What the court actually decided
On July 24, 2026, Justice Amit Bansal of the Delhi High Court refused Asian News International's bid for an interim injunction against OpenAI, holding — on a prima facie basis, in ANI Media (P) Ltd. v. Open AI OpCo LLC, CS(COMM) 1028/2024 — that storing ANI's news content to train ChatGPT falls within the fair dealing exception under Section 52(1)(a) of the Copyright Act, 1957, and therefore does not infringe under Section 51. The court also rejected OpenAI's threshold objection to Indian jurisdiction, holding that because OpenAI actively offers ChatGPT to users in Delhi and the disputed outputs are generated there, Section 62(2) of the Copyright Act and Section 20 of the Code of Civil Procedure both supply territorial jurisdiction — regardless of where training itself physically occurred. Crucially, this is an interim order on an injunction application, not a final verdict: the underlying suit, filed by ANI as a 287-page complaint on November 18, 2024, continues toward trial, and claims concerning ChatGPT's outputs — including allegations that it fabricated a nonexistent interview with Congress leader Rahul Gandhi and attributed invented quotes to ANI — remain live.
Steelmanning ANI's case
ANI's underlying grievance deserves to be stated plainly, not waved away. News-gathering is expensive: reporters, stringers, verification desks, and legal review all cost money that AI labs did not pay for. If a chatbot can answer a question using the substance of a wire report without linking to it, paying for it, or crediting it, the commercial incentive to fund original reporting erodes — a point the Digital News Publishers Association, representing outlets including The Times of India, Hindustan Times and NDTV, pressed hard when it intervened in the suit, warning the court that unchecked scraping and reproduction "reduces the incentive to create" journalism. The false-attribution claim is a separate and arguably stronger harm: an AI system inventing quotes and pinning them on a named news agency is a reputational injury that has nothing to do with fair use of a training corpus — it's closer to defamation-adjacent misrepresentation, and the court explicitly left it, along with broader output-reproduction claims, open for trial rather than closing the door on it.
Why the fair-dealing finding is still the right call
Separating training from output is the correct legal move, and the court's interim reasoning tracks it. Ingesting a news article to learn statistical patterns of language is not the same act, legally or functionally, as a chatbot reproducing that article's substance verbatim to a user. India's Copyright Act, unlike the EU's more codified text-and-data-mining carve-outs, relies on an open illustrative fair-dealing list — "private or personal use, including research" — and Bansal J.'s order extends that language to machine learning as a form of research, a reading consistent with how courts in the US have approached training under fair use (the Authors Guild v. Google precedent on transformative indexing is the closest analogue, though it predates generative AI). That is a defensible, incremental extension of existing doctrine rather than an invention of new law, and it avoids the trap of treating every act of reading copyrighted material — by a human or a machine — as inherently infringing merely because the reader is commercial. A rule that required licensing every news article ever indexed for model training would not meaningfully compensate journalism; it would functionally require AI developers to negotiate with an entire industry before writing a line of code, a barrier that legacy incumbents with the deepest catalogs — not scrappy Indian AI startups — would be best positioned to survive.
What actually matters going forward
The part of this ruling likely to get lost in "OpenAI wins" headlines is that ANI didn't actually lose on the claims that matter most for accountability. The court found, at the interim stage, that ANI's cited examples of ChatGPT outputs did not show the kind of verbatim memorization that would make the training claim easy to prove — but it did not dismiss the output and attribution claims themselves, which proceed to trial alongside the core training-infringement question. That is the right sequencing: courts should not enjoin a widely used product on a preliminary record before establishing what, precisely, it is doing wrong. But it also means OpenAI is not out of the woods. If ANI can show at trial that ChatGPT reliably reproduces its copyrighted expression, or that hallucinated ANI-attributed content causes demonstrable reputational harm, fair dealing's "fairness" prong — which weighs market substitution and does not shield every commercial use — could cut the other way on a fuller record.
The regional stakes
This is one of the first substantive rulings by an Indian court on AI training and copyright, arriving as similar disputes work through courts from New York to Toronto with no global consensus yet. For a market where OpenAI, Google and homegrown labs are racing to build India-specific models, a clear jurisdictional rule — Indian courts can hear these disputes even against US-based platforms — paired with room to develop under fair dealing, is a workable middle path. It neither hands publishers of undifferentiated wire copy a veto over AI development nor gives AI companies a blank check to launder plagiarized outputs through a training defense. The trial, not this order, is where that balance gets tested for real.