Introduction

The rapid growth of generative artificial intelligence has raised a pressing legal question across the world, whether feeding copyrighted material into a machine learning system amounts to copyright infringement. The Delhi High Court has now weighed in on this question in a dispute between ANI Media Pvt. Ltd., one of India’s prominent news agencies, and Open AI, the company behind ChatGPT. ANI accused OpenAI is using its news articles without permission to train ChatGPT and of generating responses that reproduce its content. OpenAI defended its practice as a transformative and research-orientated activity protected under the fair dealing provisions of Indian copyright law.

In a detailed judgement running into more than a hundred pages, the Court examined how large language models work, how Indian copyright law applies to their training, and whether an interim injunction should be granted while the underlying suit is pending trial. The judgement is significant not only for the news and publishing industry but for the entire artificial intelligence ecosystem operating in or serving users in India.

Case Summary

Factual and Procedural Background

ANI is engaged in gathering, producing and distributing news content, which it licenses to subscribers, including media houses and digital platforms. It claimed that OpenAI, without obtaining any licence, scraped and stored its articles as part of the raw data used to train the large language models that power ChatGPT and that ChatGPT’s responses to user queries at times closely mirrored its reports.

ANI filed a commercial suit seeking a permanent injunction and damages, along with an interim application asking the court to restrain OpenAI from continuing to store, use or reproduce its works during the pendency of the suit.

Open AI resisted the application, arguing that its servers and training operations are located outside India, that ANI’s material forms an infinitesimally small part of a vast training corpus drawn from the internet, and that the entire process is protected as fair dealing under the Copyright Act.

During the proceedings, several parties were allowed to intervene, including bodies representing digital news publishers and the music industry supporting ANI, and organisations representing the broadband and artificial intelligence industry supporting OpenAI. The Court also had the benefit of assistance from two amici curiae on the technical workings of large language models and the applicable legal principles.

Dispute Before The Court

The Court identified four core questions that needed to be answered before deciding whether to grant an interim injunction.

• The first was a threshold question of jurisdiction, since OpenAI’s servers where the actual training takes place are located in the United States; the court had to decide whether an Indian court could still entertain a copyright claim relating to that training activity.
• The second question was whether the answers that ChatGPT generates for its users, if they resemble ANI’s news reports, amount to infringement of ANI’s copyright.
• The third was whether the very act of collecting and storing ANI’s articles as training data, independent of what ChatGPT later generates, itself amounts to infringement.
• The fourth, and perhaps the most consequential, question was whether such storage and use, even if it would otherwise infringe copyright, is protected by the fair dealing exception available for private or personal use, including research, under the Copyright Act.

In simple terms, ANI’s case was that its hard earned journalistic work was being taken without payment or permission and turned into a commercial product that competes with its own business, while Open AI’s case was that training an AI model on publicly available text is no different from a student reading widely to learn, and that the law specifically protects such research oriented use.

Reasoning And Analysis Of The Court

Jurisdiction

On jurisdiction, the Court held that since OpenAI conducts commercial activity accessible to and monetised from users located in Delhi, including subscription payments, the Court could exercise jurisdiction over the dispute even though the actual training servers are situated abroad. The Court reasoned that the effects of the allegedly infringing activity are felt within its jurisdiction and that this was sufficient at the interim stage.

ChatGPT Outputs and Copyright

On the question of whether ChatGPT’s outputs infringe ANI’s copyright, the Court examined the instances of alleged copying placed on record by ANI and found that these had been generated after OpenAI’s training process on the relevant data had already concluded. The Court held that isolated post training examples could not establish that ChatGPT actually memorises and reproduces ANI’s articles as a general practice, and that in any event the responses shown did not amount to a substantial reproduction of ANI’s original expression, since news reporting largely conveys factual events which enjoy thinner copyright protection than works of greater creative content.

The Court distinguished several foreign precedents relied upon by ANI, including cases involving verbatim extracts of news articles and cases involving musical works, on the ground that those situations involved either actual verbatim copying or subject matter with a much higher degree of original expression than typical news reporting.

Storage and Fair Dealing

The heart of the judgement lies in its treatment of the storage and fair use questions, which the Court examined together. Open AI took shelter under Section 52(1)(a)(i) of the Copyright Act, which exempts fair dealing for the purpose of private or personal use, including research.

ANI argued that this defence could not apply because Open AI’s use was commercial, because the underlying copy of the work had to be a lawfully obtained copy, and because training an AI model is not really research in the sense contemplated by the statute.

The Court rejected the argument that commercial use automatically falls outside fair dealing, noting that wherever the legislature intended to restrict a Section 52 exception to non-commercial use, it said so expressly, as seen in other clauses of the same section, and that this restriction is conspicuously absent from Section 52(1)(a).

The Court also drew support from the Canadian Supreme Court’s decision in CCH Canadian Ltd v Law Society of Upper Canada, which held that research must be given a large and liberal meaning and is not confined to non-commercial contexts.

The Court further reasoned that the term research, as used in the Copyright Act, is broad enough to include the process by which a machine learning system is trained to recognise patterns in language, since such training is fundamentally an exercise in extracting information and patterns rather than republishing the original text as such.

Fairness Test

The Court then applied what it described as a fairness test, considering whether OpenAI’s use was limited to training its models, whether it caused economic harm to ANI by substituting for its own products in the market, and whether the activity served the larger public interest.

Drawing upon recent American decisions concerning the use of copyrighted books to train large language models, the Court accepted that the purpose and character of training an LLM is highly transformative, since the resulting model does not seek to replicate or supplant the original works but instead learns from them to generate new and different kinds of output, capable of translation, summarisation, drafting assistance and a wide range of other tasks quite different from simply reading a news report.

The Court noted that ANI had not placed any concrete evidence of lost subscription revenue or market share attributable to OpenAI’s activities and therefore could not show that the use displaced ANI’s own commercial position.

On public interest, the Court observed that large language models contribute meaningfully to education, research, accessibility and the dissemination of knowledge, and that these considerations weigh in favour of treating the training activity as a fair dealing.

Having found that both the purpose test and the fairness test were satisfied, the Court concluded that Open AI’s storage of ANI’s works for training purposes falls within the protection of Section 52(1)(a) of the Copyright Act.

On the question of balance of convenience, the Court weighed ANI’s claim, which it noted was capable of being compensated in monetary terms if the suit ultimately succeeds, against the practical consequences of an injunction that would require Open AI to identify and delete specific training data, which could disrupt its operations and would also affect the wider public that relies on ChatGPT for a variety of purposes.

Final Decision Of The Court

The Court held, on a prima facie basis, that Open AI’s storage of ANI’s works for training the large language models underlying ChatGPT is protected under Section 52(1)(a) of the Copyright Act and does not amount to infringement.

It further held that ANI had not established that ChatGPT’s outputs are a substantial reproduction of its news articles or that any memorisation and regurgitation of its content had taken place.

Concluding that ANI had failed to make out a prima facie case and that the balance of convenience and the risk of irreparable injury both favoured OpenAI and the public at large, the court dismissed ANI’s application for an interim injunction.

The Court was careful to clarify that its observations were made only for the purpose of deciding the interim application and would have no bearing on the final outcome of the suit, which remains to be decided after trial.

Point Of Law Settled

The judgement clarifies, at least at the interim stage, that the use of copyrighted literary works, including news content, to train large language models can fall within the fair dealing exception for private or personal use, including research, under Section 52(1)(a) of the Copyright Act, 1957.

It settles that commercial motive alone does not disqualify a party from claiming this defence, since the statute expressly restricts other specific exceptions to non-commercial use but does not do so for Section 52(1)(a).

It also underlines that a copyright holder seeking to resist such a defence on grounds of economic harm must produce concrete evidence of market substitution or lost revenue, rather than relying on general apprehension.

The decision is likely to significantly influence how Indian courts approach similar disputes involving artificial intelligence companies and copyright owners and will be closely watched as the broader suit proceeds to trial.

Case Details

• Can ChatGPT Be Forced to Display Website Links?
• The Illusion of Safety: How Google’s Gemini and OpenAI’s ChatGPT Enable Non-Consensual Digital Undressing
• The OpenAI Civil War: Elon Musk vs Sam Altman and the Fight for AI’s Future
• Fake News and Misinformation: The Threat to an Informed Society
• The UAE’s Digital Shield: Combating Fake News, Cybercrime, and Online Harm