- within Intellectual Property topic(s)
- in United States
- with readers working within the Banking & Credit and Law Firm industries
- within Law Department Performance, Real Estate and Construction and International Law topic(s)
ANI Media Pvt. Ltd. v. OPEN AI OPCO LLC, I.A. 45300/2024 in CS(COMM) 1028/2024 (Delhi High Court, 24 July 2026, Bansal J.)
Introduction
On 24 July 2026, a single judge of the Delhi High Court, Justice Amit Bansal, dismissed the interim injunction application filed by the Indian news agency Asian News International (ANI) against OpenAI. The decision is the first substantive judicial engagement in India with the central legal question of the generative-AI era: whether ingesting copyright-protected material to train a large language model (LLM) infringes the rights of the content owner. Reported to run to roughly 135 pages, the interim order is expressly a prima facie determination made at the injunction stage rather than a final adjudication of the suit. Even so, because there was no Indian precedent on point, its reasoning is likely to be treated as an important reference by developers, publishers and policymakers until the suit is decided on merits or the ruling is revisited on appeal.
This article sets out what the court actually held, distinguishes the findings from the surrounding commentary, and then places the ruling alongside the leading international decisions from the United States, the United Kingdom and Germany. A recurring theme is that outcomes in this field turn less on grand pronouncements about AI and more on the specific statutory framework of each jurisdiction and, critically, on the quality of the evidence a rights-holder brings.
The Dispute
ANI filed suit in the Delhi High Court in November 2024, the first action of its kind brought by an Indian media organisation against an AI company. Its case rested on two distinct allegations. The first was that OpenAI had used ANI's copyrighted news content, without a licence, to train the models underlying ChatGPT. The second was that ChatGPT sometimes produced responses that reproduced or drew upon ANI's work, and, separately, that the system attributed fabricated statements and events to ANI. ANI framed these false attributions, commonly described as hallucinations, as a distinct reputational harm capable of spreading misinformation.
ANI sought an urgent interim injunction restraining OpenAI from storing, reproducing or otherwise using its material for training or operating ChatGPT, and asked that the system's access to ANI's published content be disabled. At the first hearing OpenAI made a voluntary disclosure, without conceding wrongdoing, that it had already blocklisted ANI's website (aninews.in) in October 2024, excluding that domain from future training. That step addressed prospective crawling but not ANI's claims about past use, and it was recorded as a concession on the record.
Recognising the novelty of the questions, Justice Bansal framed four issues: (i) whether storing ANI's content to train ChatGPT amounts to infringement; (ii) whether generating responses using that content is itself infringing; (iii) whether any such use is protected as fair dealing under Section 52 of the Copyright Act, 1957; and (iv) whether Indian courts have territorial jurisdiction given that OpenAI's servers are located in the United States. The court appointed two amici curiae (the intellectual property advocate Adarsh Ramanujan and Professor Arul George Scaria of NLSIU) to assist. Findings on the first and third issues were delivered together.
What the Court Held
Jurisdiction: ANI succeeded
It is worth correcting a common misreading of the result at the outset. OpenAI did not win on jurisdiction; it lost that point. OpenAI's threshold objection was that, because its models are not trained in India and its servers sit in the United States, Indian courts lacked territorial jurisdiction to hear the suit. Justice Bansal rejected that objection and held, in ANI's favour, that the Delhi High Court had jurisdiction to entertain the case. In pronouncing the order the judge stated in terms that on territorial jurisdiction he had held for the plaintiff. So while the overall result went decisively to OpenAI, ANI carried the one issue that keeps the underlying suit alive in India.
Storing content for training: prima facie fair dealing
On the core training question, the court took the prima facie view that OpenAI's act of storing ANI's original literary works for the purpose of training the LLMs underlying ChatGPT falls within the fair dealing exception in Section 52(1)(a) of the Copyright Act, and therefore does not amount to infringement under Section 51. The judge located the use within the statutory carve-out for private or personal use, including research, under Section 52(1)(a)(i). He also reasoned, at the balance-of-convenience stage, that India is regarded as a frontrunner in AI and that data functions as the fuel on which LLMs operate, and that an injunction restraining the use of such content would be detrimental to the growth of AI and of LLMs being developed in India.
Outputs: not substantially similar; no proven memorisation
On the output side, the court held that ChatGPT's responses were not substantially similar to ANI's original literary works, and so did not infringe under Section 51. It extended the same conclusion to responses generated using retrieval-augmented generation (RAG), an architecture in which a model fetches external material at query time rather than reproducing content stored in its weights. Crucially, the court found that ANI had failed to satisfy it that any memorisation or regurgitation of ANI's works had occurred in the responses generated by ChatGPT. This was an evidentiary failure as much as a legal one: the claim of verbatim reproduction was not made out on the record before the court.
Balance of convenience and public interest
Having found no prima facie case on either training or outputs, the court held that the balance of convenience also favoured OpenAI, and that irreparable injury would be caused not only to OpenAI but to the public at large if an injunction were granted at this stage. The judge observed that generative AI has transformed how people seek and obtain information, and that halting the service on an interim basis would harm the millions of users in India who rely on it. The interim application was accordingly dismissed. The false-attribution or hallucination grievance, which sounds more in the nature of a reputational or misinformation harm than classical copying, was not the basis on which relief was granted and remains a live matter for the suit.
Reading the Ruling Carefully
Three cautions follow from the text. First, this is an interim order expressed in prima facie terms. It refuses an injunction; it does not finally decide that training on copyrighted news is lawful in India. The suit continues, and the findings can be revisited at trial or on appeal. Second, the reasoning is anchored in the specific language of Section 52 of the Indian statute, in particular the research limb, rather than in an open-ended American-style fair use balancing test. That statutory grounding matters when comparing the decision with foreign rulings. Third, the output findings rest heavily on the absence of evidence of memorisation and substantial similarity. A future claimant that can actually demonstrate verbatim regurgitation would present the court with a materially different record.
The International Landscape
The Delhi ruling arrives after a cluster of decisions abroad, and it is broadly consonant with the emerging pattern: courts have tended to treat the act of training as capable of protection, while scrutinising outputs, the provenance of the data, and evidence of market harm. But the routes differ, and the differences are instructive.
United States: transformation, provenance and market harm
The United States analyses these disputes through the four-factor fair use test in Section 107 of its Copyright Act. In Thomson Reuters v. ROSS Intelligence (D. Del., February 2025), Judge Bibas rejected a fair use defence where ROSS had used Westlaw's proprietary headnotes to build a competing legal-research tool. The use was found non-transformative and directly harmful to the market for the original, and the tool was not generative AI but a substitute search product. ROSS illustrates that where an AI product competes head-on with the source, the defence is fragile.
The generative-AI cases came out differently. In Bartz v. Anthropic (N.D. Cal., June 2025), Judge Alsup held that training an LLM on lawfully acquired books was exceedingly transformative and thus fair use, likening the process to a person reading in order to learn to write. But he sharply distinguished the source of the data: Anthropic's downloading and retention of over seven million pirated books for a permanent central library was not fair use. That split (lawful training use protected, piracy not) carried enormous stakes; the piracy exposure was later resolved through a class settlement announced in 2025 reported at approximately US$1.5 billion. Two days later, in Kadrey v. Meta (N.D. Cal., June 2025), Judge Chhabria also found training to be fair use, but reasoned differently, holding that the authors had failed to produce meaningful evidence of market dilution. He pointedly signalled that stronger economic evidence in a future case could change the outcome.
Read together, the American cases send a consistent signal that maps onto the Delhi result: training may be defensible, but lawful acquisition of the data and proof of market harm are where cases are won and lost. The Indian court reached a compatible destination by a different statutory road, relying on the Section 52 research exception rather than transformativeness, and, like Kadrey, resolved the output question against the rights-holder largely on the evidence.
United Kingdom: territoriality and a narrow statutory grip
The most important non-US comparator is Getty Images (US) Inc & ors v. Stability AI Ltd [2025] EWHC 2863 (Ch), handed down on 4 November 2025, the first UK trial judgment on copyright and generative AI. Tellingly, by the close of evidence Getty had abandoned its primary copyright claims relating to training, in part because it could not establish that the training and development had taken place in the UK, and because of limited evidence of infringing output. The court therefore declined to rule on the central training question, leaving the point unresolved. On the surviving secondary-infringement claim, the court held that the Stable Diffusion model was not an infringing article for the purposes of the Copyright, Designs and Patents Act 1988, so importing or making it available in the UK did not constitute secondary infringement. Getty secured only extremely limited findings on its trade mark claims relating to watermarks in certain earlier model versions. Permission to appeal was granted in December 2025.
The Getty case underscores a point that resonates with the Delhi ruling: where servers, training and data acquisition sit outside the forum, territoriality becomes decisive, and evidence of what the model actually reproduces is often the weak link in a rights-holder's case. The difference is that in Delhi the jurisdictional objection failed and the court proceeded to address the merits, whereas in London the claimant's own retreat left the merits of training largely untouched.
Germany and the EU: a divergent path
Continental Europe is moving in a different direction. In November 2025 the Munich Regional Court ruled in favour of the German collecting society GEMA, holding that OpenAI had infringed copyright in song lyrics that were found to have been memorised by its models and reproduced in near-identical form in response to simple prompts. That reasoning, liability precisely because the model had memorised and could regurgitate protected text, is the mirror image of the Delhi finding, where the absence of proven memorisation was fatal to ANI. The EU framework, with its text-and-data-mining exception under the Digital Single Market Directive and an opt-out mechanism, gives European courts levers that neither the Indian nor the US statute contains, and a reference to the Court of Justice of the EU on the scope of that exception remains possible.
Finally, the most closely watched action, The New York Times v. OpenAI in the United States, remains in litigation and has not produced a final merits ruling on training. It is best treated as pending rather than as authority, and the Delhi court itself noted that no authoritative foreign ruling yet existed for it to draw upon.
Where the Delhi Ruling Fits
Placed against this backdrop, the Delhi High Court's order is neither an outlier nor a radical departure. It aligns with the prevailing tendency (visible in Bartz, Kadrey and, by omission, Getty) to treat model training as capable of legal protection while leaving outputs, data provenance and demonstrable harm as the real battlegrounds. Where India diverges is in its statutory mechanism: rather than the elastic US fair use standard or the EU's opt-out-based mining exception, the court leaned on the enumerated research exception in Section 52 of a 1957 statute drafted long before machine learning existed. Whether that provision can bear the weight now placed on it is precisely the kind of question an appellate court, or the legislature, may need to resolve.
Two features of the ruling deserve emphasis for anyone reading it as a precedent. It is interim and prima facie, so its holdings are provisional by design. And it turned substantially on evidence: ANI could not show memorisation, regurgitation or substantial similarity. As the GEMA decision in Munich demonstrates, a claimant who can prove that a model reproduces protected text may obtain a very different result even on similar facts. The lesson from the global picture is consistent: in AI-copyright litigation, the evidentiary record, the lawfulness of the data source, and the jurisdictional footprint frequently matter more than any sweeping statement about the technology itself.
Conclusion
ANI Media v. OpenAI gives India its first judicial marker in the AI-copyright debate, and it favours the developer at the interim stage: storing content for training is prima facie fair dealing, ChatGPT's outputs were not shown to be substantially similar to ANI's work, and no injunction will issue against a service the court regarded as serving a broad public interest. Yet the win is narrower than the headlines suggest. ANI prevailed on jurisdiction, keeping its suit firmly within the reach of the Delhi High Court; the false-attribution grievance survives; and the entire order is provisional. Read alongside ROSS, Bartz, Kadrey, Getty and the Munich GEMA decision, the ruling confirms that a rough international consensus is forming around the training question, even as the sharper contests over outputs, licensing and data provenance are only beginning.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
