High Court of Delhi at New Delhi, Single Judge | Date of Decision: 24th July, 2026 (Reserved on: 27th March, 2026) Case Number: CS(COMM) 1028/2024, I.A. 45301/2024 & I.A. 26192/2025 (order passed on I.A. 45300/2024, application under Order XXXIX Rules 1 and 2 CPC) Bench: Hon’ble Mr. Justice Amit Bansal
BACKGROUND
The Plaintiff, ANI Media Pvt. Ltd. (“ANI”), a news agency, filed the present suit against the Defendant, Open AI OpCo LLC (“Open AI”), alleging unauthorised use of ANI’s copyrighted news content in connection with Open AI’s large language model (“LLM”) application, ChatGPT. ANI’s claim rested on two distinct grounds: (i) a “training claim” that Open AI copied and stored ANI’s data to train its LLM without authorisation; and (ii) a “reproduction/output claim” that Open AI’s LLM reproduces ANI’s copyrighted works in the responses (outputs) generated for ChatGPT users. The application under consideration, I.A. 45300/2024, sought interim injunction under Order XXXIX Rules 1 and 2 of the Code of Civil Procedure, 1908.
On 19th November 2024, the Court framed four issues for determination and appointed two Amici Curiae (Mr. Adarsh Ramanujan and Professor Arul George Scaria) to assist the Court. Multiple intervention applications were subsequently filed and allowed (limited to the framed legal issues): the Federation of Indian Publishers, the Digital News Publishers Association (“DNPA”) and the Indian Music Industry (“IMI”) broadly supported ANI’s case; Flux AI Labs, the Indian Governance and Policy Project (“IGAP”) and the Broadband India Forum (“BIF”) broadly supported Open AI’s case. Extensive hearings took place over twenty-one dates between February 2025 and March 2026, after which judgment was reserved with liberty to file closing submissions.
The judgment begins with a detailed technical description of “Working of LLMs” covering training data collection (scraping of publicly available web data alongside licensed material, subject to quality and safety filtering and respecting website owners’ crawling restrictions), data preparation (tokenisation and embedding into vectors), model training (iterative prediction and refinement), post-training fine-tuning, user interaction via prompts and Retrieval-Augmented Generation (“RAG”) a feature allowing the LLM to retrieve information from external, authoritative sources outside its training data before generating a response. The Court clarified that “LLM(s)” and “ChatGPT” are used interchangeably in the judgment except where the “ChatGPT search” function is specifically referenced.
ISSUES FOR DETERMINATION
The Court addressed, in the following order (determined by their logical/foundational interrelationship), the four issues framed on 19th November 2024:
Issue No. 4: Whether the Courts in India have jurisdiction to entertain the present suit, considering that Open AI’s servers are located in the United States of America.
Issue No. 2: Whether the use by Open AI of ANI’s copyrighted data to generate responses for its users amounts to infringement of ANI’s copyright.
Issue No. 1: Whether the storage by Open AI of ANI’s data (news content, claimed to be protected under the Copyright Act, 1957) for training ChatGPT amounts to infringement of ANI’s copyright.
Issue No. 3: Whether Open AI’s use of ANI’s copyrighted data qualifies as “fair dealing” (the Court clarified that although the issue, as originally framed, used the words “fair use,” it should be read as “fair dealing,” the expression used in Section 52(1)(a) of the Copyright Act, 1957) in terms of Section 52 of the Copyright Act, 1957.
Issues 1 and 3 were considered together, given their interconnection (Open AI’s fair dealing defence being raised specifically in response to the storage/training allegation).
KEY HOLDINGS OF THE COURT
On Issue No. 4 (Jurisdiction): The Court held, at the prima facie stage, that it had territorial jurisdiction to entertain the suit under both Section 20 of the CPC and Section 62(2) of the Copyright Act, 1957, since ANI’s principal place of business and registered office are within the Court’s jurisdiction, Open AI specifically targets/offers its services to users in India (including within the Court’s jurisdiction) and the alleged infringing activities (responses generated for ANI’s prompts) occurred within the jurisdiction. On the separate question of whether the Copyright Act could apply to the training claim given that training occurs on servers outside India, the Court held that the storage of ANI’s works on US servers is merely the “terminal step in the chain of events” beginning with access of copyrighted works from India and their transmission abroad and that the Act does not require the Court to sever this chain and examine only the last step; accepting Open AI’s argument would permit evasion of Indian copyright law by relocating the terminal link abroad. Relying on Blueberry Books v. Google India (Division Bench) and Neetu Singh v. Telegram (Coordinate Bench, holding that merely locating servers outside India does not divest Indian courts of jurisdiction over copyright disputes), the Court held that the training and output claims could not be artificially severed, since the output claim is based on the training of the LLM and the output is reproduced within the Court’s jurisdiction.
On Issue No. 2 (Reproduction/Output Claim): The Court held, at the prima facie stage, that ANI had failed to make out a case of infringement based on ChatGPT’s generated responses. The Court’s analysis proceeded in two parts:
(i) Memorization/regurgitation: The Court held that the nine illustrative instances in the plaint (and specifically the detailed example involving an interview of Neeraj Chopra’s mother) could not be the product of memorization, because the training cut-off dates for the relevant LLMs (April 2022 for GPT-4; April 2024 for GPT-4o) predated the publication of all fourteen works ANI alleged were infringed (published August–September 2024). The Court held that such post-training-cutoff responses were, prima facie, attributable to the RAG technique (retrieving live information from ANI’s website) rather than memorization of training data and that whether RAG-based outputs amount to infringement was an unpleaded issue not before the Court in this application.
(ii) Substantial reproduction: Applying the principles from R.G. Anand v. Delux Films (comparison of works as a whole; no copyright in idea/theme/facts, only in form, manner and expression) and Eastern Book Company v. D.B. Modak (the “skill and judgment” test for originality, as opposed to “sweat of the brow”; no copyright in facts per se), the Court examined the specific comparison relied upon by ANI (the Neeraj Chopra’s mother interview) and held that ChatGPT’s first response (to a non-adversarial prompt) was not a substantial or near-exact reproduction but added its own commentary; that ANI’s second prompt (“tell me exactly what she said”) was an adversarial prompt designed to elicit a specific response; and that even then, the only similarity was a quoted portion of the interview subject’s own words in respect of which, applying Section 17(cc) of the Copyright Act (first ownership of copyright in a public address/speech vesting in the speaker), the Court held that Neeraj Chopra’s mother, not ANI, would prima facie be the first owner of copyright in the quotation, with no evidence of assignment to ANI. Comparing the articles and ChatGPT outputs as a whole (following Leo Burnett and Bikramjeet Singh Bhullar, which disapprove of dissecting works into parts for comparison) rather than isolated extracts and noting that Open AI’s rebuttal note on the remaining illustrations went unrebutted by ANI, the Court found no substantial similarity. The Court distinguished the foreign authorities relied upon by ANI (GEMA v. Open AI, Associated Press v. Meltwater, Advance Local Media v. Cohere, Infopaq, Meltwater (UK)) on the basis that those cases involved either verbatim/near-verbatim reproduction on non-adversarial prompts or different factual/legal contexts, none of which were established on the facts before the Court.
On Issues No. 1 and 3 (Training/Storage Claim and Fair Dealing): The Court held these issues together and found in Open AI’s favour, holding that the storage of ANI’s literary works for LLM training purposes falls within Section 52(1)(a) of the Copyright Act, 1957 (fair dealing) and therefore does not constitute infringement. The Court’s reasoning proceeded through the following sub-holdings:
(a) Scope of Section 14(a)(i): The Court held that the right of reproduction under Section 14(a)(i), which expressly includes “storing of it in any medium by electronic means” (introduced by the 1994 amendment), is broad and does not distinguish between temporary and permanent storage; however, Section 14 operates “subject to the provisions of this Act,” including Section 52, such that an act falling within Section 52 is not infringing. Following the Division Bench in B.D. Bhandari and the Single Bench in Rameshwari Photocopy Services, the Court held that Section 52 is not a “proviso” or “exception” to Section 51 but an integral, independently operative part of the Copyright Act defining user rights and therefore must be given a liberal, non-restrictive interpretation consistent with the Supreme Court of Canada’s reasoning in CCH Canadian v. Law Society of Upper Canada.
(b) Purpose Test under Section 52(1)(a)(i) (“private or personal use, including research”): The Court rejected ANI’s contention that commercial use is categorically excluded from Section 52(1)(a)(i), holding that wherever the legislature intended to exclude commercial use, it did so expressly (citing Sections 52(1)(ad), 52(1)(k)(ii), 52(1)(l), 52(1)(n), 52(1)(o)) and that this limitation is conspicuously absent from Section 52(1)(a). The Court distinguished Jiwan Publishing House and TIPS Industries v. Wynk Music (both involving distribution of copyrighted material to third parties for commercial exploitation) as inapplicable, since Open AI’s use is an internal training process not involving distribution of ANI’s works to any third party. On the “non-infringing copy” requirement in the Explanation to Section 52(1)(a), the Court held, based on the placement of commas in the statutory text, that this limitation applies only to the “incidental storage of any computer programme,” not to other categories of works and in any event ANI had not alleged that Open AI accessed its content through unauthorized means (e.g., bypassing a paywall) rather than freely available website content. On “private or personal use,” the Court held, relying on dictionary definitions and the Supreme Court’s decision in Academy of General Education, Manipal v. B. Malini Mallya, that “private” is broader than “personal” and can encompass use by a closed entity (not confined to individuals) where the relevant data/process is not accessible to the public; since the training data is accessible only to the LLM itself and not to any human third party, the Court held this use to be private. On “research,” applying the doctrine of updating construction (citing State (Through CBI) v. S.J. Choudhary) to interpret the term in light of modern technological developments, the Court held that the training process involving iterative analysis, pattern extraction and refinement from literary works constitutes a form of research and that research exceptions need not be confined to human researchers.
(c) Fairness Test: Noting that there is no single, uniformly applied Indian test for “fair dealing” (and that the US four-factor test, while occasionally referenced by Indian courts, is not directly applicable given doctrinal differences), the Court formulated three fact-specific factors for this case, informed by Article 9(2) of the Berne Convention: (i) whether Open AI’s use of ANI’s works is limited to LLM training; (ii) whether such use results in economic competition causing actual or potential damage to ANI’s legitimate interests; and (iii) whether ChatGPT’s functions serve the overall public interest. Applying these factors, the Court held: (i) no instance was shown of ANI’s works being used by Open AI for any purpose beyond training; (ii) ChatGPT’s functions (content creation, research assistance, translation, education, etc.) are fundamentally different in character and purpose from ANI’s core business of news reporting and syndication, such that ChatGPT’s outputs do not substitute for ANI’s news articles and no evidence was placed on record showing any loss of market share or subscription revenue to ANI a conclusion supported by reasoning from the US decisions in Bartz v. Anthropic, Kadrey v. Meta Platforms and Authors Guild v. Google regarding the “transformative” character of LLM training use; and (iii) the societal benefits of LLM technology (education, healthcare, research, accessibility, knowledge dissemination) establish the public interest factor.
On Balance of Convenience and Irreparable Injury: The Court held that ANI had the technical ability to block its website from Open AI’s web crawlers (an “opt-out” mechanism) but had not exercised it, while Open AI had, in fact, internally blocked ANI’s website from both training-data scraping and the ChatGPT search/RAG function. The Court noted ANI’s own prior offer (3rd October 2024) to license its content to Open AI for a fee of USD 7.5 million, holding this demonstrated that ANI’s claim, if ultimately successful, was quantifiable and compensable in damages whereas an injunction against Open AI could cause injury not compensable in monetary terms, including to the broader public (millions of Indian ChatGPT users) and to the developing Indian AI/LLM ecosystem, a factor the Court treated as an important “fourth factor” in assessing interim relief in intellectual property matters.
Final holding and operative order: The Court held that ANI had failed to make out a prima facie case for interim injunction on either the training/storage claim (falling within the Section 52(1)(a) fair dealing exception) or the output/reproduction claim (no substantial reproduction shown and no memorization established given the training-cutoff timeline). Balance of convenience and irreparable injury considerations also weighed against injunction. I.A. 45300/2024 was accordingly dismissed, with the Court clarifying that its observations were made solely for the purposes of adjudicating this interim application and would have no bearing on the final outcome of the suit.
STATUTORY PROVISIONS INVOLVED
Section 13(1)(a) of the Copyright Act, 1957, establishing that copyright subsists in “original literary… works,” was applied to hold that ANI’s news articles and interviews, as published, qualify as original literary works (a position not disputed by Open AI), with the Court separately finding, based on a sample Professional Services Agreement, that ANI prima facie holds ownership under Section 17 in respect of works created by its personnel.
Section 14(a)(i) and 14(a)(iii) of the Copyright Act, 1957, conferring the exclusive rights of reproduction (including electronic storage) and communication to the public, were the rights ANI claimed were infringed; the Court held these rights operate “subject to” Section 52.
Section 17 (including the proviso Section 17(cc), concerning first ownership of copyright in a publicly delivered address/speech) was applied specifically to hold that the quoted remarks of Neeraj Chopra’s mother in an interview would, prima facie, belong to her as speaker, not to ANI as reporting agency, absent evidence of assignment.
Section 51 of the Copyright Act, 1957, defining when copyright is infringed, was the operative infringement provision, read together with Section 2(m) (definition of “infringing copy,” which incorporates the phrase “in contravention of the provisions of this Act,” thereby excluding acts falling within Section 52).
Section 52(1)(a) of the Copyright Act, 1957 (and its Explanation), providing the “fair dealing” exception for “private or personal use, including research,” “criticism or review,” and “reporting of current events and current affairs,” was the central provision construed and applied by the Court in Open AI’s favour on the training/storage claim, with extensive discussion of its legislative history (1957 original text; 1994 amendment broadening “private study” to “private use, including research”; 2012 amendment adding “personal use” and the Explanation on electronic storage).
Section 62(2) of the Copyright Act, 1957, permitting suits to be instituted where the plaintiff resides or carries on business, was applied (together with Section 20 CPC) to establish territorial jurisdiction.
Section 20 of the Code of Civil Procedure, 1908, governing territorial jurisdiction generally, was applied in conjunction with Section 62(2) of the Copyright Act.
REASONING OF THE COURT
The Court’s reasoning throughout reflected a deliberate methodology of first establishing the Court’s jurisdictional competence (since this was foundational), then addressing the output/reproduction claim (since a finding of no substantial reproduction materially narrowed the scope of the fair dealing inquiry on the training claim) and finally undertaking the detailed statutory construction exercise on Sections 14 and 52 together with the factual assessment of fairness.
On jurisdiction, the Court’s reasoning emphasized that copyright infringement, where part of a continuous process spanning data access in India and ultimate processing/storage abroad, cannot be artificially fragmented to immunize the India-based conduct merely by locating the final storage step offshore a reasoning explicitly aimed at preventing circumvention of Indian copyright law through server-location arbitrage.
On the output claim, the Court’s reasoning was tightly anchored to the documented training-cutoff dates, which the Court treated as dispositive of the memorization theory for the specific illustrations pleaded (since the underlying works were published after training had concluded), redirecting the analysis toward RAG-based retrieval as the more plausible technical explanation while expressly declining to rule on whether RAG-sourced outputs could themselves constitute infringement, since this theory had not been pleaded. The Court’s close, side-by-side textual comparison of ChatGPT’s responses to ANI’s original articles (including noting the shift in ANI’s own prompting strategy from a neutral query to an explicitly adversarial “tell me exactly what she said” prompt) reflected a methodologically rigorous application of the “compare as a whole, not dissected parts” principle from R.G. Anand and Leo Burnett and the Court treated ANI’s failure to rebut Open AI’s consolidated comparison note on the remaining illustrations as effectively conceding the point for those instances.
On the fair dealing analysis, the Court’s reasoning was structured and comparative: having first established (through close reading of related sub-clauses of Section 52(1), several of which explicitly exclude commercial use) that the absence of such exclusionary language in Section 52(1)(a) was a deliberate legislative choice rather than an oversight, the Court then applied the doctrine of updating/”always speaking” construction (drawing on S.J. Choudhary) to extend “research” to encompass machine-based learning processes undertaken for the ultimate benefit of human users reasoning explicitly that the 2012 legislature “could not have imagined” AI/LLM technology, but that this gap should be filled through purposive, technology-neutral interpretation rather than treated as foreclosing the exception’s application. In assessing fairness, the Court’s three-factor framework was expressly derived from first principles (balancing the author’s rights against the Article 9(2) Berne Convention-derived concern for “normal exploitation” and “legitimate interests”) rather than mechanically importing the US four-factor test, which several Indian authorities (and all parties, by broad consensus) agreed was not directly transplantable into the Indian statutory scheme; the Court nonetheless drew persuasive support from the reasoning (rather than the holdings) of recent US decisions (Bartz, Kadrey, Authors Guild) regarding the transformative character of using copyrighted works to train a general-purpose generative tool, reasoning that the purpose and character of ChatGPT’s varied functions differ fundamentally from the purpose and character of ANI’s news-reporting business, such that no market substitution effect was established.
On balance of convenience, the Court’s reasoning drew a pointed inference from ANI’s own commercial conduct its unexercised ability to block web-crawler access and its own prior monetary licensing offer to Open AI to conclude that ANI’s claimed harm was not of a character requiring extraordinary interim protection, while treating the potential disruption to the broader Indian AI ecosystem and millions of ChatGPT users as a weighty countervailing public-interest consideration distinct from, but reinforcing, the balance-of-convenience analysis.
DOCTRINAL SIGNIFICANCE
Within the bounds of what was actually decided, this is a landmark, closely reasoned interim-stage judgment described by the Court itself as addressing “uncharted areas” where courts must “catch up with technology” and is among the first substantive judicial treatments in India of whether and how the training of large language models on copyrighted news content and the generation of AI outputs based on such training, implicate the Copyright Act, 1957. Its most significant and likely most-cited holdings are: (i) that electronic storage of copyrighted works for the purpose of training an LLM can fall within the “private or personal use, including research” exception under Section 52(1)(a)(i), interpreting “private” to extend beyond individual human use to a closed, non-publicly-accessible process and “research” to extend to machine learning via the doctrine of updating construction; (ii) that Section 52(1)(a) is not subject to an implied non-commercial-use limitation, since the legislature expressly imposed such limitations elsewhere in Section 52(1) where intended; (iii) that comparison for substantial-similarity purposes must be of the works as a whole and that post-training-cutoff illustrations cannot support a memorization theory; and (iv) the formulation of a bespoke, Berne Convention-informed three-factor fairness test for Indian copyright law, explicitly departing from mechanical reliance on the US four-factor test. The judgment is, however, expressly and repeatedly characterized by the Court as a prima facie, interlocutory determination, with the Court cautioning that its observations bear no weight on the final outcome of the suit, which remains to proceed to trial on full evidence including on disputed factual questions the Court declined to resolve at this stage (such as whether RAG-based retrieval and output generation could independently constitute infringement, an issue the Court noted was not pleaded and would require trial). As a judgment emanating from a single bench with no appellate pronouncement reflected in the text and given the extraordinary scale and diversity of intervention (encompassing publishers, the music industry, AI start-ups and policy think tanks, together with two Amici Curiae), this decision is likely to be of substantial persuasive significance for subsequent AI-copyright litigation in India, pending any appellate review or legislative clarification of the application of Sections 14, 51 and 52 of the Copyright Act to generative AI technologies.
Compulsory Licensing Copyright Act Copyright Act 1957 copyright case laws Copyright Enforcement Copyright Infringement Copyright Law Copyright Rules Descriptive Marks India Indian IP Framework Indian Patent Act Indian Patent Law Indian Trademark Law Intellectual Property Law IP Law India Patent Claims Patent Enforcement Patent Infringement Patent law Patent Revocation Patents Act Patents Act 1970 Pharmaceutical Patents Section 9 Trade Marks Act Section 29 The Patent Act 1970 Trademark Infringement Trademark Registration Trade Marks Act 1999 TRIPS Compliance