Can an AI company scrape a newspaper’s archive, feed it into a language model, and call it fair dealing? That question is sitting in front of the Delhi High Court right now, and nobody in India has a settled answer yet. This piece breaks down where the law stands, what the biggest pending case says, and what a licensing route might look like if fair dealing does not survive judicial scrutiny.
The Problem, Stated Plainly
Large language models are trained on enormous datasets scraped from the open internet. That dataset almost always includes copyrighted material: news reports, books, blog posts, photographs with captions, song lyrics, academic papers. The model does not “store” this content the way a photocopier does, but it does ingest it, break it into tokens, and adjust billions of parameters based on patterns found in that text. The output is new text, but the process depends entirely on copying protected expression at the ingestion stage.
Under Section 14 of the Copyright Act, 1957, the owner of a literary or artistic work has the exclusive right to reproduce the work in any material form, including storage in any medium by electronic means. Scraping and storing text for training purposes fits this definition on a plain reading. The real fight is not over whether copying happens. It is over whether that copying falls within an exception.
Section 52 and the Fair Dealing Question
Section 52(1)(a) of the Copyright Act permits fair dealing with a literary, dramatic, musical or artistic work for purposes such as private use, research, criticism or review. Indian law uses “fair dealing,” a narrower and more purpose-bound concept than the American “fair use” doctrine under Section 107 of the US Copyright Act. American fair use runs on a four-factor balancing test that expressly considers commercial character and market effect. Indian fair dealing is enumerated. The statute lists specific purposes, and AI model training for a commercial product is not one of them.
This distinction matters enormously for LLM developers. A company can plausibly argue that training on text for research purposes is fair dealing. It becomes much harder to argue that once the same training run powers a paid subscription product used by millions of people. Indian law currently does not carve out any explicit exception for AI model training as a category of copyright use. That gap is precisely why courts are being asked to decide the issue case by case, and why the outcome is so consequential for the entire AI industry operating in India.
ANI Media v. OpenAI: The Case to Watch
The clearest test of this question is playing out in ANI Media Pvt. Ltd. v. OpenAI Inc. & Anr., CS(COMM) 1028/2024, before Justice Amit Bansal at the Delhi High Court. Asian News International, a major domestic news agency, sued OpenAI claiming that ChatGPT was trained on ANI’s news content without permission or licence fees, and that the model occasionally reproduces ANI’s reporting, including material meant only for subscribers. ANI has argued that OpenAI’s use falls outside fair dealing under Section 52 precisely because ChatGPT is a commercial product built for profit, not a research tool. ANI has also raised the problem of hallucinated attribution, where the model invents quotes or stories and falsely credits them to ANI, which the agency says damages its reputation independently of the copying issue.
OpenAI’s defence runs on two tracks. First, a jurisdictional objection: the company argues that neither it nor its servers are based in India, so Indian courts have no authority over the dispute. Second, a merits defence built around how language models actually function. Counsel for OpenAI told the court that these systems generate answers through predictive pattern matching rather than by retrieving stored facts, and that identical prompts can yield different answers depending on when they are asked, which undercuts the claim that the model is simply reproducing a fixed copy of ANI’s articles.
The court appointed two amici curiae, Dr. Arul George Scaria of NLSIU and advocate Adarsh Ramanujan, to assist on the technical and doctrinal questions involved. Final arguments concluded on 27 March 2026 and judgment stands reserved as of this writing. The matter has already run through 32 hearings since ANI filed suit on 19 November 2024, making it the longest-running and most closely watched AI copyright dispute in the country. Whatever Justice Bansal decides on the interim relief application will function as India’s first real judicial signal on how Section 52 applies to model training, even though it will not be a final verdict on infringement itself.
Why Authors Guild v. Google Does Not Settle This
Litigants and commentators keep reaching for the American precedent in Authors Guild v. Google, where mass digitisation of books for a searchable index was held to be transformative fair use. The comparison is tempting but weak. Google’s book-scanning project turned physical books into a searchable database, but a search index does not compete with the books themselves in the market; a chatbot that can summarise, paraphrase, or in some cases reproduce a news article competes directly with the journalism it was trained on. A search tool points a user back to the source. A generative model can substitute for the source. That difference goes to the heart of market harm, which is where both the American four-factor test and the practical economics of Indian news publishing converge, even though Indian statutory language does not use the word “transformative” at all.
What a Licensing Model Would Actually Look Like
If courts eventually hold that large-scale, uncompensated scraping for commercial model training falls outside Section 52, the industry will need a lawful alternative. Global AI companies have already started building one, and the shape of it is fairly consistent:
- Content licensing agreements where a publisher grants an AI company the right to train on and cite its archive for a negotiated fee, often structured as a flat annual sum or a per-token royalty.
- Opt-out and opt-in crawler protocols, where publishers use robots.txt-style signals or newer standards to block or permit AI crawlers, giving rights holders a technical lever alongside the legal one.
- Attribution and traffic-back arrangements, where the AI product surfaces a citation or link to the original publisher, partially replicating the “search index” logic that worked in the Google Books case.
- Statutory licensing schemes, a route the Copyright Act already recognises in other contexts, which Parliament could extend to AI training through amendment rather than leaving the entire question to litigation.
For an Indian publisher or content owner negotiating one of these deals today, the practical checklist looks like this: confirm ownership and registration of the copyrighted works before negotiating, define training use separately from output use in the licence text, insist on audit rights over what was actually ingested, cap the licence to a defined dataset version rather than an open-ended crawl, and build in a renegotiation clause tied to the eventual outcome of pending litigation like ANI v. OpenAI.
Where This Leaves Practitioners
Nothing in Section 52 was written with transformer models in mind, and Parliament has not yet amended the Copyright Act to address AI training directly. The Commerce Ministry has responded by setting up a panel of experts to examine whether the existing statute can handle these questions or needs legislative change. Until that panel reports, or until the Delhi High Court’s reserved order comes down, Indian law on this point remains genuinely unsettled. Any advice given to a client on this issue today has to be framed as risk allocation rather than a confident prediction of how a court will rule. Lawyers advising AI companies operating in India should be pushing for licensing arrangements now, not waiting for a judgment that could go either way and that will, in any event, almost certainly be appealed regardless of outcome.
Frequently Asked Questions
Is it illegal to train an AI model on copyrighted content in India? There is no direct answer yet under Indian statute. The Copyright Act, 1957 does not mention AI training as a category, so the question turns on whether such use qualifies as fair dealing under Section 52(1)(a). That provision lists specific permitted purposes such as private use, research, criticism and review, and commercial AI training does not sit comfortably within any of them. Courts have not yet ruled on this squarely, which is exactly why ANI v. OpenAI matters so much.
What is the difference between fair use and fair dealing? Fair use, under American copyright law, applies a four-factor balancing test that looks at purpose, nature of the work, amount used, and market effect. Fair dealing, under Indian law, is a closed list of specific permitted purposes. A use either fits one of the listed purposes or it does not. India’s narrower, enumerated approach gives judges less room to declare a new commercial use “fair” simply because it seems transformative.
What is the ANI v. OpenAI case actually about? Asian News International sued OpenAI in the Delhi High Court, alleging that ChatGPT was trained on ANI’s news content without a licence and that it sometimes reproduces or misattributes ANI’s reporting. OpenAI has raised both a jurisdiction objection and a defence on how the technology functions. Final arguments concluded on 27 March 2026 and the order remains reserved.
Has the Delhi High Court decided the case yet? No. As of this publication, judgment on the interim relief application is still reserved. No final ruling on infringement has been delivered.
Can Indian news publishers stop AI companies from training on their content? Not through a blanket legal right at the moment, since there is no settled precedent either way. Publishers currently rely on a mix of technical blocking measures, such as robots.txt directives aimed at AI crawlers, and negotiated licensing deals. A ruling against fair dealing in ANI v. OpenAI would strengthen publishers’ hand considerably.
Will licensing become mandatory for AI companies training on Indian content? Not automatically, and not yet. Licensing is currently a voluntary commercial solution that AI companies are adopting in some markets to reduce legal risk. Whether it becomes something closer to mandatory will depend on how courts read Section 52, and on whether Parliament amends the Copyright Act to address AI training directly, which the Commerce Ministry’s expert panel is currently examining.
Does this issue affect only OpenAI, or all AI companies operating in India? It affects every AI company that trains or fine-tunes models using Indian copyrighted material, or that offers AI products to users in India. ANI v. OpenAI is the leading case because it is furthest along, but the underlying legal question of how Section 52 applies to model training is common to the entire industry.
As of publication, judgment in ANI v. OpenAI remains reserved. This piece will be updated the moment the Delhi High Court’s order is out.
