The Delhi High Court's decision dated July 24, 2026, in ANI Media Pvt. Ltd. vs. Open AI OPCO LLC [CS(COMM)1028/2024] marks India's first significant judicial examination of copyright and large language model (LLM) training. Although the suit remains pending on the merits, the refusal of interim relief provides the first indication of how an Indian court may approach the use of copyrighted material for LLM training and AI-generated outputs. The Court was assisted by two amici curiae and six interveners.
ANI has alleged that OpenAI used its copyrighted news reports to train ChatGPT without authorization, thereby infringing its copyright. It also claimed that ChatGPT generated responses that reproduced ANI's content. OpenAI disputed these allegations, arguing that Indian courts lacked jurisdiction over the training claim because its servers were located outside India, that AI training constituted a lawful ‘fair dealing’ use under the Copyright Act, 1957, and that ChatGPT did not regurgitate or substantially reproduce ANI's original work.
Before turning to the legal issues, the judgment explains how a large language model (LLM) works. Training involves processing large volumes of data, including information obtained from websites and licensed sources. This data is tokenised (broken into units such as words, parts of words and punctuation marks) and processed so that the model learns statistical relationships and patterns. When responding to a user prompt, the model generates text based on those learned patterns; it does not ordinarily retrieve the underlying training documents. Retrieval-Augmented Generation (RAG), by contrast, enables the model to retrieve external information when generating a response.
Thereafter, the Court identified four key questions that would determine whether interim relief should be granted:
Territorial jurisdiction
OpenAI argued that the offshore location of its servers defeated jurisdiction over the training claim and that ChatGPT's terms of use conferred jurisdiction on courts in San Francisco.
It also relied on the UK case of Getty Images v. Stability AI (2025 EWHC 38 (Ch)) in which the claimant did not pursue a training claim in comparable circumstances.
The Court rejected OpenAI's objection that Indian courts lacked jurisdiction merely because its servers were located abroad. ANI’s principal place of business and registered office were within the territorial jurisdiction of the Delhi High Court and the allegedly infringing outputs were generated within India for Indian users. It characterised storage on US servers as a “terminal link” in an inseverable chain involving access to and transmission of copyrighted works from India and the generation of outputs for users in India.
Consequently, the Delhi High Court held that it possessed jurisdiction to entertain the suit. Relying on Neetu Singh & Anr. v. Telegram FZ LLC & Ors. (Delhi High Court, 2022) the Court observed that “in the present age of cloud computing and diminishing national boundaries in data storage, conventional concepts of territoriality cannot be strictly applied”, and took the view that acceptance of such an argument would “lead to evasion of Indian copyright law by infringers who shift the terminal link to servers abroad”.
Infringement in outputs generated by ChatGPT
The Court found that ANI had failed to establish a prima facie case of copyright infringement based on ChatGPT's responses. Copyright protects the expression of ideas rather than the underlying facts or information in news reports. The material placed before the Court did not demonstrate substantial reproduction of ANI's protected expression. Significantly, the 14 articles relied upon by ANI were published after the training cut-off dates of the relevant GPT models. The Court therefore considered RAG a more plausible explanation for the responses than memorisation during training. It observed that, on the evidence before it, the contention that OpenAI had permanently stored and memorised ANI's works for regurgitation could not be accepted. The question of memorisation was, however, left open for trial.
Storage of copyrighted works for AI training and the fair-dealing defence
The two connected issues were taken up together. It first held that electronic storage of copyrighted works, whether temporary or permanent, constitutes reproduction under Section 14(a)(i) of the Copyright Act, 1957. The issue was therefore whether OpenAI’s storage of ANI’s works for LLM training fell within the ‘fair dealing’ exception contained in Section 52(1)(a) of the Copyright Act, 1957.
Indian fair dealing differs from the broader US doctrine of fair use. US law applies a statutory four-factor inquiry, namely, (1) the purpose and character of the use, including whether such use is of a commercial nature or is for non-profit educational purposes; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and (4) the effect of the use upon the potential market for or value of the copyrighted work.
On the other hand, Section 52(1)(a) of the Copyright Act, 1957 specifies the purposes for which fair dealing is permitted. It exempts fair dealing with works for purposes including “private or personal use, including research”. The provision does not prescribe a fixed set of factors for assessing fairness. However, Indian courts have previously recognized that ‘transformative character’ of use can be crucial.
OpenAI argued for Section 52 to be ‘liberally construed’ as AI training constitutes research and private use and Section 52(1)(a) does not restrict private use to non-commercial purposes. The words ‘private’ should not be limited to personal use by an individual and ‘research’ should not be limited to a private context. Relying on the transformative character of training, OpenAI argued that ChatGPT is not a substitute for subscription to news agencies.
ANI, on the other hand, contended that a commercial AI developer cannot invoke the defence of private or personal use as they are making commercial use of the data collected. ANI argued that OpenAI’s claim of transformative use is flawed as “the expressive and non-expressive element of raw data are not segregated before feeding information in the LLM”.
The Court examined fair dealing through two stages: purpose and fairness. On purpose, it rejected ANI’s contention that commercial use necessarily fell outside Section 52(1)(a). It noted that the Act expressly uses “non-commercial” limitations elsewhere but not in this provision. It also distinguished “private” from “personal”: while personal use may relate to an individual, private use could extend to activity within a closed corporate environment. Since ANI’s material was used internally in the training process and was not itself made available to the public, the use could qualify as private.
Lastly, applying the principle of “updating construction”, the Court held that “research” need not be undertaken by a human and could extend to machine learning.
On fairness, the Court drew on existing jurisprudence, including transformative use, and identified three considerations relevant to this dispute: (1) the extent of use of ANI’s works; (2) economic competition and prejudice to ANI’s legitimate interests; and (3) the wider public interest. On the first consideration, memorisation, regurgitation or communication of exact copies was not prima facie established. The use was limited to training LLMs. As to the second, it was held that “it cannot be said that responses produced by ChatGPT are substitutes for news articles published by ANI” and that “nothing has been placed on record on behalf of ANI to show that they have lost any market share or there has been a reduction in subscription revenues due to operation of Open AI”. On the third, the Court considered the public benefits flowing from such training to be considerable. Commenting on the importance of AI, the Court remarked that “AI has brought about a transformational change in the lives of the people” and that it has “completely revolutionized the manner in which people seek and obtain information”. The Court was of the opinion that any interim injunction would be “detrimental to the growth of AI” and “have adverse impact on public interest, including millions of users of ChatGPT in India, many of whom would not be paid subscribers”.
The Court thus concluded, on a prima facie basis, that “Open AI's acts of storage of the literary works of ANI for the training of its LLMs would fall under Section 52(1)(a) of the Copyright Act and hence would not amount to infringement”.
The Court separately held that the balance of convenience and irreparable injury weighed against an interim injunction. ANI had not demonstrated any loss of subscribers, advertising revenue or news-syndication business attributable to OpenAI. Its earlier offer to license its content to OpenAI for USD 7.5 million also indicated that any loss could be quantified and compensated through damages if ANI ultimately succeeded. An injunction, by contrast, could significantly disrupt OpenAI's operations and adversely affect the public interest. The Court therefore declined interim relief.
Conclusion
The ruling marks an important first judicial view in India on the application of copyright law to LLM training. Its conclusions, however, are expressly prima facie and confined to the interim application; the Court clarified that they would have no bearing on the final outcome of the suit. The legality of using copyrighted works for AI training in India is therefore far from finally settled, but the decision provides an important early framework for how these questions may be approached. ANI has appealed the order before a Division Bench of the Delhi High Court, which issued notice to OpenAI on September 15, 2026. The appeal remains pending and is next listed for hearing on December 8, 2026.
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.







