ANI v. OpenAI: India’s First Copyright Ruling On LLM Training

In an interim ruling, the Delhi High Court has examined whether the use of copyrighted news content for LLM training and output generation amounts to infringement or qualifies as fair dealing.

ANI Media Pvt. Ltd. (ANI) alleged that OpenAI OpCo LLC (OpenAI) had used its copyrighted news content without authorisation in two ways: by copying and storing it to train the large language models underlying ChatGPT, and by reproducing it in responses generated by ChatGPT. In a ruling dated July 24, 2026, the Delhi High Court rejected ANI’s application for an interim injunction and delivered India’s first substantive judicial analysis of copyright infringement and fair dealing in the context of generative AI. The dispute drew interventions from publishing, news, music, technology and policy bodies, reflecting its wider significance for copyright owners and AI developers.

Before turning to the legal questions, the Court explained how LLMs work. A large language model predicts the word most likely to come next in a sentence. It is trained on a vast collection of text taken from websites and licensed sources, referred to as ‘raw data’. This data is filtered and broken into smaller units called tokens, which may be words, parts of words or punctuation marks. Each token is assigned a number so that the computer can process it. During training, the model is shown sentences with words hidden and attempts to predict the missing or next word. It compares its answer with the original text and repeatedly adjusts itself, gradually becoming better at recognising language patterns. It is then further trained to answer questions, prepare summaries, translate text and write code. The Court also explained Retrieval-Augmented Generation, through which a model looks up external information when a question is asked.

Four issues were framed. The Court first addressed jurisdiction, then output-side infringement, and finally considered storage and fair dealing together, since the two questions were closely connected.

Issue: Whether Indian Courts have jurisdiction considering that OpenAI’s servers are located in the US

ANI's suit combined two distinct causes of action: (i) OpenAI's training of ChatGPT's underlying LLMs on scraped data, admittedly carried out on servers outside India (the ‘training claim’), and (ii) the generation of allegedly infringing responses, which occurred in India in reply to ANI's prompts (the ‘reproduction claim’). OpenAI argued the offshore location of training defeated jurisdiction over that claim regardless of jurisdiction over the reproduction claim, relying on Getty Images v. Stability AI (2025 EWHC 38 (Ch)), where the UK claimant did not press a training claim for the same reason. It further argued that ChatGPT's terms of use bind users to the jurisdiction of the courts in San Francisco.

The Delhi High Court held that territorial jurisdiction was established. ANI’s principal place of business and registered office were within its jurisdiction under Section 62(2) of the Copyright Act. OpenAI targeted Indian users, and the allegedly infringing outputs were generated within the jurisdiction, attracting Section 20 of the Code of Civil Procedure.

The Court also declined to sever the training and reproduction claims. It characterised storage on US servers as merely the terminal step in a chain beginning with access to copyrighted material from India. Since the reproduction claim was premised on training and the output was generated in India, the training claim could not be excluded merely because its final stage occurred abroad.

Accepting OpenAI’s argument, the Court observed, would allow infringers to evade Indian copyright law by locating their servers overseas. Relying on Neetu Singh & Anr. v. Telegram FZ LLC & Ors. (Delhi High Court, 2022), it held that copyright owners could not be left remediless and that territoriality must be applied flexibly in an era of cloud computing.

Issue: Does generating ChatGPT responses using ANI's copyrighted articles infringe copyright?

The Court accepted that ANI holds copyright in its articles and that unauthorised reproduction or communication to the public would infringe under Sections 14 and 51. But the infringement claim turned on whether ChatGPT memorises and reproduces ANI's protected expression — and on that, the Court found against ANI.

No case for memorisation. LLMs generally learn statistical patterns from training data rather than storing text verbatim, though the Court accepted that verbatim memorisation can occur with sparse or unique data - and that reproducing memorised content would amount to infringement. The 14 articles relied upon by ANI, however, were published after the training cut-off dates for the relevant GPT models. The Court considered it more likely that the responses resulted from RAG—live retrieval from ANI’s website—rather than memorisation during training. It also noted that ANI’s expert could elicit only insignificant regurgitation.

No substantial similarity, in any event. Copyright in news protects expression, not the underlying facts - as recognised in the Indian Supreme Court’s ruling in Eastern Book Company v. D.B. Modak (2008). Also, per the apex court’s decision in R.G. Anand v. Delux Films (1978), works must be compared as a whole rather than dissected into parts. On that comparison, ChatGPT's responses did not substantially reproduce ANI's expression.

The Court left memorisation open as an issue for trial if ANI subsequently produced evidence, but held that no prima facie case of output-side infringement had been established.

Issue: Does storing ANI’s copyrighted works for training ChatGPT amount to infringement?

The Court stated that ANI’s copyright includes the exclusive right under Section 14(a)(i) of the Copyright Act to reproduce its literary works “in any material form, including the storing of [them] in any medium by electronic means”. Electronic storage of ANI’s articles, whether temporary or permanent, would therefore amount to reproduction, irrespective of the purpose of storage.

However, Section 14 expressly makes these rights “subject to the provisions of this Act”, which includes fair dealing exceptions set out under Section 52. Thus, the Court held that the storage question would be considered together with OpenAI’s fair-dealing defence (see below).

Issue: Does OpenAI’s use of ANI’s copyrighted works qualify as fair dealing under Section 52 of the Copyright Act?

Section 52(1)(a) provides that fair dealing with a work, other than a computer programme, for specified purposes—including “private or personal use, including research”—does not constitute copyright infringement.

To determine whether OpenAI’s storage of ANI’s works fell within this exception, the Court applied a two-stage examination: first, whether the storage was for a purpose recognised under Section 52(1)(a) (the purpose test); and second, whether the use was fair (the fairness test).

The Court stated that Section 52(1)(a) must be given a broad and liberal interpretation and that OpenAI’s use satisfied the purpose test. It made three findings. First, ‘private use’ is not restricted to non-commercial purposes. Where the Copyright Act intends to distinguish between commercial and non-commercial use, it does so expressly; no such limitation could therefore be read into Section 52(1)(a).

Second, private use is not confined to individuals and may include internal use by companies. OpenAI’s training process was private because the material was used within a closed system and was not made available to the public. Its commercial purpose was irrelevant.

Third, applying the principle of updating construction, the Court held that ‘research’ is not confined to human activity and may include machine-led research.

On fairness, the Court observed that ‘fair dealing’ is undefined and that Indian courts have adopted different tests depending on the facts of each case.

Reliance was placed on the Delhi High Court’s ruling in The Chancellor, Masters & Scholars of the University of Oxford v. Narendera Publishing House (2008), concerning guidebooks that reproduced excerpts from academic textbooks. The Court there emphasised the transformative character of the use - the guidebooks served a purpose substantially different from that of the original work. A transformative use would ordinarily not act as a market substitute and could therefore qualify for the fair-dealing defence.

Also referred to was Bartz v. Anthropic (US District Court, N.D. California, 2025), which observed: “the purpose and character of using copyrighted works to train LLMs to generate new text was quintessentially transformative. Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to…replicate or supplant them—but to turn a hard corner and create something different.”

Similarly, the Court found that OpenAI’s training did not seek to substitute ANI’s works or economically compete with its business. ChatGPT performs numerous functions, whereas ANI’s business is news gathering and syndication. ANI had produced no evidence of lost market share, reduced subscriptions or other actual or potential damage.

The Court also considered the public interest. While copyright owners’ rights remain important, the societal benefits of scientific and technological research are relevant when assessing fairness. LLM training, in its view, advanced research, education, accessibility, dissemination of information and technological innovation.

Accordingly, OpenAI’s storage of ANI’s works for training fell within the fair dealing exception and did not prima facie amount to infringement.

No interim injunction

The Court held that the balance of convenience and irreparable injury weighed against an interim injunction. ANI had produced no evidence that OpenAI’s activities had caused loss of subscribers, advertising revenue or news-syndication business. Further, since ANI had itself offered OpenAI a licence for USD 7.5 million, any loss was quantifiable and could be compensated if ANI ultimately succeeded.

By contrast, an injunction could significantly disrupt OpenAI’s operations and affect the wider development of LLMs, including in India. The Court also noted that ANI could block its website from OpenAI’s crawlers and that OpenAI had, in any event, stopped accessing ANI’s website for training and RAG.

The Court therefore declined interim relief, while clarifying that its findings were only prima facie and would not determine the final outcome of the suit.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.