Delhi court tells ANI: the price of training data is real, the per-source market is not

The Delhi High Court did not tell OpenAI it could use ANI's stories for free. Justice Amit Bansal told ANI on Friday that the price of those stories is real, the market for them is not, and the only remedy on offer is money, not a stop sign.

The order in ANI Media Pvt Ltd v. OpenAI Opco LLC (Verdictum case 2026:DHC:5900) draws that line, and it is the line a single-sentence "OpenAI wins" headline will skip. The training question is not closed. The injunction question is.

Justice Bansal refused.

The first was jurisdiction. OpenAI had argued that training happens on servers outside India, so a Delhi court could not hear the case at all. The court disagreed: ChatGPT is offered in India, the alleged output infringement happened here, and the suit proceeds in Delhi (The Hindu, 24 July 2026).

The second was the heart of the case: whether OpenAI's training of ChatGPT on ANI's literary works was copyright infringement at all. Under Indian copyright law, Section 51 of the Copyright Act, 1957, sets out what counts as infringement; Section 52(1)(a) carves out a "fair dealing" exception for private or personal use, research, criticism, or review. Justice Bansal held that OpenAI's storage of ANI's works for the purpose of training its LLMs is, on the face of it, covered by that exception. That is a prima facie holding, not a final judgment, but it is the first time an Indian court has applied fair-dealing to large-scale LLM training (MediaNama).

The output side went the same way. ANI showed the court specific ChatGPT responses that reproduced ANI bylines and text. Justice Bansal found those outputs were not "substantially similar" to ANI's works under Section 51. The instances ANI cited were published after OpenAI's training cut-offs of April 2022 (for GPT-4) and April 2024 (for GPT-4o), which means the model could not have memorised them during training. The likelier source, the court reasoned, was retrieval: ChatGPT searches the live web, including ANI's own site, when it answers a question, a technique known in the field as retrieval-augmented generation. ANI failed to show that what it saw was training memorisation rather than retrieval (The Leaflet).

Even where a plaintiff shows a prima facie case, an Indian court will refuse interim relief if the harm to the defendant, and to the public, outweighs the harm to the plaintiff. Justice Bansal found that an injunction "would be detrimental to the growth of AI and more particularly, to the LLMs being developed in India" and that "it would be economically unviable to develop an LLM if training of an LLM would require licenses from multiple sources."

The court then pointed to a concrete number: ANI itself offered OpenAI a license for USD 7.5 million in a communication dated 2 October 2024. OpenAI declined. ANI had no record evidence of lost subscribers or syndication revenue. The court called ANI's claimed injury "quantifiable" and capable of being compensated in money, which is the standard test for refusing an interim injunction in favour of damages. The order does not say training is free. It says the per-source licensing model is not workable at the scale of an LLM.

ANI is seeking Rs 2 crore in damages (roughly USD 240,000 at mid-2026 rates against the dollar) in the main suit, which continues. The reputation and fabrication claim is still live: ANI alleged that ChatGPT had fabricated ANI-bylined content and damaged its standing. Justice Bansal did not adjudicate that on the merits this round; the only memorisation question he decided was the one ANI put before him, and ANI lost it. The fabrication theory returns to the docket.

Five intervenors lined up behind ANI at the hearing: the Digital News Publishers Association, the Federation of Indian Publishers, T-Series, Saregama, and Sony Music. Two amici curiae also appeared. The news and music industries are using this case to set a template. ANI's own counsel cited OpenAI's existing licensing deals with the Financial Times, the Associated Press, and Condé Nast as evidence that comparable licensing is industry practice. Justice Bansal's holding accepts that such deals exist while refusing to require them.

The case now runs on three tracks. The next hearing will move to damages and to the reputation and fabrication theory. The training question can still be reopened on final judgment, on appeal, or in a parallel suit. The marker the Delhi court has planted, that per-source licensing of training data is economically unworkable, will be priced in by every Indian news organisation, AI lab, and downstream licensing negotiation that follows.