ANI v. OpenAI: The Delhi High Court ’s On Whether Training Data and AI Outputs Infringe Copyright

On 24 July 2026, the Delhi High Court dismissed ANI Media’s application for an interim injunction against OpenAI in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC. The decision is one of India’s first detailed judicial treatment of copyright claims arising from the training and operation of a large language model. It is significant, but its limits are equally important: the Court’s findings are expressly prima facie, were made at the interlocutory stage, and will not determine the final outcome of the suit.

Background: AI and Copyright Infringement

ANI argued that OpenAI copied and stored ANI’s news content while preparing data used to train the large language models underlying ChatGPT, and the training and outputs constituted infringement of its copyright.

It alleged that ChatGPT reproduced ANI’s works in responses generated for users. OpenAI disputed both claims and relied, among other defences, on fair dealing under Section 52(1)(a) of the Copyright Act, 1957 (the “Copyright Act”).

The Court considered four issues:

• territorial jurisdiction;

• whether ChatGPT’s outputs infringed ANI’s copyright;

• whether electronic storage of ANI’s works for training amounted to reproduction; and

• whether such use was protected as fair dealing under Section .

  
  

Since the application sought interim relief, the Court assessed whether ANI had established a prima facie case, whether the balance of convenience favoured an injunction, and whether refusal of relief would cause irreparable injury.

Jurisdiction: foreign servers did not place the dispute beyond Indian courts

OpenAI argued that its models were trained and its data stored on servers outside India. The Court nevertheless held, at a prima facie stage, that it had territorial jurisdiction under Section 62(2) of the Copyright Act and Section 20 of the Code of Civil Procedure. ANI’s principal and registered offices were in Delhi; OpenAI offered services to users in India, including Delhi; and the allegedly infringing outputs were generated within the Court’s jurisdiction.

The Court also declined to isolate the final act of storage on overseas servers from the preceding chain of events, including access to works from India and transmission abroad. It considered the training and output claims sufficiently connected to reject, at this stage, the contention that applying Indian copyright law would necessarily be extra-territorial.

Output-infringement: no substantial reproduction of protected expression

The Court accepted that original literary works may remain copyright-protected even when freely accessible online, and that ANI had prima facie established ownership of original works published on its website. At the same time, copyright in news extends to the original form and manner of expression  and not to the underlying facts. The relevant question was therefore whether ChatGPT reproduced a substantial part of ANI’s protected expression.

ANI relied on examples of ChatGPT responses said to reproduce its reports. A central difficulty was temporal: the cited ANI articles were published in August and September 2024, whereas the relevant model-training cut-off dates were April 2022 for GPT-4 and April 2024 for GPT-4o. The Court therefore held that those examples could not establish memorisation arising from training on the cited works. It considered the responses more consistent with retrieval from external sources, potentially through retrieval-augmented generation, or RAG. Since infringement through RAG had not been pleaded, that issue was left open. However, it was of the prima facie view that the outputs generated by ChatGPT using RAG technique did not amount to infringement under Section 51 of the Copyright Act since the outputs were not substantially similar to ANI’s original literary works.

On the comparisons presented, the Court found no prima facie substantial similarity between ANI’s articles and ChatGPT’s responses when the works were considered as a whole. It also noted that one closer response followed a prompt asking ChatGPT to state “exactly” what an interviewee had said. Even then, the overlap concerned part of a quotation attributable to the interviewee, not protected expression shown to be owned by ANI. The Court accordingly held that ANI had not established a prima facie output-infringement claim. Questions concerning model memorisation and regurgitation more generally were treated as disputed matters requiring evidence at trial.

Storage for training: may be reproduction, but may be saved as non- infringing under Section 52

The judgment interprets Section 14(a)(i) of the Copyright Act broadly: electronic storage of a literary work, whether temporary or permanent, is reproduction. The purpose or intention behind storage is not relevant to the initial infringement inquiry. The decisive question was whether the use was nevertheless rendered non-infringing by Section 52 of the Copyright Act.

Fair dealing under Section 52

The Court treated Section 52 as an integral part of the statutory balance between owners’ rights and permitted uses. It formulated a two-stage inquiry under Section 52(1)(a): first, whether the dealing serves one of the specified statutory purposes; and second, whether the dealing is fair.

1. The purpose test

OpenAI relied on “private or personal use, including research” under Section 52(1)(a)(i). The Court held that “private” use is not confined to use by an individual and may include internal use by a private company. The opposing argument was that although the training process may have been private, the ultimate use and product were public and commercial. However, the Court considered the training process to be private, because the training corpus was not made available to the public as such.

The Court also construed “research” to not be confined to human activity and could include machine learning undertaken to advance AI systems. It rejected the proposition that commerciality alone excludes a use from Section 52(1)(a), noting that the provision does not expressly impose a non-commercial limitation, unlike several other clauses in Section 52.

The Court further held that the requirement that a copy must not itself be infringing, in the Explanation to Section 52(1)(a), applies to incidental storage of a computer program rather than to every electronically stored work. In any event, ANI’s works were taken from material freely available on its website; ANI had not alleged that OpenAI circumvented a paywall or obtained the works from an unauthorised source. On that reasoning, the purpose test was prima facie satisfied.

2. The fairness test

Rather than importing the US four-factor fair-use test wholesale, the Court derived three factors for the facts before it:

• whether use of ANI’s works was limited to training;

• whether the use competed with ANI or prejudiced its legitimate commercial interests; and

• whether ChatGPT’s functions served the public interest.

On the record then available, the Court found no instance of ANI’s works being used for a purpose other than model training, and no evidence that the training material was reproduced or communicated to the public. It also regarded ChatGPT’s functions as materially different from ANI’s news-gathering and syndication business. ChatGPT could summarise or synthesise information, but the Court did not consider it a substitute for ANI’s reporting. ANI had not produced evidence of lost subscribers, reduced licensing income or diminished advertising revenue attributable to ChatGPT.

Public interest also weighed in OpenAI’s favour. The Court referred to the use of LLMs in education, research, accessibility, software development, translation and access to information, as well as the practical burden that a requirement to license every item of training data could impose on AI development. These considerations led it to conclude, prima facie, that both the purpose and fairness requirements were satisfied.

Why the injunction was refused

The Court held that ANI had not established a prima facie case on either the training or output claim. The remaining injunction factors also favoured OpenAI. ANI had offered to license its content to OpenAI for USD 7.5 million, indicating that any proven loss could potentially be quantified. OpenAI had also stated that it had blocked ANI’s website for both training and ChatGPT search/RAG. By contrast, an injunction could significantly affect the operation of ChatGPT and its users.

Conclusion

The decision provides an important judicial framework for AI training under Indian copyright law, though not a final rule of general application. Its immediate holding is narrow: on the evidence before it, ANI did not establish a prima facie case for restraining OpenAI. The Court treated model training as electronic reproduction, but considered that reproduction prima facie protected under Section 52 of the Copyright Act as private use for research.

What the decision does, and does not do:

• It recognises electronic storage as reproduction. AI developers cannot assume that intermediate copying sits outside Section 14 of the Copyright Act; any protection must be found within Section 52 or another statutory limitation.

  
  

• It gives “private use” and “research” an expansive, technology-neutral reading. On the facts before the Court, internal model training by a commercial entity was capable of satisfying the statutory purpose test.

  
  

• It does not authorise all training on publicly available content. The findings were fact-specific and turned on the absence of proven regurgitation, market substitution or acquisition from unauthorised sources.

  
  

• It does not finally determine RAG liability. The Court treated RAG as distinct from training-based memorisation and left the issue open because it had not been pleaded.

  
  

• It does not resolve the treatment of other categories of works. News reports are heavily factual and receive protection primarily in their original expression. More creative works, paywalled materials, concentrated use of a single creator’s corpus, or outputs that reproduce protected expression may raise materially different issues.