I.     Introduction

1.    On July 24, 2026, the Delhi High Court declined to grant an interim injunction against OpenAI OPCO LLC (“OpenAI“) in a copyright suit filed by ANI Media Private Limited’s (“ANI“) (CS(COMM) 1028/2024). This interim order is India’s first judicial examination of how territorial jurisdiction principles apply to cross-border AI training, artificial intelligence (“AI“)-generated outputs, and the legality of training large language models (“LLMs“) under the (Indian) Copyright Act, 1957 (“Copyright Act“).

2.     ANI alleged infringement on two grounds: (a) OpenAI’s unlicensed use of its copyrighted news content to train the LLMs underlying ChatGPT; and (b) ChatGPT’s reproduction of its works in responses to user prompts.

II.    Issues Framed

1.     The Court framed four issues for determination at the interim stage:

(a)   Whether OpenAI’s storage of ANI’s data (news, claimed to be protected under the Copyright Act, 1957) for training its software, i.e., ChatGPT, would amount to infringement of ANI’s copyright.

(b)   Whether OpenAI’s use of plaintiff’s copyrighted data to generate responses for its users would amount to infringement of ANI’s copyright.

(c)   Whether OpenAI’s use of ANI’s copyrighted data qualifies as ‘fair use’ (to be read as ‘fair dealing’, the expression actually used in Section 52(1)(a)) in terms of Section 52 of the Copyright Act, 1957.

(d)   Whether the Courts in India have jurisdiction to entertain the present lawsuit, given that OpenAI’s servers are located in the United States of America.

2.     The Court appointed two amici curiae – Mr. Adarsh Ramanujan and Professor Arul George Scaria – and permitted intervention by the Federation of Indian Publishers, Digital News Publishers Association and Indian Music Industry (supporting ANI), and Flux Labs AI Private Ltd., IGAP Project LLP and Broadband India Forum (supporting OpenAI).

III.   Key Findings of the Court

A.     Jurisdiction: Foreign Servers Do Not Oust Indian Copyright Jurisdiction

1.      OpenAI’s jurisdictional objection raised two questions:

(a)   Whether the Court had territorial jurisdiction to entertain the suit at all: The Court held it had jurisdiction under Section 62(2) of the Copyright Act and Section 20 of the Code of Civil Procedure, 1908, since ANI’s registered office and principal place of business are in Delhi, and OpenAI marketed its services to, and generated the allegedly infringing responses for, users across India, including Delhi. Accordingly, a substantial part of the cause of action arose there.

(b)   Whether the Copyright Act applies to the training claim given that OpenAI’s servers are located outside India: The Court held that the server’s overseas location was not decisive, characterising it as merely the ‘terminal step’ in a chain that began with access to the works from India. Drawing on the Division Bench’s rejection of a similar defence in Blueberry Books v. Google India, 2016 SCC OnLine Del 3338, the Court held that the Copyright Act does not require looking only at the final step while disregarding preceding acts within India. Relying on Neetu Singh v. Telegram, 2022 SCC OnLine Del 2637, the Court added that accepting OpenAI’s argument would let infringers evade Indian law merely by locating their servers abroad.

2.     Accordingly, the Court held, on a prima facie view, that the acts formed a continuous chain connected with India, making it inappropriate to reject jurisdiction at the threshold.

Key Takeaway: The Court’s prima facie observations suggest that foreign AI companies cannot avoid Indian jurisdiction merely by locating their training servers outside India. So long as the copyright owner is based in India and the AI company’s services and outputs reach Indian users, both the training and output limbs of an infringement claim can, at least at the prima facie stage, be brought before courts in India. However, whether the Copyright Act governs all aspects of overseas training remains open for determination at trial.

B.     AI Outputs: ChatGPT’s Responses Do Not, Prima Facie, Infringe ANI’s Copyright

1.      Whether ANI’s Copyright Subsists in Its News Content: The Court held, prima facie, that ANI’s news articles, reports and interviews qualify as “original literary works” under Section 13(1)(a) – a point undisputed by OpenAI. On ownership, which OpenAI contested, the Court relied on a sample Professional Services Agreement vesting copyright in ANI for works created by its personnel, holding ANI was prima facie the owner despite the content being freely available on ANI’s website. Ownership of individual works was left for trial.

2.     Whether the outputs generated by ChatGPT infringed ANI’s copyright: The Court distinguished between ‘reproduction of protected expression’ and ‘use of underlying ideas or facts’, reiterating that copyright protects only expression, not facts or news itself – a point particularly relevant to news reporting, where facts remain free for public use while the expression used to convey them is protected. The Court therefore compared ANI’s articles with ChatGPT’s outputs for substantial reproduction of expression, not mere factual overlap, examining: (a) whether OpenAI’s models had memorised and “regurgitated” ANI’s works; and (b) whether the outputs substantially reproduced ANI’s protected expression. The Court found as follows:

(a)   No prima facie evidence of memorisation: The training cut-off dates for the underlying models – April 2022 (GPT-4) and April 2024 (GPT-4o) – preceded publication of the ANI articles relied on in the plaint, so the cited examples more likely reflected ChatGPT’s search / Retrieval-Augmented Generation (“RAG“) function retrieving live content, rather than memorisation. The Court held that ANI’s examples were insufficient, at the prima facie stage, to establish systematic memorisation or verbatim reproduction of its works.

(b)   No substantial similarity: The Court invoked the judgment in G. Anand v. Delux Films, AIR 1978 SC 1613, for the principle that similarity must be assessed between works as a whole, rather than isolated parts, and held that ANI’s examples did not, prima facie, demonstrate substantial reproduction of protected expression. The Court observed that some factual overlap was inevitable in news reporting and declined to presume infringement merely from training or from isolated examples, holding that each allegedly infringing output must be independently examined for substantial reproduction – an analysis for trial, not for this interim stage.

The Court also distinguished foreign authorities cited by ANI in support of their contention. For instance, it distinguished Munich Regional Court’s decision in GEMA v. OpenAI as it involved verbatim lyric reproduction on non-adversarial prompts, unlike ANI’s adversarial prompting; and distinguished Associated Press v. Meltwater and Advance Local Media v. Cohere because those cases involved direct scraping or verbatim reproduction which was absent in the present case.

3.     In light of the above, the Court held that ANI had failed, at the prima facie stage, to make out infringement on the output claim, leaving final determination to trial.

Key Takeaway: Adversarial or extraction prompting that elicits only paraphrased summaries, rather than verbatim or near-verbatim reproduction, may be insufficient, at the prima facie stage, to establish output-based copyright infringement. If this reasoning is upheld at trial, AI platforms that present factual information in their own expression, without substantially reproducing the claimant’s protected expression, may not be held liable on output-based infringement claims.

C.     AI Training and Storage: Use for Training Falls Within the Fair Dealing Exception under Section 52

1.     Fair Dealing: This interim order is the first time an Indian court has considered whether storing copyrighted works for AI training engages the exclusive right of reproduction, and whether such storage qualifies as fair dealing. The Court held that this storage does engage ANI’s exclusive right under Section 14(a)(i), but since Section 14 operates “subject to” the Act, including Section 52, it would not amount to infringement if it falls within the Section 52(1)(a) fair dealing exception.

2.    The Twin Tests – ‘Purpose Test’ & ‘Fairness Test’: Rather than treating fair dealing narrowly, the Court held Section 52 is not a proviso to Section 51 but an integral part of the Act that independently defines users’ rights. This requires purposive interpretation that balances copyright owners’ rights against the public interest in research, learning and innovation. The Court articulated a two-step test for Section 52(1)(a): the ‘Purpose Test’ (whether the use falls within an enumerated purpose) and the ‘Fairness Test’ (whether the use amounts to ‘fair dealing’).

(a)     Purpose Test

(i)     Fair dealing is not confined to non-commercial activities: A central issue was whether OpenAI, despite running a commercial AI platform, could invoke Section 52(1)(a)(i), which protects fair dealing for “private or personal use, including research”. The Court rejected ANI’s contention that commercial use is incompatible with such protection, holding that the legislature – aware of the distinction between commercial and non-commercial use – had expressly excluded commercial use only in specific other provisions of Section 52(1). This limitation could not be read into Section 52(1)(a)(i), so OpenAI’s commercial character alone did not disentitle it from the fair dealing defence absent a specific statutory bar.

(ii)    The Explanation to Section 52(1)(a) and the ‘Non-Infringing Copy’ Requirement: The Court also addressed the Explanation to Section 52(1)(a), which states that storing a work in electronic medium for the purposes mentioned in the clause – “including the incidental storage of any computer programme which is not itself an infringing copy” – does not constitute infringement. Based on the placement of commas in the Explanation, the Court held that the “not itself an infringing copy” qualification applies only to the incidental storage of computer programmes, and not to other electronically stored works. In any event, drawing on the distinction recognised in Bartz v. Anthropic between lawfully acquired works and copies taken from unauthorised ‘shadow libraries’, the Court found that since OpenAI had obtained ANI’s works from information freely available on its website, without resorting to unauthorised sources or bypassing paywalls, therefore, no infringing copy had been stored.

(iii)   Meaning of ‘Private Use, Including Research’:

• Private Use: ANI argued that ‘private use’ under Section 52(1)(a)(i) applies only to individuals, given the overlapping meanings of ‘private’ and ‘personal’. The Court disagreed, citing dictionary definitions that extend ‘private’ to a ‘particular person, group, or class’ or ‘company’, and Academy of General Education, Manipal v. B. Malini Mallya (2009) 4 SCC 256, which recognised institutional activity as ‘private use’. Since OpenAI’s training data was accessible only to its own LLM models, the Court found this to be private use, declining to treat ‘private’ and ‘personal’ as interchangeable, which would render the term ‘private’ needless.

• Research: The Court held that ‘research’, though undefined under the Copyright Act, is an intermediary investigative process that precedes an output, and found that LLM training – screening, extracting and converting literary works into machine-readable inputs to generate the most probable statistical output – could constitute research. Applying the doctrine of updating construction, the Court held that ‘research’ extends to machine learning and is not confined to human actors, since such a limitation would be regressive. Accordingly, OpenAI’s training was, prima facie, capable of constituting research, satisfying the Purpose Test under Section 52(1)(a).

(b)     Fairness Test: To determine fairness, the Court formulated three factors broadly in line with Article 9 of the Berne Convention – the first two addressing whether ANI’s rights were affected by OpenAI’s use of its works and the third addressing public interest:

(i)     Scope of Use – Training Versus Reproduction in Outputs: First, on whether OpenAI’s use was limited to training or extended to communicating ANI’s works to the public, the Court – building on its earlier finding that ANI had not established memorisation or regurgitation – found no instance of OpenAI using ANI’s works beyond training its LLMs.

(ii)    Market Impact on ANI: Second, on whether training placed OpenAI in economic competition with ANI or prejudiced its commercial interests, the Court found that ANI had placed nothing beyond bare averments on record to show any loss of market share or subscription revenue, and that OpenAI’s use would not result in market substitution of ANI’s works. The Court reasoned that ANI’s business is limited to news reporting and syndication, whereas ChatGPT serves multiple functions – including research, translation and summarisation – so its responses could not be treated as substitutes for ANI’s news articles.

(iii)   Public Interest in Generative AI: Third, on the wider public interest served by generative AI, the Court found substantial public benefits – including improved access to information and support for education, scientific research and software development, etc. – holding this to be a relevant factor in assessing fairness, though not an independent defence in itself.

3.     On a cumulative assessment, the Court held that both the Purpose Test and the Fairness Test under Section 52(1)(a) were satisfied at the prima facie stage – OpenAI’s storage of ANI’s works for training thus fell within the fair dealing exception and did not infringe under Section 51.

Key Takeaway: If upheld at trial, an AI developer that trains its models on data that is lawfully and freely accessible online – without circumventing paywalls or crawler blockers, and confining that data’s use to internal training rather than public reproduction – may be able to invoke the Section 52(1)(a) fair dealing exception even where the platform is commercial. Content owners seeking to restrict such training should treat technical access controls, such as paywalls and crawler blockers, as a practical first line of defence.

IV.    Balance of Convenience and Irreparable Injury

1.     Having found that OpenAI had raised substantial, arguable defences on both claims, the Court turned to the principles governing interim relief and noted:

(a)    Absence of Demonstrated Damage to ANI: ANI could have blocked its website from OpenAI’s web crawlers, for both training and ChatGPT’s search/RAG function, but did not; nor was there evidence that OpenAI’s activities caused ANI to lose subscribers or suffer any loss to its news syndication business.

(b)    Adequacy of Monetary Compensation: Any injury ultimately suffered by ANI could be compensated through monetary relief if infringement were established at trial – reinforced by ANI’s own October 2024 offer to license its content to OpenAI for USD 7.5 million, confirming that its claim was quantifiable – whereas OpenAI could not be adequately compensated if wrongly restrained by an interim injunction.

(c)     Public Interest in AI Development: An injunction restraining training would have far-reaching consequences beyond this dispute, given generative AI’s transformative role across education, healthcare and governance, and its dependence on broad access to publicly available data. Restraining training would be detrimental to AI and LLM development in India and would adversely affect millions of ChatGPT users, many of whom are not paid subscribers.

The balance of convenience therefore favoured OpenAI, and ANI’s application was dismissed.

Key Takeaway: Courts are unlikely to grant an interim injunction against AI training or deployment where the content owner has not deployed available technical safeguards and cannot show quantifiable, demonstrated loss. Parties seeking urgent interim relief in AI-related copyright disputes should come prepared with concrete evidence of market harm. This dismissal, however, disposes of only the interim injunction application – the underlying suit continues to trial, and the order remains open to challenge in appeal.

V.     Conclusion

This interim order marks India’s first substantive judicial engagement with the interplay between copyright law and generative AI. While declining interim relief to ANI, the Court left significant questions open for trial, including the scope of copying in AI training, whether memorisation or regurgitation may occur elsewhere, and ownership of individual works.

For AI developers, the interim order offers a favourable position in India, at least for now: training on publicly accessible data may qualify as fair dealing under Section 52(1)(a), even for a commercial platform, provided the data is not obtained from unauthorised sources or paywalls and the outputs do not substantially reproduce protected expression. Given that many issues remain open for trial, safeguards like respecting crawler blockers and paywalls, and monitoring outputs for substantial reproduction, remain relevant to sustaining this defence.

This decision also arrives against the backdrop of the Department for Promotion of Industry and Internal Trade’s (“DPIIT“) Working Paper on Generative AI and Copyright (Part I), ‘One Nation One License One Payment’, released for public consultation in December 2025, which proposes a mandatory blanket licensing and royalty framework for AI training – departing from a pure fair dealing-based approach as being considered under this case. Since the Working Paper lacks the force of law and no legislative amendment has been enacted, the Court’s interim findings on the applicability of Section 52(1)(a) to AI training remain the operative legal position for now, though future statutory changes along DPIIT’s proposed lines, or the final judgment, could alter this position.

Interestingly, a related question arises under India’s Digital Personal Data Protection Act, 2023 (“DPDP Act“). Section 17(2)(b) exempts processing for “research, archiving or statistical purposes,” subject to a bar on using personal data to make decisions about specific individuals, and prescribed safeguards including purpose limitation and data minimisation. On one view, the Court’s reasoning on “research” may guide interpretation of the DPDP Act exemption, as both provisions bundle a stated purpose with limiting conditions. On another view, the regimes differ fundamentally: copyright focuses on reproduction of protected expression (the training/output inquiry), whereas the DPDP Act addresses processing of personal data and the risk of individual harm. Whether the Court’s approach extends to the DPDP Act, whether AI training can satisfy its purpose-specific requirements, and whether AI-generated outputs used in decision-making would breach the bar on decisional use, remain open questions.