ANI vs OpenAI: Delhi HC Calls ChatGPT Training Fair Dealing

The Delhi High Court on July 24 refused to stop OpenAI from using ANI’s news content, holding that storing copyrighted articles to train ChatGPT prima facie qualifies as “fair dealing” under India’s Copyright Act. It is the first substantive ruling by an Indian court on whether AI training needs a licence — and it lands on the opposite side of the question from India’s own government policy paper. The order is interim, the main suit continues, and one important issue actually went ANI’s way.

What the court actually held

Justice Amit Bansal ruled on ANI’s application for an interim injunction in the news agency’s copyright suit against OpenAI, reserved since March 27 — the decision Bar and Bench reports as India’s first judicial finding on AI training and copyright. ANI, the country’s largest wire service, had asked the court to restrain OpenAI from using its content and to order its articles deleted from training datasets while the case runs.

The court answered the suit’s core questions in OpenAI’s favour at this stage. On training, it held that “OpenAI’s act of storing ANI’s original literary works for training the LLMs underlying ChatGPT falls under Section 52(1)(a)” of the Copyright Act, 1957 — the fair dealing exception — and so does not amount to infringement under Section 51, as both Bar and Bench and LiveLaw quote the order. An LLM, or large language model, is the statistical system behind chatbots like ChatGPT; “storing” here refers to ingesting text during the training process.

On outputs, the court found ChatGPT’s answers — generated through retrieval-augmented generation, where the model looks up current information rather than reciting memorised text — were not substantially similar to ANI’s articles. “ANI has failed to satisfy this Court that any memorisation or regurgitation of ANI’s copyrighted literary works has happened,” the order reads as quoted by SpicyIP and multiple legal outlets.

ANI did win something significant: jurisdiction. OpenAI had argued Indian courts could not hear the case because its servers sit in the US and no training happens in India. The court rejected that objection and held the suit maintainable, a finding reported consistently by Storyboard18, Medianama and Outlook Business. Every foreign AI company serving Indian users now knows an Indian courtroom is a live venue. One caveat for everything above: the detailed written order was still awaited when outlets published, so all quotes are as reported by the legal press present in court.

Why ANI lost the injunction

Indian courts grant interim injunctions on a three-part test: a prima facie case, balance of convenience, and irreparable injury. The court found ANI failed all three. “ANI has failed to make out a prima facie case… Balance of convenience is also against grant of interim injunction,” Justice Bansal wrote, per Bar and Bench — adding that an injunction would cause “irreparable injury” not only to OpenAI but “to the public at large.”

The economic reasoning went further than most observers expected. “It would be economically unviable to develop an LLM if training of an LLM would require licenses from multiple sources,” the judge observed per Bar and Bench’s report of the order, warning that an injunction “would be detrimental to the growth of AI and more particularly, to the LLMs being developed in India” and would hurt millions of Indian ChatGPT users, “many of whom would not be paid subscribers.”

That reasoning matters because of who ANI is and what it asked for. The agency filed the suit in November 2024 — India’s first generative-AI copyright case — claiming ₹2 crore in damages and alleging ChatGPT reproduced its work verbatim and even hallucinated stories falsely attributed to it, per TechCrunch’s report of the filing. The court heard two amici curiae, IP advocate Adarsh Ramanujan and Prof. Arul George Scaria, before ruling. The hallucination grievance was noted but did not carry the copyright injunction, LiveLaw reports.

How the ruling compares globally

India has just produced one of the more AI-friendly first rulings anywhere — through the unlikely vehicle of a closed statutory list written in 1957. The global scorecard as of July 2026:

The US comparison is the sharpest. In June 2025, Judge William Alsup held that training on copyrighted books is fair use — but that hoarding pirated copies is not, which is what pushed Anthropic into a settlement that received final approval on July 20, 2026: $1.5 billion, roughly $3,000 per covered work, per the Authors Guild. Days later, Judge Vince Chhabria sided with Meta while stressing his ruling was narrow — and that news content might present a stronger case against fair use. The one US decision to reject fair use outright, Thomson Reuters v. Ross, involved a direct competitor and is now before the Third Circuit, argued June 11, 2026.

Europe is running the opposite experiment. The UK’s High Court handed Stability AI a broad win over Getty in November 2025 mostly because the training happened outside the UK, and the government’s March 2026 report then abandoned a proposed opt-out exception, leaving licensing as the default. The EU allows commercial mining unless rights holders reserve their rights, an obligation the AI Act now polices — and Munich’s court still found OpenAI liable in November 2025 for song lyrics the model memorised and reproduced. Delhi’s reasoning — no memorisation shown, training protected — effectively splits the difference: India’s regulatory instinct on Big Tech has often followed Brussels, as we saw when the EU forced Google to open Android to rival AI assistants, but this ruling reads closer to the American cases than anything European.

The India angle

The ruling freezes an already lopsided bargaining table. No Indian publisher or news agency has a disclosed training-data licensing deal with OpenAI, while News Corp’s deal is reportedly worth over $250 million across five years and Dotdash Meredith’s is at least $16 million a year, per Press Gazette’s and LLM Pulse’s deal trackers. Indian publishers chose litigation instead: the 22-member Digital News Publishers Association intervened in ANI’s case, and its counsel told the court last August that “physical newspapers are disappearing, digital news will disappear, and only ChatGPT will remain,” per Business Standard. Even that coalition has cracks — the Times of India stayed out and instead joined Google’s paid AI Overviews pilot as one of nine publishers worldwide, per BestMediaInfo.

The court’s public-interest logic also reflects scale: India is OpenAI’s second-largest market with 100 million weekly ChatGPT users, Sam Altman wrote in February, per TechCrunch. An injunction on training would have reverberated far beyond one wire service — which is precisely why publishers wanted it, since what happens to content fed into chatbots is now a commercial question as much as a privacy one.

The quietest beneficiaries are India’s own model builders. A licensing mandate would have hit Sarvam AI — a $1.5 billion unicorn as of June per Medianama — along with Krutrim and other sovereign-model ventures training on Indian-language corpora, the same ecosystem competing with China’s fast-rising open models. None of them has publicly commented on the training-data question; the ruling shields their practices without their having argued for it.

Here is the twist we find most consequential for Policy & Society readers: the judiciary and the executive are now pulling in opposite directions. The government’s own DPIIT working paper of December 2025 proposes a mandatory licensing regime — AI firms could train without prior consent but would owe statutory royalties through a new body, the CRCAT, per Ikigai Law’s analysis. MeitY has backed that framework, per Medianama, rejecting the idea that public good grants “unconditional immunity against legal challenges.” Justice Bansal’s order, as Bar and Bench notes, expressly differs with the white paper’s premise. A court says fair dealing covers training; the government is drafting a regime that assumes it should be paid for. Parliament, not this lawsuit, will likely have the last word.

What to watch

The written order’s full text, first. Everything quoted so far comes from courtroom reporting; the detailed reasoning — especially how a closed 1957 exceptions list stretched to cover machine learning — will shape every argument that follows. Second, an appeal: ANI can take the interim order to a Division Bench of the Delhi High Court, though it had not announced a decision as of July 25, and neither ANI nor OpenAI had issued public statements. Third, the main suit itself proceeds to trial, where a fuller evidentiary record on memorisation could produce a different answer — the prima facie finding does not bind the final judgment, as Medianama notes. Fourth, the policy track: DPIIT’s Part II paper on AI outputs is pending, and this ruling raises the stakes for whether the government legislates its licensing model. And globally, watch whether OpenAI does in India what it did in Brazil this May — settle a publisher lawsuit by signing a licensing deal, per LLM Pulse’s tracker. A negotiated licence would make this landmark ruling moot between these parties while leaving its reasoning on the books for everyone else.