CS(COMM) 1028/2024, I.A. 45301/2024 & I.A.26192/2025

Date of decision 24th July 2026

This judgment is among the first significant Indian decisions to examine how traditional copyright principles apply to generative artificial intelligence and large language models. In refusing interim relief to ANI Media Pvt. Ltd., the Delhi High Court considered whether OpenAI’s use of publicly available news content for training ChatGPT, and the generation of responses through ChatGPT, amounted to copyright infringement or was protected as fair dealing under Section 52 of the Copyright Act, 1957. The ruling is notable for its treatment of AI training, retrieval-augmented generation, commercial use, market substitution and public interest in the development of AI technologies.

Background

Against this backdrop, ANI Media Pvt. Ltd. filed a copyright infringement suit against OpenAI OpCo LLC before the Delhi High Court, alleging unauthorised use of ANI’s copyrighted news content in connection with ChatGPT. ANI advanced two principal claims: (a) that OpenAI copied and stored ANI’s works to train its large language models; and (b) that ChatGPT generated outputs that reproduced, or were substantially similar to, ANI’s copyrighted works.

The Court noted that the case raised novel questions at the intersection of copyright law and artificial intelligence. Since the Copyright Act, 1957 was enacted long before the emergence of generative AI and large language models, the Court approached the dispute by examining how existing copyright principles should apply to technologies that rely on large-scale data collection, storage, tokenisation, model training and retrieval-augmented generation.

Issues

1. Whether OpenAI’s storage of ANI’s copyrighted data for training ChatGPT amounted to infringement of ANI’s copyright.
2. Whether ChatGPT’s generation of responses using ANI’s copyrighted data amounted to infringement of ANI’s copyright.
3. Whether OpenAI’s use of ANI’s copyrighted data qualified as fair dealing under Section 52 of the Copyright Act, 1957.
4. Whether Indian courts had jurisdiction to entertain the suit, given that OpenAI’s servers and training infrastructure were located outside India.

Contentions of the Parties

ANI’s Contentions

ANI’s case was that OpenAI had used its copyrighted news content without permission. ANI alleged that OpenAI copied, stored and processed ANI’s news articles and other literary works while training ChatGPT. According to ANI, the fact that its works were available on the internet did not make them free for unrestricted use, nor did it dilute ANI’s copyright in the original expression of its reports.

ANI also characterised the technical steps involved in AI training, scraping, storage, tokenisation and vectorisation as acts of reproduction or adaptation in electronic form. ANI argued that the conversion of text into machine-readable formats did not take the use outside copyright law; it merely changed the form in which the copyrighted work was copied or processed.

Separately, ANI objected to ChatGPT’s outputs. It argued that certain responses reproduced, paraphrased or substantially copied ANI’s interviews and news reports, thereby amounting to unauthorised communication of its works to the public. ANI said this was not merely a case of using underlying facts, but of appropriating the form and substance of its protected expression.

On fair dealing, ANI’s principal objection was that OpenAI’s use was commercial and could not properly be described as private or personal use, or as research, within the meaning of Section 52(1)(a). ANI also urged the Court not to import the broader US fair use doctrine into Indian law, particularly where large-scale copying of copyrighted works was being justified as AI training.

ANI further submitted that OpenAI’s conduct had commercial consequences for ANI’s business. By making summaries or extracts of ANI’s content available through ChatGPT, ANI argued that OpenAI diverted traffic away from ANI’s own platforms, reduced licensing opportunities and undermined the market for ANI’s news content.

OpenAI’s Contentions

OpenAI first objected to the Court’s jurisdiction over the training claim. It argued that the relevant acts of training and storage took place outside India, on servers located in the United States, and that Indian copyright law should not be applied extraterritorially merely because ChatGPT was accessible in India.

On the merits, OpenAI’s central answer was that ChatGPT does not store or reproduce training data once training is complete. Any copying during training, it said, was only an intermediate and non-expressive step in teaching the model statistical relationships, linguistic patterns and contextual associations. OpenAI maintained that copyright protects original expression, not facts, ideas, news events, grammar, syntax or language patterns.

OpenAI also challenged ANI’s examples of alleged infringement. It pointed out that the articles relied on by ANI were published after the relevant training cut-off dates for OpenAI’s models. On that basis, OpenAI argued that those examples could not prove memorisation or regurgitation from training data. At most, the responses reflected facts, short summaries or material generated through search or RAG-type functionality, rather than reproduction of ANI’s protected expression.

For the fair dealing defence, OpenAI submitted that its use fell within “private or personal use, including research” under Section 52(1)(a) of the Copyright Act. It emphasised that the training process was internal and closed, and that the training material was not made available to the public. OpenAI also argued that commercial use is not automatically excluded from fair dealing, particularly where the statute does not expressly impose a non-commercial limitation.

Finally, OpenAI relied on the broader utility of ChatGPT. It submitted that the technology serves public-facing functions such as research, education, translation, accessibility, summarisation and innovation, and that restricting model training through interim relief would have consequences beyond the parties to the suit.

Contentions of the Intervenors

Intervenors Supporting ANI

1. The Digital News Publishers Association, Indian Music Industry and Federation of Indian Publishers broadly supported ANI.
2. They argued that AI training involves unauthorised copying and storage of copyrighted works and that such use cannot be treated as fair dealing merely because it is technologically complex or machine-readable.
3. They submitted that Section 14(a)(i) gives copyright owners an exclusive right over reproduction, including storage in electronic form, and that temporary or intermediate copying can still amount to infringement.
4. They contended that OpenAI’s use was commercial, large-scale and substitutive, and should not be protected under Section 52(1)(a).
5. They further argued that Indian law does not contain a broad US-style fair use defence and that transformative use should not be imported into Indian copyright law without statutory basis.

Intervenors Supporting OpenAI

1. IGAP Project LLP, Broadband India Forum and Flux AI Labs broadly supported OpenAI.
2. They submitted that LLMs do not ordinarily reproduce training material but learn statistical patterns, relationships and linguistic structures from large datasets.
3. They argued that AI training is a form of research or private use and that requiring licences from every copyright owner would make AI development economically unviable, particularly for Indian AI developers.
4. They contended that ChatGPT’s outputs are materially different from ANI’s articles and do not substitute ANI’s news syndication business.
5. They also emphasised the public interest in access to AI tools, technological innovation and the development of domestic AI systems.

Court Ruling

Jurisdiction

The Court held, prima facie, that it had territorial jurisdiction. ANI’s principal office was in Delhi, OpenAI made its services available to users in India, and the alleged infringing outputs were generated within the jurisdiction of the Court. The Court rejected the argument that the location of servers outside India, by itself, defeated jurisdiction.

Output/Reproduction Claim

The Court held that ANI had not established, at the interim stage, that ChatGPT’s outputs were substantial reproductions of ANI’s copyrighted works. The Court noted that copyright does not protect facts or news events, but only the original form, manner and arrangement of expression. It found that the examples relied on by ANI were not substantially similar when compared with ANI’s works as a whole.

The Court also found that the illustrative ANI articles were published after the relevant training cut-off dates for OpenAI’s models. Therefore, the examples could not support ANI’s allegation that ChatGPT memorised those articles from training data. The Court considered that such outputs were more likely connected to live retrieval or RAG-type functionality, an aspect not specifically pleaded as the foundation of ANI’s claim.

Training/Storage Claim and Fair Dealing

The Court held that storage of a literary work in electronic form is covered by the reproduction right under Section 14(a)(i) of the Copyright Act. However, Section 14 is subject to Section 52, which sets out acts that do not constitute infringement. The Court therefore considered the storage and fair dealing issues together.

Applying a broad and liberal interpretation of Section 52(1)(a), the Court held, prima facie, that OpenAI’s storage of ANI’s works for training LLMs fell within “private or personal use, including research”. It reasoned that AI training is a closed, internal process and can qualify as research when understood in light of technological developments. The Court held that “private” does not mean only use by an individual. It can include use by a company or closed group, provided the use is internal and not made available to the public. The Court adopted a broad reading of “research,” observing that research may be undertaken in commercial settings.

The Court further held that commercial use is not automatically excluded from Section 52(1)(a), since the provision does not expressly restrict the defence to non-commercial use. It also held that the “non-infringing copy” limitation in the Explanation to Section 52(1)(a) applies only to incidental storage of computer programmes, and not to every electronically stored work.

Fairness and Public Interest

For the fairness analysis, the Court considered whether OpenAI’s use was limited to training, whether it caused economic competition or market substitution, and whether ChatGPT served broader public interest. It held, prima facie, that OpenAI’s use was limited to training, that ChatGPT’s functions were fundamentally different from ANI’s news syndication business, and that ANI had not shown loss of market share or subscription revenue.

The Court also found that LLMs and ChatGPT serve important public interests, including education, research, translation, accessibility, software development and access to information. On this basis, the Court held that the purpose and fairness tests under Section 52(1)(a) were satisfied at the interim stage.

Interim Injunction

The Court refused to grant an interim injunction. It held that ANI had not made out a prima facie case of infringement either on the training/storage claim or on the output claim. It also found that the balance of convenience favoured OpenAI, since an injunction could affect the functioning of ChatGPT and have wider implications for AI development.

The Court noted that ANI’s claim was quantifiable in monetary terms, especially since ANI had itself offered a licence to OpenAI. It also considered that ANI and its subscribers had technical options to block crawling or scraping.

Key Takeaways

1. AI training is not treated the same as public reproduction: The Court drew a practical distinction between internal use of works for model training and outward-facing reproduction of those works in outputs. At the interim stage, storage for training could fall within Section 52(1)(a) if it remained internal and did not substitute the copyright owner’s market.
2. Indian fair dealing remains purpose-based: The Court did not import the US four-factor fair use test as binding law. Instead, it assessed the statutory purpose under Section 52(1)(a) and then examined fairness through factors such as market substitution, functional difference and public interest.
3. Commercial use is relevant, but not decisive: OpenAI’s commercial character did not, by itself, defeat the fair dealing defence. The Court focused instead on what the use was for, whether the training material was exposed to the public, and whether the use caused market harm.
4. Evidence of market harm will matter: The Court gave weight to the absence of material showing loss of subscribers, reduced licensing revenue, traffic diversion or market substitution. Future plaintiffs may need stronger evidence of commercial impact, not merely assertions of unauthorised use.
5. News copyright protects expression, not facts: The judgment reinforces that news publishers must show copying of original expression, arrangement or presentation not merely use of underlying facts or events. This makes substantial similarity a demanding test in AI-output disputes involving news content.
6. Technical controls may influence injunctions: The Court noted that ANI and its subscribers had options to block crawling or scraping. That observation may become relevant in future cases when courts assess urgency, balance of convenience and equitable relief.