Training Data on Trial: What ANI Media v. OpenAI Signals for Businesses Using and Building AI in India

By Lucy Rana and Pranit Biswas

What the Court actually decided

In ANI Media Pvt. Ltd. v. Open AI OpCo LLC, the news agency alleged that its copyrighted content had been used without authorisation to train a large language model. On 24 July 2026, Justice Amit Bansal declined to grant an interim injunction, holding on a prima facie, interim-stage view that the storage and use of the material for training could fall within the fair-dealing exception under Section 52(1)(a) of the Copyright Act, 1957. The Court noted, among other things, that the agency had not demonstrated lost subscribers or revenue, and that it retained the ability to block the model’s crawlers, which weighed against the need for a restraint.

Two framing points give the decision its significance. First, it is one of the first Indian rulings to examine the training pipeline itself, not merely what a chatbot outputs to a user. Second, it is careful about its own limits: an interim order refusing temporary relief, decided before the evidence stage, which the Court itself treated as non-precedential and which remains open to appeal before a Division Bench. The main suit continues, and the central question — whether training on copyrighted works is lawful — is left for trial.

The decision is a signal, not a settlement. It tells us how one court weighed the interim balance; it does not resolve, for all cases, whether training on protected content infringes.

Input versus output — and why the distinction matters commercially

The reasoning turned in part on separating the ingestion of material for training from the generation of responses. That distinction matters to businesses on both sides of the data question. For content owners — publishers, news agencies, music and image libraries — it frames the practical question of whether value is better protected through technical measures such as crawler controls and access terms, and through licensing, than through litigation over ingestion alone. For companies building or deploying AI, it highlights that the analysis may treat what a model memorises and reproduces differently from what it merely learns from, and that demonstrable, near-identical reproduction of protected expression sits on more contested ground than abstract learning.

The international picture reinforces the point that outcomes are fact- and forum-specific. In late 2025 a German court found infringement where a model was shown to have memorised and reproduced protected song lyrics almost verbatim — close to the mirror image of the Indian court’s emphasis on the absence of proven memorisation. Closely watched proceedings continue in the United States. The lesson is less that any one result will prevail globally than that the treatment of training data is being decided case by case, on the specific evidence of what a model ingested and what it can be shown to reproduce. For instance, the United States settlement in Bartz v. Anthropic PBC illustrates the first of those limbs. In June 2025 the Northern District of California held that training on lawfully acquired books was transformative fair use, but that Anthropic’s downloading and retention of a permanent library from pirate sources was not. It was that finding on provenance, and the statutory damages exposure it created across some 482,460 registered works, which produced the USD 1.5 billion settlement approved in July 2026 — the largest in the history of United States copyright litigation. Thus, litigation on generative AI vis-à-vis copyright may differ based on the specific facts and circumstances.

The policy track running alongside the litigation

While the courts work through individual disputes, the policy question is being addressed separately. The Department for Promotion of Industry and Internal Trade published, in December 2025, the first part of a working paper on generative AI and copyright, canvassing options that include a centralised or compulsory-licensing approach to the use of copyrighted works for training, with a public consultation. Whatever emerges, the direction of travel is towards an eventual framework for the input side of AI — which means that today’s judicial answers may in time be overtaken, or codified, by legislative or regulatory ones.

For a fuller case note on the ruling, see the firm’s earlier analysis, “Copy, Right? ANI Media Pvt. Ltd. v. Open AI OpCo LLC.”

Frequently Asked Questions