91视频

Delhi High Court’s first substantive ruling on AI training and copyright did not restrain ChatGPT. But the reasoning, not the result, is what in-house teams need to read closely, writes IP expert Modhura Roy

Justice Amit Bansal delivered a 135-page interim order, in ANI Media Pvt Ltd v OpenAI Inc & Anr, at Delhi High Court on 24 July 2026, declining to injunct OpenAI from storing and using the ANI news agency’s content to train large language models behind ChatGPT.

This was the first substantive judicial engagement in India, with the question being litigated in parallel across the US, UK and EU: is training a generative model on someone else’s copyrighted material an infringement, or is it permitted?

For general counsel skimming the headlines, the takeaway looked simple: OpenAI won, AI training is fine in India, move on. That reading is wrong, and acting on it would be a mistake. The order is interim, not final; it turns as much on a failure of evidence as on a point of law, and on one issue that matters enormously for cross-border enforcement: ANI actually won.

The value of the judgment for corporate IP teams lies in its reasoning and what it signals about how Indian courts will handle these disputes for years to come.

Inside the ruling

ANI filed the suit in November 2024, seeking an injunction and about INR20 million (USD210,000) in damages. Its case had two limbs: that OpenAI unlawfully stored and used ANI’s news reports to train ChatGPT without a licence; and that the chatbot generated fabricated statements falsely attributed to ANI – so-called hallucinations – that the agency argued threatened its reputation.

Due to the wide implications, the court heard from a range of intervenors representing broader industry interests including publishers, digital news outlets, music producers, AI companies and technology policy groups.

The court framed and answered four questions, each at the prima facie stage.

1. 1. On territorial jurisdiction. OpenAI argued that its models are trained on servers outside India, so Indian courts could not hear the claim. The court rejected that objection and held it had jurisdiction, reasoning that ANI is based in Delhi and that ChatGPT is accessed and used in India. The judge was unwilling to accept a logic that would let any developer escape Indian law simply by hosting its infrastructure abroad. On this issue ANI succeeded.
   2. On storage of copyrighted works for training. The court took the prima facie view that OpenAI’s storage of ANI’s material to train its models falls within the fair dealing exception in section 52(1)(a) of the Copyright Act, 1957, and therefore does not amount to infringement under section 51.
   3. On outputs. The court found that the responses ChatGPT generated were not substantially similar to ANI’s original reporting, and that ANI had not shown the model had memorised or regurgitated its works. The illustrations ANI relied on appeared to be produced through retrieval at query time rather than reproduced from the training corpus, which undercut the memorisation theory.
   4. On balance of convenience and public interest. The court concluded that restraining a widely used AI system at the interim stage, before the merits had been tested, could harm innovation and the wider public interest. Having found no prima facie case of infringement, the court refused the injunction.

But crucially, it did not decide the ultimate legality of AI training, so the suit proceeds to trial, and ANI’s hallucination and misattribution claim remains alive.

Not just a publisher’s problem

It is tempting to file ANI v OpenAI under “publisher problems” and assume it is irrelevant to a manufacturing, technology or financial services legal department. That assumption would be myopic. Almost every large enterprise is now simultaneously a content owner (it produces manuals, research, code, marketing, databases and proprietary reports), a content licensor or licensee, and a deployer of generative AI tools.

The principles the court articulated on jurisdiction, evidence, scope of 91视频 copyright exceptions and the reputational fallout of AI misattribution reach directly into all three roles.

The following implications are the ones worth carrying into your next strategy meeting.

Implication 1: A double-edged jurisdiction finding. The jurisdiction finding may prove the most durable part of the order. By holding that effects felt in India are enough to bring a foreign platform before an Indian court, the judgment gives Indian rights holders a forum to pursue global AI developers without being turned away at the threshold. For a domestic company whose proprietary content has been scraped, that is genuinely useful; claimants are not obliged to litigate in California to protect an Indian copyright.

The same reasoning, however, exposes companies that build or fine-tune AI models. If your organisation trains models on Indian data, or deploys AI products accessible to Indian users, the location of your servers will not shield you from litigation in India. In-house teams should map where their AI training and deployment touch Indian users and content, and assume Indian courts are an available venue for anyone who objects.

Implication 2: No injunction without evidence. The heart of ANI’s defeat was not a doctrine, it was proof. The court was not persuaded that ChatGPT had memorised or reproduced ANI’s articles, and it distinguished between material retrieved and synthesised at the time of a query and material reproduced from the training set.

For any IP team contemplating enforcement against an AI developer, this sets the bar. General assertions that “the model must have trained on our content” will not move an Indian court. What is needed is concrete technical evidence: documented instances of verbatim or near-verbatim output; controlled and reproducible prompting that surfaces protected expression; expert analysis of memorisation; and a clear demonstration of substantial similarity between output and the original work.

This has practical consequences before any dispute arises. Companies that regard their written and structured content as valuable IP should build the capability to detect and preserve evidence of infringing outputs, i.e., logging queries and responses, retaining originals with provenance, and being able to show substantial similarity. Enforcement strategy against AI now looks less like a cease-and-desist exercise and more like a technical investigation.

Implication 3: A narrow exception, generously read. It is easy to read the section 52 finding as India adopting a broad, US-style fair-use position for AI. It has not. Indian copyright law does not contain an open-ended fair-use doctrine. Section 52 is a closed, enumerated list of permitted acts, and the courts have generally treated it as exhaustive.

The judge’s willingness to place training-related storage within the fair dealing exception is significant precisely because it stretches a narrow statutory category to a use its drafters never contemplated, that is, a prima facie view at the interim stage. It could be narrowed at trial, distinguished on different facts, or revisited on appeal.

The lesson for in-house counsel is caution. Do not treat this order as a green light to train models on third-party content in India without regard to permissions. The legal position seems unsettled, the exception is being read expansively for now, and the final word has not been spoken. Where your organisation relies on ingesting external data to build AI capability, the conservative course by using licences, permissions or clearly non-expressive use still remains the defensible one.

Implication 4: Licensing, opt-outs and data governance. A quiet but important fact in the case is that OpenAI had, from October 2024, placed ANI’s domain on a blocklist and undertook to exclude it from future training. The dispute, in other words, was largely about historic use. This points to the direction of travel: opt-out mechanisms, robot-style exclusion signals and negotiated licences are becoming the practical infrastructure of the AI-content relationship, filling the space that litigation cannot.

For corporate IP teams this suggests a two-sided agenda. As a content owner, decide deliberately whether your material should be available for AI training and, if not, implement and document technical opt-outs because a court may later ask whether you took reasonable steps to signal your position.

As a content user or licensor, revisit your contracts. Licences, vendor agreements, employment and contributor terms, and customer contracts increasingly need express provisions on whether data may be used to train AI, who owns AI output, and how liability for infringing training or output is allocated.

Indemnities and warranties around AI training rights are fast becoming standard, and the department that drafts them early will avoid disputes later.

Implication 5: When the model gets it wrong. The part of ANI’s claim that ChatGPT attributed fabricated statements to it survives. This is a reminder that harm from generative AI is not confined to copyright. False attribution engages reputation, potential passing-off and, depending on the content, defamation and consumer protection concerns.

Corporate IP groups tend to think in terms of patents, trademarks and copyright, but AI misattribution sits at the intersection of brand protection and IP, and the court’s decision to keep this claim alive signals that Indian courts take it seriously.

The practical response is monitoring. Brand and IP teams should periodically test how major AI systems represent their company, products and executives, document instances of false attribution, and have an escalation path – to the platform, and if necessary to court – when hallucinated content causes real harm. This is brand watch adapted for the AI era, and it belongs on the IP function’s radar.

The policy horizon

The judgment does not exist in a vacuum. The scope of copyright exceptions for AI training is already under examination by an expert committee constituted under the Ministry of Commerce and Industry. The outcome of that process, which would potentially be a purpose-built text and data mining exception or licensing frameworks, may ultimately matter more than any single interim order.

India is closely watching the EU’s text and data mining regime, and parallel ongoing litigation in the US and UK. In-house teams should treat the current legal position as provisional and keep a watching brief on both the trial in ANI v OpenAI and the policy track, because the ground will shift.

Turning the ruling into action

Translating the judgment into action, several steps follow.

1. 1. Audit your content assets and classify what is genuinely valuable IP versus commodity information, since only the former justifies enforcement investment.
   2. Decide, as a matter of policy, whether your content should be available for AI training, and implement documented opt-outs where it should not.
   3. Build the technical capability to detect, reproduce and preserve evidence of infringing AI output before you need it.
   4. Refresh all forms of contract templates, including (but not limited to) vendor, customer, employment and licensing to address AI training rights, output ownership, indemnities and warranties.
   5. Govern your own internal use of generative AI to prevent leakage of your trade secrets and proprietary material into third-party models.
   6. Extend brand monitoring to cover AI misattribution, with a clear escalation route.

Key takeaways

ANI v OpenAI is best understood not as a verdict on the legality of AI training, but as an early map of the terrain.

It tells corporate IP teams that Indian courts will assert jurisdiction over global platforms; that enforcement will rise or fall on technical evidence; that 91视频 narrow statutory exceptions are being read generously but unpredictably; that licensing and opt-outs are the emerging norm; and that AI misattribution is a live reputational risk.

None of this is settled law. However, the direction is clear enough – corporate IP functions that start adapting their evidence, contracts and governance now, instead of waiting for the final judgment, will be the ones best placed when it arrives.

Modhura Roy is director – legal and corporate administration (intellectual property & open-source compliance) at Cognizant