AI Regulation Tracker  /  Courts

Delhi HC Refuses to Injunct OpenAI's Training on ANI Content

The Delhi High Court has dismissed ANI's application for an interim injunction against OpenAI, holding on a prima facie view that storing ANI's news copy to train large language models falls inside the fair-dealing exception at section 52(1)(a) of India's Copyright Act. It is the most substantial Indian reasoning yet on AI training and copyright. It is also expressly interim: the judgment states that its observations have no bearing on the final outcome of the suit.

By Anthony Guerriero  ·  Published July 29, 2026  ·  Last updated July 29, 2026  ·  Courts

This is the ruling the Indian AI-copyright docket has been waiting on since November 2024, when the court framed four issues and appointed two Amici Curiae. It arrived on 24 July 2026, four months after argument closed.

The headline is accurate as far as it goes. On a prima facie view, the court held that OpenAI's storage of ANI's literary works for training purposes falls inside India's fair-dealing exception. What the headline omits is that the court dismissed the application on three independent grounds, only one of which is doctrinal, and then went out of its way to disclaim any effect on the trial.

What exactly did the court hold on fair dealing?

India's fair-dealing provision is narrower than the American fair-use standard. Section 52(1)(a) of the Copyright Act 1957 lists specific purposes rather than an open-ended balancing test, and the court worked through a purpose test and a fairness test in sequence.

On the fairness side, the court treated the public interest in research and technological development as a relevant consideration, reasoning that objectives including "developing innovative computational tools, disseminating information, promoting education, enhancing accessibility, and advancing artificial intelligence research" carry weight alongside the rights of copyright owners. At paragraph 255 it concluded:

Thus, on a prima facie view, all the factors for establishing the aspect of fair dealing stand satisfied in the present case and the fairness test stands fulfilled.

The findings section then states the holding in full at paragraph 256:

In light of the discussion above, both the purpose test as well as the fairness test under Section 52(1)(a) stand fulfilled. Hence, in my prima facie view, Open AI's acts of storage of the literary works of ANI for the training of its LLMs would fall under Section 52(1)(a) of the Copyright Act and hence, would not amount to infringement.

Note the precise scope. The finding is about storage for training. It is prima facie. It is about ANI's works in this proceeding. It is not a declaration that AI training is lawful in India.

Why did ANI actually lose the application?

Because it did not prove the output-side case, and this is the part practitioners should read most carefully. At paragraph 271 the court recorded that the outputs "were not substantially similar to ANI's original literary works," and that ANI "failed to satisfy this Court that any memorization and regurgitation of ANI's original literary works has happened through the responses generated by ChatGPT."

That is an evidentiary failure, not a doctrinal immunity. A different claimant with logs showing verbatim reproduction at scale is not answered by this judgment. The court also examined a specific example in the pleadings and found a separate ownership problem: quotes attributed to a third party were owned by that third party, with nothing to show assignment to ANI, so their reproduction "would not amount to infringement of ANI's copyright." It further noted that the ChatGPT response did not reproduce the remainder of the much longer original article.

Then came the discretionary grounds. At paragraph 272 the court held that ANI "failed to make out a prima facie case for grant of interim injunction," that balance of convenience ran against an injunction, and that irreparable injury "would be caused not only to Open AI but also to the public at large" if one were granted. Paragraph 273: the application is dismissed.

Who was heard, and why does that matter?

Unusually for an interlocutory application, the court built a record wider than the two parties. It appointed two Amici Curiae when it framed the issues in November 2024, and it admitted a series of intervenors, including the Federation of Indian Publishers as a national representative body of the Indian publishing industry.

One amicus submission is worth knowing because it names the theory the judgment ultimately leans toward. Professor Arul George Scaria argued that OpenAI's use of the works is "non-expressive use," on the basis that training data is used only to extract information such as patterns, trends and correlations in phrases and sentences rather than to reproduce expression. He also invited the court to consider whether an LLM can feasibly be developed without access to copyrighted material, and what follows if every developer must license from every rightsholder.

An intervenor put the narrower statutory point: that "storage" in section 14(a)(i) has to be read in the context of reproduction, so mere storage would not amount to infringement. Reading those submissions alongside the findings tells you which arguments the court found workable, which is more useful for the next case than the outcome is.

What should publishers and AI developers take from it?

For publishers, the strategic lesson is about evidence, not doctrine. The court refused relief partly because the record did not show verbatim regurgitation, and partly because it could not connect the pleaded example to a right ANI actually owned. Both are fixable problems in a differently prepared case, and both are cheaper to fix before filing than after.

For developers, the temptation is to read this as an Indian safe harbour. It is not, and the court said as much. The finding is provisional, it is confined to storage for training, and the fairness analysis leaned on public-interest considerations that a court could weigh differently where the defendant's use looks more substitutional and less research-like.

For US and European counsel, the useful point is structural. India reached a training-friendly interim result through a closed-list fair-dealing provision rather than an open fair-use balance, and it did so while leaving the output-side question open on the facts. That is a different route to a familiar destination, and it means Indian outcomes cannot be predicted by analogy to American reasoning.

One thing I am not going to do is characterise what this means for the other Indian proceedings involving AI training, or for the parties' prospects at trial. The court reserved the merits, and so will I.

Last verified: July 29, 2026

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any ruling applies to your situation with qualified counsel in the relevant jurisdiction.