The Delhi High Court’s decision on July 24, 2026, dismissing the interim application filed by ANI Media Private Limited (“ANI”), a leading Indian news agency, against OpenAI OPCO LLC (the U.S.-based developer of ChatGPT) (“OpenAI”), brings into focus the delicate balance between copyright protection and AI innovation. This is the first significant judicial decision by an Indian court on copyright infringement arising from both the training of large language models (“LLMs”) and their output generation.
The Court’s interim order, which prioritizes AI development and public benefit over interim copyright protection, makes it incumbent on both content owners and AI developers to reassess their contractual frameworks, terms of use and data practices.
From Jurisdiction to Fair Use: The Court’s Decision
The Court framed four key issues, i.e. whether: (i) OpenAI’s storage of ANI’s content for LLM training amounted to copyright infringement; (ii) OpenAI’s output generation infringed ANI’s copyright; (iii) OpenAI could avail of the “fair use” defense under the Indian Copyright Act, 1957 (“Copyright Act”); and (iv) the Indian courts had territorial jurisdiction considering that OpenAI’s servers are located in the U.S. The Court’s reasoning on each issue carries distinct implications for rights holders and AI developers operating in or targeting India.
Jurisdiction
On jurisdiction, the Court held that territorial jurisdiction was established under the Copyright Act as well as the Code of Civil Procedure, 1908 due to ANI’s principal place of business being in Delhi and ChatGPT being accessible to users in India.
Notably, the Court rejected OpenAI’s argument that no infringing activity took place within the jurisdiction of the Court, because training occurred on U.S.-based servers. The Court held that the storage of ANI’s works in the U.S. is simply a terminal step in a chain of events beginning with accessing of copyrighted works from India and their transmission abroad. Accepting OpenAI’s position would sever that chain of events, and in turn undermine the rights of copyright owners in India.
This reasoning is significant as it effectively precludes an AI company that accesses Indian content remotely from avoiding the jurisdiction of Indian courts by merely processing the data offshore.
LLM outputs
In the context of the output claim, ANI had presented various responses generated by ChatGPT as evidence of reproduction. The Court observed that the referenced responses were not substantially similar to ANI’s copyrighted works and hence did not support a finding of copyright infringement. Evidence indicating that ChatGPT did not reproduce the article in its original form, even on adversarial prompting (i.e., prompts specifically designed to induce verbatim reproduction), further bolstered this conclusion. The Court also distinguished ANI’s case from the Munich Regional Court’s decision in GEMA v. OpenAI, where GEMA was able to obtain a verbatim reproduction of song lyrics using non-adversarial prompting.
Further, the articles that ANI alleged were infringed were published after the training cut-off dates for OpenAI’s models, which significantly weakened ANI’s claim that the models had memorized and reproduced its copyrighted content. This was a critical flaw in ANI’s evidence and a cautionary note that rights holders seeking interim relief against AI developers will need to align their evidentiary strategy with how and when these models are actually trained.
LLM training and fair use
The most significant part of the decision is the Court’s prima facie view that training LLMs using stored literary works does not constitute infringement. The Court applied a two-step framework, a “purpose test” and a “fairness test,” and held that OpenAI’s use falls under the exemption provided for ‘private or personal use, including research’ under Section 52(1)(a)(i) of the Copyright Act.
Under the purpose test, the Court gave “private” an expansive interpretation, holding that it extends to companies and cannot be limited to individuals. Similarly, the Court applied the doctrine of updating construction and observed that “research” encompasses machine learning and that confining this exception to human acts alone would impede societal progress. The Court clarified that OpenAI’s use is entirely internal, i.e., training data is not made available to a third party in any form once training is complete. Critically, the Court took the view that use of the copyrighted materials for commercial purposes would not preclude reliance on this defense, particularly since the legislature imposed a non-commercial use requirement elsewhere in Section 52 but conspicuously omitted it from Section 52(1)(a).
On the fairness test, the Court again relied on the internal nature of OpenAI’s use, noting that the LLMs are not designed to reproduce or communicate the training material to the public, and also found that ANI had not demonstrated any loss of market share or a reduction in subscription revenues. The Court also observed that trained LLMs serve considerable public benefits. In dismissing ANI’s application, the Court held that an interim injunction at this stage would be detrimental to the growth of AI and to LLMs being developed in India.
The interpretation that internal use for commercial purposes qualifies as “private or personal use” under Section 52(1)(a)(i), which for the most part shields AI developers from infringement claims so long as training data is not communicated externally, will potentially be the most contested aspect of the order.
Implications for the Industry
Content owners
The decision recognizes the role of technological barriers (such as paywalls or crawler blockers) by copyright owners to prevent scraping by LLMs. Pertinently, the Court’s decision was heavily influenced by the fact that ANI’s content was freely and publicly available and that ANI had opted not to block web crawlers, despite such options being available to it. Copyright owners should treat this as a lesson and consider implementing technological barriers as a first line of defense, which would not only restrict access, but also strengthen any future infringement claims. It may be harder for a court to characterize any access to content by an AI developer in circumvention of active restrictions deployed by the content owner as “fair” dealing.
In addition, while the Court did not reference ANI’s website terms of use, content owners should not overlook contractual protection as a complementary layer. Clear and enforceable terms of use that expressly prohibit scraping, automated access and the use of website content for AI training may reserve the option for an independent contractual claim, apart from a copyright infringement claim. Such terms, where possible, should require affirmative consent, e.g., in a clickwrap manner, prior to access to website content.
Copyright owners should also ensure that any infringement evidence relied on is consistent with the training cut-off dates for the relevant models and demonstrates that the relevant content is reproduced by the LLM through ordinary use or non-adversarial prompts. As noted above, ANI’s case was significantly weakened due to shortcomings in the evidence adduced in support of its infringement claim. Content owners should also be able to document market harm systematically, including through analytics on traffic diversion, loss of subscribers or advertising revenue decline.
Finally, copyright owners should consider entering into commercially reasonable licensing deals with organizations such as OpenAI. OpenAI’s existing arrangements with the Associated Press, Financial Times and Condé Nast were noted by the Court. ANI itself had offered OpenAI a license for USD 7.5 million, a fact the Court referenced in holding that ANI’s claim is quantifiable and thus capable of compensation in monetary terms. The Court’s observations in relation to licensing also indicate that Indian courts may view this as the appropriate market-based solution. Such licensing agreements with AI developers should clearly delineate permitted uses, restrict onward training (i.e., training of subsequent or derivative models) and include audit rights to verify compliance.
AI developers
The Court’s decision certainly provides significant relief to AI developers, but it should not be treated as a blanket permission to train LLMs on all available content without restrictions. The decision was premised on specific factual findings – the use was entirely internal; there was no substantial reproduction in the LLM outputs; and ANI could not demonstrate market harm. AI developers should ensure that their training practices are structured to comply with these conditions and that scraping and data collection practices are limited to content that is genuinely freely available. Accessing content through unauthorized means or by bypassing technological restrictions deployed by copyright owners may substantially weaken any fair dealing defense.
AI developers should also continue to incorporate adversarial testing during the training and implement other safeguards such as ‘noise injection’ i.e., introducing random variations to prevent verbatim memorization of training data, output filtering and memorization detection tools to minimize risk of regurgitation.
The most reliable risk mitigation strategy is undoubtedly to enter into licensing arrangements with publishers and content owners. Key negotiation points include the scope of permitted training uses, the treatment and ownership of outputs, data retention and deletion timelines and the continued use of licensed content by models that have already been trained, post-termination of the agreement.
Companies deploying AI tools
While the decision does not directly concern companies deploying third-party AI tools, such entities may be exposed to risk if the underlying models were trained on infringing content. This risk may be higher where AI-generated outputs substantially reproduce copyrighted material. Companies integrating AI through API access, SaaS platforms or enterprise licenses should ensure that vendor agreements include suitable representations that training data was lawfully obtained, appropriate indemnification for IP-related claims and implement internal usage policies that require human review of AI-generated content before commercial use.
Conclusion
While decided at the interim stage, the decision provides the first significant judicial guidance on the application of Indian copyright law to AI training and offers a certain level of business certainty for AI developers operating in or targeting India. This is, however, unlikely to be the final word and will inevitably invite further litigation testing the boundaries of fair dealing in the context of AI training. The Court itself observed that its findings would have no bearing on the final outcome of the suit and several important issues await determination at trial. Stakeholders should therefore not treat this as settled law, but only as an indicative framework within which their practices, technical safeguards, contractual arrangements and litigation strategies should be reviewed.
This insight has been authored by Reshma (Vaidya) Gupte (Counsel) and Prakriti Anand (Senior Associate). They can be reached on rgupte@snrlaw.in and panand@snrlaw.in, respectively, for any questions. This insight is intended only as a general discussion of issues and is not intended for any solicitation of work. It should not be regarded as legal advice and no legal or business decision should be based on its content.
© 2026 S&R Associates