1. Introduction
The advent of large language models (“LLMs”) in late 2022, followed by their rapid growth, has introduced significant complexities into the copyright landscape. Creators have resisted the unauthorized use of copyrighted works to train LLMs which can produce outputs that substantially mimic, and in some cases, reproduce their works. In ANI Media (P) Ltd. v. Open AI OPCO LLC,[1] the Delhi High Court (“DHC”) issued an important interim ruling where it examined for the first time the extent to which the use of copyrighted works to train LLMs is protected under the Copyright Act, 1957 (“Act”).
This newsletter examines the above ruling and its implications in India.
2. The Dispute
OpenAI is an American AI research organization that developed the GPT series of LLMs, including ChatGPT. In doing so, OpenAI procured and processed publicly available literary works, including those in which ANI, an Indian news agency, claimed copyright. In 2024, ANI filed a suit before the DHC alleging that OpenAI had infringed its copyright through the unauthorized use of its works. There were two principal infringement claims. First, OpenAI copied and stored ANI’s copyrighted works for training its LLMs. Second, ChatGPT reproduced substantial portions of ANI’s copyrighted works in response to user prompts. Pending adjudication of the suit, ANI sought an interim relief viz. a direction requiring OpenAI to delete all of ANI’s works held and stored in “whatever form.” OpenAI contested on the ground that its use of ANI’s works for training its LLMs constituted a “fair dealing” within the meaning of section 52(1)(a)[2] of the Act.
3. The Decision & Analysis
There were four principal issues before the court, as discussed below.
3.1 Territorial jurisdiction: OpenAI contended the DHC could not entertain the suit since its LLMs were trained on servers located outside India and ChatGPT’s Terms of Use contained a clause vesting exclusive jurisdiction in the courts of San Francisco. The court rejected this contention and held that jurisdiction lay squarely with the DHC under both section 62(2)[3] of the Act and section 20[4] of Code of Civil Procedure since (i) ANI was based in Delhi; (ii) OpenAI actively marketed and monetized ChatGPT to users within India; and (iii) effects of the alleged infringement, including consumption and reproduction of outputs, were felt in India.
3.2 Use of copyrighted works constitute infringement: Next, the court considered whether OpenAI’s use of ANI’s content to generate responses for ChatGPT users amounted to copyright infringement. ANI contended that ChatGPT, when prompted, produced outputs that substantially reproduced its copyrighted articles. In response, OpenAI claimed that ChatGPT was trained not to reproduce but to learn and utilize the grammatical rules and syntax from ANI’s work, to which ANI could not be said to have copyright. The court decided the issue by applying the test for infringement laid down by the Supreme Court in R.G. Anand v. Deluxe Films,[5] which requires the two works to be compared as a whole, rather than examining selected portions of the works. Applying this test, the court found in favour of OpenAI and categorically noted that when viewed as a whole, ChatGPT’s output bore no substantial similarity to ANI’s work.
3.3 Storage of copyrighted works and application of the “fair dealing” exception: ANI contended that OpenAI had, without a license or authorization, stored its copyrighted articles on its servers for training ChatGPT. It argued that, under section 14(a)(i)[6] of the Act, storage of copyrighted works by electronic means amounted to reproduction of work. Thus, the act of unauthorized storage itself constituted an infringement, irrespective of what ChatGPT subsequently generated for end users. OpenAI did not dispute that such storage had occurred but submitted that it was only temporary and LLMs did not have access to the content upon completion of their training. It further argued that this act of storage was protected as “fair dealing” under section 52(1)(a). The issue of the legality of storage of ANI’s work for training and the applicability of the “fair dealing” exception was thus tested together as intertwined questions.
The court agreed that by virtue of section 14(a)(i), the storage of copyrighted work by electronic means, whether temporary or otherwise, amounted to reproduction. Therefore, its unauthorized storage would constitute infringement. However, it also noted that section 14(a)(i) was subject to the exception of “fair dealing” under section 52(1)(a) which permitted the use of copyrighted works without the owner’s consent for (i) private or personal use, including research; (ii) criticism or review; and (iii) reporting current events. OpenAI claimed protection under the exception “private or personal use, including research.”
ANI raised three threshold arguments to keep OpenAI outside this exception. First, a commercial entity like OpenAI could not claim “private” use. The court disagreed, noting that section 52(1)(a), unlike some other clauses of section 52 contain no explicit requirement that the use be non-commercial. Second, ANI contended the fair dealing exception could not apply unless the copies used to train were themselves lawfully obtained and non-infringing. The court disagreed and held that section 52(1)(a) does not contain such a requirement, and the court could not import a condition that legislature had not imposed. Further, ANI had placed nothing on record to show that OpenAI obtained its works from unauthorized sources or by circumventing ANI’s paywall. Rather, it was undisputed that OpenAI sourced the material from content freely available on ANI’s website. Third, ANI argued that “private use or personal use, including research” could only mean use by an individual human being, not a corporation processing data at industrial scale. The court rejected this too and held that while the term “personal” in section 52(1)(a) may be confined to individuals, “private” is broader and can extend to a closed group or a company. It held that “research” was not an exclusively human activity, and machine learning can itself constitute research within the meaning of the statute. On this basis, the temporary storage of ANI’s articles for LLM training was held to qualify as private use.
Finally, to determine the “fairness” of the use by OpenAI, the court examined whether (i) the use was confined only to training; (ii) it caused actual or potential economic harm to ANI’s legitimate commercial interests; and (iii) it served the broader public interest in research, innovation and the dissemination of knowledge. On the first factor, the court held that the usage was indeed limited to train ChatGPT. On the second, it held OpenAI operated in a distinct market with limited overlap not capable of substituting ANI’s services and that ANI had failed to produce evidence of any actual economic loss. On the third factor, it held that LLMs advance scientific knowledge, innovation, education and contribute to AI research, thereby satisfying the public interest test. The court ultimately held that OpenAI’s use of ANI’s works amounted to “fair dealing.”
3.4 Analysis: Though the DHC’s decision may appear to signal a liberal stance toward AI training, this reading requires some qualification. Several of the court’s findings in OpenAI’s favor were only driven by insufficient evidence before the court at the interim stage. For instance, the finding that ChatGPT’s outputs did not substantially reproduce ANI’s articles rested on ANI’s failure to demonstrate memorization or regurgitation. Similarly, the court’s assessment on harm to ANI’s commercial interests leaned on the absence of demonstrated economic harm rather than a finding that AI training causes no cognizable harm to news publishers as a class. The decision, therefore, does not close the door on creators, but places a greater evidentiary burden on them. However, the court’s treatment of OpenAI’s commercial use of ANI’s works as immaterial was arguably too liberal. Given the ever-increasing scale of commercial returns generated by AI companies, treating such use as immaterial does not strike a balanced approach.
The decision may also have an unintended practical consequence. Creators and publishers may increasingly resort to self-help measures, such as stricter paywalls, robots.txt exclusions and crawler-blocking technologies to prevent their content from being used for AI training. This, in turn, may push AI developers towards licensing agreements with creators for access to their content. In fact, a market for licensing content for AI training is already emerging. OpenAI has entered into licensing arrangements with several major publishers.[7] Therefore, the question of unlicensed use, presently put to rest, may resurface.
4. Conclusion
The DHC’s ruling signals a pivotal moment in copyright law’s encounter with artificial intelligence. By affirming that the use of copyrighted works in training large language models can be transformative, the judgment protects innovation while underscoring AI outputs must not substantially reproduce or regurgitate original works. Its purposive reading of a pre‑AI statutory framework demonstrates judicial agility in addressing novel challenges. Yet, AI jurisprudence will inevitably evolve case by case. As systems grow more sophisticated, disputes over the provenance of training data and the economic impact on creators will intensify. Sustaining equilibrium between technological progress and copyright protection will demand ongoing dialogue among courts, technologists, creators and policymakers.
Author
Saumya Tiwari
[1] 2026 SCC OnLine Del 5291
[2] Section 52(1)(a), the Act provides the following acts shall not constitute an infringement of copyright, namely (a) a fair dealing with any work, not being a computer program (i) private or personal use, including research; (ii) criticism or review, whether of that work or any other work; (iii) reporting of current events and current affairs, including lecture delivered in public. The explanation states storing of any work electronically for the foregoing purposes including incidental storage of any computer program which in itself is not an infringing copy for the listed purpose shall not constitute copyright infringement.
[3] Section 62(2) of the Act permits a copyright owner to institute an infringement suit before a court within whose jurisdiction the plaintiff resides, carries on business, or personally works for gain.
[4] Section 20 of CPC permits a suit to be instituted in a court within whose jurisdiction the defendant(s) reside, carry on business, or personally work for gain; or where the cause of action, wholly or in part, arises.
[5] (1978) 4 SCC 118
[6] Section 14(a)(i) of the Act gives the copyright owner the exclusive right to reproduce the work in any material form, including storing it in any medium by electronic means.
[7] Condé Nast, The Financial Times, The Washington Post, The Atlantic & Vox Media et. al.