Copyright in the AI Era: the Delhi High Court’s ANI V. OpenAI Decision

The relationship between copyright law and artificial intelligence has been difficult to define. Copyright law was developed around familiar concepts- authors, publishers, copies and readers. Generative artificial intelligence has introduced a very different way of creating and processing information.

The recent decision of the Delhi High Court in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, decided on 24 July 2026, brings this tension directly before an Indian court.

The dispute arose from ANI’s allegation that OpenAI had used its copyrighted news content to train the large language models underlying ChatGPT without its permission. ANI also raised concerns about the ability of ChatGPT to generate responses based on its content.

The Court declined to grant ANI an interim injunction. At the interim stage, it found, prima facie, that OpenAI’s storage of ANI’s works for the purpose of training its language models fell within the fair-dealing exception under Section 52(1)(a) of the Copyright Act, 1957. The Court also found that ANI had not established, on the material before it, that ChatGPT was reproducing its copyrighted works in a manner that amounted to infringement.

What was the dispute about?

ANI’s case essentially concerned two different uses of its content.

The first related to training. ANI alleged that OpenAI had accessed, copied and stored its copyrighted works and used them in training the models underlying ChatGPT.

The second related to outputs. ANI placed before the Court instances where ChatGPT had generated responses relating to ANI’s reporting and argued that such outputs infringed its copyright.

The distinction between these two issues is important. Using a copyrighted work as part of the process of training a model raises one set of questions. Producing an output that reproduces a substantial portion of the original work raises another.

The Court therefore had to consider not merely whether ANI’s works had been copied, but why they were copied, how they were used and what ultimately came out of that process.

Copyright protection does not disappear because content is available online

One point on which the Court was clear is that material being freely available on the internet does not make it free from copyright.

ANI’s news content was publicly accessible. That, however, did not deprive ANI of copyright in its original literary works. This is an important distinction. The internet has made information easy to access, but accessibility is not the same as permission to reproduce.

The more difficult question was whether the particular act of storing and using that material for training an AI model could nevertheless be protected by an exception to copyright infringement. That brought Section 52 of the Copyright Act into focus.

Can AI training amount to “research”?

Section 52(1)(a) of the Copyright Act permits certain acts of fair dealing, including fair dealing with a work for “private or personal use, including research”. ANI argued that this provision could not reasonably extend to the activities of a commercial AI company.

The Court did not accept that argument at the interim stage.

A particularly interesting part of the judgment is the Court’s treatment of the word “research”. The Court considered how a statutory provision enacted in a very different technological era should be understood in the context of modern developments. The Court applied the principle of “updating construction” and held, prima facie, that research cannot necessarily be understood as an activity undertaken only by human beings. In the Court’s view, modern research and learning increasingly involve machine learning and computational systems. On that basis, the Court found that OpenAI’s use of ANI’s works for training its language models could fall within the statutory reference to research. This is arguably one of the most consequential aspects of the decision.

The Court was not simply asking whether the Copyright Act expressly mentions artificial intelligence. It was asking whether an existing statutory expression can be applied to a technology that did not exist when the legislation was enacted.

Commercial use did not automatically take OpenAI outside Section 52

ANI also relied on the fact that OpenAI is a commercial enterprise. The Court, however, did not treat the commercial nature of OpenAI’s business as sufficient, by itself, to take the use outside Section 52(1)(a). The Court instead looked at the purpose of the particular use. According to the Court’s prima facie assessment, the material was being stored within OpenAI’s system for the purpose of training the models and was not being made publicly available as a substitute for ANI’s original works. This distinction matters.

The judgment should not be understood as saying that commercial entities are generally free to use copyrighted works without permission. Rather, it suggests that commerciality alone may not determine whether a particular use qualifies as fair dealing. The nature, purpose and consequences of the use remain relevant.

Fair dealing is more than simply identifying a permitted purpose

Finding that AI training could fall within “research” did not end the Court’s analysis. The Court also considered whether the use was fair. In doing so, it examined the nature of OpenAI’s use, whether the use resulted in market substitution and the broader public interest involved in the development and use of the technology. The Court found that OpenAI’s use of ANI’s works for training was different in character from ANI’s own business of creating and supplying news content. It also found that ANI had not, at the interim stage, produced sufficient evidence of actual market substitution or demonstrated economic harm caused by OpenAI’s use.

This part of the decision is likely to be particularly important in future cases. Copyright disputes involving AI may increasingly turn on evidence of actual or likely economic harm, rather than simply establishing that copyrighted material was included somewhere in the training process.

What about the answers generated by ChatGPT?

The second part of ANI’s case concerned the outputs generated by ChatGPT. Here again, the Court drew an important distinction. The mere fact that a system has been trained using copyrighted material does not necessarily mean that every answer generated by that system reproduces the copyrighted work. ANI was required to demonstrate that the outputs in question actually reproduced a substantial part of its protected expression. On the material before the Court, ANI was unable to establish sufficient similarity between the relevant ChatGPT outputs and its original works. The Court also found that ANI had not established that OpenAI’s models had memorised or “regurgitated” its works.

The Court similarly considered responses generated using Retrieval-Augmented Generation (RAG) and found, on the examples before it, that they were not substantially similar to ANI’s original literary works.

This does not mean that an AI-generated response can never infringe copyright. If a system produces a response that reproduces a substantial part of a copyrighted work, the legal position may be very different. The question will ultimately depend on the content of the output and the nature and extent of the reproduction.

The location of OpenAI’s servers was not enough to keep the case outside India

Another interesting aspect of the decision concerns jurisdiction. OpenAI argued that its servers were located outside India and that the training process therefore took place outside the territorial jurisdiction of the Delhi High Court. The Court did not accept this as sufficient to defeat ANI’s proceedings. The Court considered the connection between the alleged acts, ANI’s presence in India and the availability and use of OpenAI’s services in India. It found, at the interim stage, that it had jurisdiction to entertain the proceedings. For technology companies operating across jurisdictions, this part of the decision is worth watching closely. The physical location of servers may not, by itself, determine whether an Indian court can exercise jurisdiction over conduct connected with India.

What the judgment does and does not decide

It is tempting to describe the decision as a ruling that “AI companies can use copyrighted content to train their models.” That would be an overstatement.

The Court was deciding an application for an interim injunction, not finally determining the copyright suit. Its findings were expressly described as prima facie, and the Court made it clear that its observations would not prejudice the final adjudication of the suit.

The judgment therefore does not establish that:

• every form of AI training on copyrighted material is lawful;
• publicly available content can always be used without permission;
• copyright owners have no remedy against AI companies; or
• AI-generated outputs can never infringe copyright.

What it does establish, at least at this stage of the proceedings, is that the existing fair-dealing provisions of Indian copyright law are capable of applying to the training of AI models.

That is an important finding in itself.

What does this mean for content owners?

For publishers, news organisations, authors and other copyright owners, the judgment highlights the importance of evidence.nSimply showing that an AI system has had access to copyrighted material may not be enough to obtain an injunction. A rights holder may need to demonstrate what was copied, how it was used, whether the material was retained or reproduced, and whether the resulting system or outputs have caused or are likely to cause meaningful commercial harm. This may also encourage copyright owners to think more carefully about the technical and contractual measures available to them.

At the same time, the Court’s recognition that publicly accessible content remains protected by copyright should not be overlooked. The judgment does not turn the internet into a copyright-free zone.

What does this mean for AI developers?

For AI developers, the decision provides some judicial support for the argument that training models using publicly accessible copyrighted material can, in appropriate circumstances, fall within Section 52.

But the decision is not a blanket exemption.

The circumstances surrounding the use will matter. Developers will therefore need to continue paying attention to issues such as the provenance of training material, access restrictions, licensing arrangements, data governance and the possibility of outputs reproducing protected expression.

The safer lesson from the judgment is not that “anything available online can be used”, but rather that the purpose and manner of use matter.

A significant first step, not the final word

Perhaps the most interesting feature of ANI v. OpenAI is that the Court has attempted to work within the existing Copyright Act rather than treating artificial intelligence as requiring an entirely separate legal regime. The judgment shows that concepts such as “research” and “fair dealing” can potentially be interpreted in a manner that accommodates technological developments.

At the same time, the decision leaves several difficult questions unanswered.

What happens where the material used for training is behind a paywall? Does the analysis change where the copyright owner has expressly prohibited automated access? How should the law deal with books, music, photographs, films or computer code? What evidence should establish that a model has memorised copyrighted material? And where should the line be drawn between learning from a work and reproducing it?

These questions are likely to occupy Indian courts, legislators and the technology industry for years to come.

Conclusion

The Delhi High Court’s decision in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC is an important first chapter in what is likely to be a much longer conversation between copyright law and artificial intelligence.

For now, the Court has indicated that training an AI model on copyrighted material may, in appropriate circumstances, qualify as fair dealing under Section 52(1)(a) of the Copyright Act. At the same time, it has reaffirmed that copyright protection does not disappear merely because a work is available online and that AI-generated outputs must still be examined for actual reproduction of protected expression.

The judgment therefore strikes a balance between two competing concerns: protecting the economic rights of creators while allowing technological development and research to continue. Whether that balance survives the final adjudication of the suit\ and whether the legislature ultimately considers it sufficient, is a question that remains open.

For businesses developing or using AI, and for creators whose works are being used by such systems, one thing is already clear: copyright and AI can no longer be treated as separate legal conversations.