Delhi High Court Denies ANI’s Injunction Against OpenAI in India’s First LLM Copyright Case
India's first court ruling on AI training and copyright denied ANI Media's injunction against OpenAI on July 24, 2026. Here's what the Delhi HC actually decided.
On July 24, 2026, Justice Amit Bansal of the Delhi High Court denied ANI Media's application for an interim injunction against OpenAI, making it the first Indian court decision to consider whether training a large language model on copyrighted content infringes copyright. ANI, one of India's largest news agencies, argued that OpenAI copied its articles without authorisation and that ChatGPT reproduced or misattributed its reporting. OpenAI won on both points, but the ruling leaves three significant legal questions unresolved, and the full trial has not yet taken place.
What happened
| Detail | Fact |
|---|---|
| Case name | ANI Media Pvt. Ltd. v. OpenAI OpCo LLC |
| Court | Delhi High Court |
| Judge | Justice Amit Bansal |
| Decision date | July 24, 2026 |
| Outcome | Interim injunction denied; OpenAI wins on both claims |
| Models named in suit | GPT-4 and GPT-4o |
| Indian statute at issue | Copyright Act, 1957, Sections 14(a)(i) and 52(1)(a) |
ANI raised two claims. The first was an input claim: that OpenAI copied and stored ANI’s articles to train ChatGPT without permission, which ANI argued fell under Section 14(a)(i) of the Copyright Act, 1957 (the provision covering reproduction of literary works). The second was an output claim: that ChatGPT reproduced or misattributed ANI’s journalism in its answers.
OpenAI countered with the fair dealing exception for research under Section 52(1)(a), arguing that training an LLM is a transformative process rather than simple reproduction. The court agreed on both counts.
Why the output claim collapsed: retrieval, not memory
The most decisive finding was also the most technical. ANI submitted examples of ChatGPT producing answers that closely matched its articles. The problem: GPT-4 and GPT-4o were trained before those articles were published. The court was direct in paragraph 124 of the order: “the illustrations given in the plaint are post the training of Open AI’s LLMs and a case for memorization… cannot be made out.”
What actually happened was almost certainly RAG, retrieval-augmented generation. This is a technique where the model pulls live content from the web while answering a query rather than recalling something baked into its weights during training. The court explained this distinction carefully, and it proved near-dispositive on the output claim.
The court did flag one question it chose not to answer: whether serving live copyrighted content through RAG amounts to “communication to the public” under Section 2(ff) of the Copyright Act. That right covers any work made available for the public to see, hear, or enjoy, and it does not require proof that the work was stored or copied during training. ANI had not properly pleaded this argument, so the court left it open. It is likely where future disputes will focus.
Why the input claim is the weaker win for OpenAI
The court accepted that LLM training qualifies as “research” under Section 52(1)(a), bringing it within the fair dealing exception. The analysis has two pressure points worth noting.
First, India has no statutory text-and-data mining exception (a specific carve-out for automated, large-scale analysis of corpora, which several other jurisdictions have introduced). The court stretched the existing “research” provision to cover AI training, doing by interpretation what Parliament has not done by legislation. Higher courts may not read the provision that broadly.
Second, the court conflated training and output when addressing jurisdiction, assuming infringement occurred wherever the model was trained. OpenAI argued the Copyright Act has no extraterritorial reach because its models were trained in the United States. The court disagreed, holding that training and the resulting outputs formed one continuous process, and since ANI is based in India and outputs were accessed here, Indian law applied. That reasoning extends an Indian court’s reach to conduct that occurred largely abroad, provided some part of the chain touches Indian territory.
Three questions the trial court left open
This is an interim order decided on a lower bar than a full trial: prima facie case, irreparable harm, and balance of convenience. ANI can still present evidence at trial. Three unresolved issues could shift the outcome:
- Does “research” cover machine-scale processing? Section 52(1)(a) was written for human researchers. An LLM does not read; it processes entire corpora at a scale no person could. Whether that distinction matters legally is unresolved.
- Is live RAG retrieval “communication to the public”? This argument bypasses the training question entirely and asks whether an AI serving live copyrighted content without a licence is an unlicensed intermediary. ANI did not press it; a future claimant likely will.
- What does a full evidentiary record show? The court’s comments on how GPT-4 is trained were made without a complete trial record. The interim order can be cited but is not a final position on any of these points.
Why it matters
This is India’s first case in what is becoming a global body of law on AI training and copyright, joining similar disputes in the United States and the United Kingdom. For AI developers, the ruling provides short-term comfort: an Indian court has accepted the “research” fair dealing argument and demonstrated a solid grasp of how RAG differs from memorisation. For publishers and content owners, the communication-to-the-public question the court left open may offer a more durable route to a future injunction, since it does not depend on proving anything about training data at all.
For businesses that use AI tools drawing on live web content, the practical implication is real. If a future court rules that RAG-based retrieval of copyrighted material is “communication to the public,” the licensing landscape for AI products could change substantially, affecting everything from AI-generated news summaries to customer-facing chatbots. Our coverage of how AI safety guardrails interact with model behaviour shows how quickly legal and technical assumptions about AI can be upended.
Our take
The training-date mismatch is a genuinely strong finding. Courts anywhere should take that kind of concrete, temporal evidence seriously, and the RAG explanation is careful. The fair dealing stretch is less convincing. Reading “research” to cover industrial-scale machine processing is a policy choice dressed as statutory interpretation, and it sets a precedent that could unravel if a higher court applies stricter textualism.
The communication-to-the-public question is the one to watch. It is structurally harder for AI companies to defend because it does not care what happened at training time. If your product retrieves and serves a news agency’s content in real time, the argument that you are not reproducing it becomes very thin. Any business building AI products that surface third-party content, whether through AI integration or custom tooling, should treat this case as an early warning rather than a green light.
Frequently asked questions
Did the Delhi High Court rule that AI training on copyrighted data is legal in India?
Not definitively. The court denied ANI's interim injunction and accepted OpenAI's fair dealing defence for research under Section 52(1)(a), but this was an interim order, not a final verdict. The full trial can still produce a different result.
Why did ANI's ChatGPT output examples fail in court?
The articles ANI cited as evidence were published after GPT-4 and GPT-4o's training cutoff dates. The court concluded the model was retrieving live content via RAG rather than reproducing memorised training data, so the output claim could not be sustained.
What is retrieval-augmented generation (RAG) and why does it matter legally?
RAG is a technique where an AI model fetches current content from the web in real time to answer a query, rather than relying solely on what it learned during training. Legally, it means the model may not have 'stored' copyrighted material during training, which undermines infringement claims based on copying. However, it raises a separate question about whether serving live copyrighted content constitutes 'communication to the public'.
Can ANI still win the case against OpenAI?
Yes. The July 24, 2026 order only denied an interim injunction; the full trial on the merits has not taken place. ANI can present new evidence, and the court explicitly left open the communication-to-the-public argument under Section 2(ff) of India's Copyright Act.
