and Authors Guild v. Google reached a similar conclusion about a search index built from scanned books. Commercial purpose in each was held not to defeat the defence by itself.
Public interest carried the third factor and then carried the injunction. The Court listed what the technology does: analysing and generating text, widening access to information, supporting education and scientific research, assisting software development and translation, and providing tools for persons with disabilities. At the injunction stage it went further, holding that requiring a licence from every source would make large language model development economically unviable, and that an order against OpenAI would reach Indian developers and millions of Indian users. Public interest has been treated as a fourth factor in intellectual property injunctions in this court before, in Zydus Lifesciences, F-Hoffman-La Roche and AstraZeneca.
The questions that decide the case are all still open. Whether retrieval-augmented outputs infringe was never pleaded and so was never decided. Whether tokenisation and vectorisation during training are further acts of reproduction, whether ANI’s articles clear the skill and judgment threshold for copyright at all, and what follows from ChatGPT attributing fabricated statements to ANI are each left for trial.
India has landed in a different place from Germany on similar facts. The Munich Regional Court granted an injunction against OpenAI in November 2025 in the GEMA case, on evidence that the models had memorised song lyrics, which is the very finding ANI could not make out here. The gap between the two outcomes is evidentiary rather than doctrinal, and it is the same gap that opens wherever AI output is checked against the source it claims.
Frequently asked questions
Did the Delhi High Court decide that training AI on copyrighted work is legal in India?
No. Everything the Court held was prima facie, decided on an interim application, and the judgment records that its observations have no bearing on the final outcome of the suit. The suit itself continues to trial on all four issues.
Can an Indian publisher still sue OpenAI in India after this ruling?
Yes, and jurisdiction was the one issue ANI won. Servers located in the United States did not take the case outside the Delhi High Court, because accessing the copyrighted works began in India and the outputs are reproduced before users here.
Does ANI v OpenAI protect every AI company that trains on Indian content?
No. The reasoning turns on material that was freely available on the rights holder’s own website, and the Court expressly noted that no paywall was broken and no unauthorised source was alleged. A corpus built from subscription content or a pirate library raises a different question.
How is fair dealing under Section 52 different from US fair use in AI training cases?
Section 52 lists specific permitted purposes, so a use must first fit one of them, whereas the American doctrine opens a general defence assessed on four factors. Indian courts have held that the four-factor test does not apply here, which is why this judgment built its own.
Does blocking a web crawler protect a publisher under Indian copyright law?
Blocking creates no right on its own, but not blocking cost ANI at the injunction stage. The Court noted that ANI had an opt-out available against both crawling and live retrieval and evidently did not use it, and treated that against it on the balance of convenience.
What happens next in ANI v OpenAI?
The suit proceeds to trial, where ownership, memorisation, tokenisation and the fair dealing defence will be decided on evidence rather than on affidavits. The interim findings do not bind that decision, and an appeal against the refusal of the injunction remains open to ANI.
References
- ANI Media Pvt. Ltd. v. Open AI OpCo LLC, CS(COMM) 1028/2024, High Court of Delhi, judgment of 24 July 2026, neutral citation 2026:DHC:5900 (official judgment PDF)
- Full text of the same judgment (Indian Kanoon)
- The Copyright Act, 1957 (14 of 1957), sections 14(a)(i), 51 and 52(1)(a) with its Explanation, including the Explanation added by the Copyright (Amendment) Act, 2012
- Reporting on the outcome and the reasoning on storage for training (LiveLaw)
- GEMA v. Open AI, Munich Regional Court I, 42 O 14139/24, judgment of 11 November 2025, on memorisation of song lyrics (Library of Congress Global Legal Monitor)
- For the practitioner-side question of using these tools without importing their errors, see how to use ChatGPT and Claude for legal research and drafting in India
This article is for informational and educational purposes only and does not constitute legal advice. For advice on a specific case, consult a qualified advocate.
