An Indian court said, for the first time, that training isn't infringement

On July 24, 2026, Justice Amit Bansal of the Delhi High Court issued a 135-page judgment with a straightforward bottom line: the interim injunction sought by news agency ANI (Asian News International) against OpenAI is refused.

One sentence carries the ruling. "OpenAI's act of storing ANI's works does not amount to copyright infringement." Storing is the operative word — the finding is about the act of collecting and processing news articles in order to train a large language model.

This is the first time an Indian court has substantively ruled on copyright in LLM training data. And it hands judges in the 20-plus similar suits running worldwide — The New York Times against OpenAI, plus actions from publishers, labels and image archives in a dozen jurisdictions — one more decision to reason against.

One caveat up front: this is not a final judgment. The main suit continues. The court itself stated that its observations are limited to the interim application and have no bearing on the final outcome. There's still a specific reason it matters, and we'll get to it.

The cast — ANI, OpenAI, and Section 52

ANI is India's largest news agency. Founded in 1971, it supplies video and text news across India and South Asia; most Indian broadcasters and newspapers run its feed, and Reuters holds a reported stake. For a wire service, content licensing is the revenue model — essentially the whole business is being paid when someone else uses your reporting. Which makes wire services the single most exposed category to the AI training data question.

ANI filed in the Delhi High Court in November 2024, alleging that OpenAI scraped its content without authorization to train ChatGPT and that ChatGPT also generated fabricated information falsely attributed to ANI. It sought damages and deletion of stored data. Indian publishers and music industry bodies later joined, expanding the case considerably.

OpenAI's defense is consistent across jurisdictions: training extracts statistical patterns rather than reproducing and distributing works, and absent reproduction in outputs there is no infringement. In India it added a jurisdictional argument — servers outside India, training conducted outside India.

The key that unlocked this ruling is Section 52 of India's Copyright Act, 1957. It lists uses exempted from infringement, and Section 52(1)(a) covers "fair dealing" for private or personal use, including research. The name resembles American fair use but the structure differs sharply. US fair use is an open, four-factor balancing test. Indian fair dealing is a closed list — the use must fall within an enumerated purpose. So the real question was whether LLM training could fit inside that list at all.

What the court actually decided

Issue The court's finding
Storing works for training Within Section 52(1)(a)(i) fair dealing — not prima facie infringement
Does ChatGPT reproduce ANI articles? ANI failed to establish reproduction or retrieval
Balance of convenience Did not favour ANI
Harm if injunction granted Irreparable harm possible to OpenAI and the wider public
Scope of the finding Limited to the interim application; no bearing on the final outcome

The first finding is the consequential one. The court held that OpenAI storing ANI's literary works and using them to train a large language model falls within Section 52(1)(a) fair dealing. India's statute exempts research and private use, and the court read model training as that kind of use. Getting that interpretation out of a closed-list regime is a notably forward-leaning move — and precisely why it will be contested.

The second finding is about evidence. ANI argued ChatGPT regurgitates its articles; the court found the supporting material insufficient. This is the practically important part. Across current global litigation, the line separating plaintiff wins from plaintiff losses generally runs right here: "they trained on it" is not enough; you have to show "it comes back out." That is exactly why the NYT complaint attached more than a hundred examples of ChatGPT reproducing its articles nearly verbatim.

The third finding is injunction-specific logic. An interim injunction restrains conduct before the merits are decided, so courts weigh the harm of granting against the harm of refusing. Here the court concluded that restricting ChatGPT in India could cause irreparable harm not just to OpenAI but to the many users and businesses already relying on it. Read that as judicial recognition that AI services have become something close to infrastructure.

And again, worth repeating: this is a prima facie determination at the interim stage. Stronger evidence at trial could produce a different result, and an appellate bench could land elsewhere. Section 52 has never been tested against foundation-model training before, so the upper courts have genuinely open ground.

What each side gets

OpenAI gets time and a citable finding. India is among ChatGPT's largest markets by user count; a service-restricting injunction would have hurt commercially and, worse, set a bad precedent for courts elsewhere. Instead OpenAI can now say a major jurisdiction's court treated training as fair dealing. Foreign courts aren't bound by it, but its weight as persuasive material is real.

ANI got less, but not nothing. The main suit survives. Refusing an injunction isn't a finding that ANI is wrong — it's a finding that the case isn't clear enough to justify immediate restraint. ANI can strengthen its output-reproduction evidence and fight the merits. That said, if the practical purpose of the suit was leverage in a licensing negotiation, that leverage just got noticeably weaker.

India's AI sector gets predictability. IT services are a large share of India's GDP, and legal uncertainty around training was a burden on domestic startups too. A signal that training is presumptively permitted helps model development inside India — relevant given the government's push to build sovereign foundation models.

News organizations worldwide lose negotiating power. Much of the leverage behind the licensing deals struck over the past two years came from a credible threat to sue. As the odds of winning fall, so do the terms. This ruling pushes in that direction.

AI companies broadly gain norm formation. As similar rulings accumulate across jurisdictions, a de facto rule forms without any explicit legislation. And legislation that arrives after a de facto rule has hardened tends to ratify practice rather than reverse it.

How other courts have ruled

This decision only makes sense next to the record built over the past two years — and that record splits.

Rulings favorable to AI. In June 2025, Judge William Alsup of the Northern District of California held in Bartz v. Anthropic that training an LLM on copyrighted works is "exceedingly transformative" and constitutes fair use. That same month Judge Vince Chhabria ruled for Meta in Kadrey v. Meta, finding plaintiffs had failed to demonstrate market harm. Both adopted the training-as-transformative-use logic.

Rulings unfavorable to AI exist just as clearly. In the same decision, Alsup treated Anthropic's sourcing of training material from pirate libraries as a separate and actionable wrong, which ultimately produced a $1.5 billion settlement in September 2025. Training may be fine; how you obtained the corpus is examined independently. And in February 2025 a Delaware federal court rejected fair use in Thomson Reuters v. Ross Intelligence, where the decisive fact was that the resulting product was a direct competitive substitute for the original work.

Three patterns recur across all of it. One, the act of training is trending toward protection as transformative use. Two, how the data was acquired creates independent liability. Three, whether outputs substitute for or reproduce the original is what most often decides the case.

The Indian ruling sits squarely on patterns one and three: training is fair dealing, reproduction wasn't proven. That a decision from an entirely different legal tradition mirrors the structure of American outcomes is the real significance here — different statutes in different systems converging on the same shape of answer.

Unresolved cases still matter. NYT v. OpenAI and Microsoft is ongoing with a large volume of output-reproduction evidence on the record, so it may well land differently. In Europe, fights over the AI Act's text-and-data-mining exception and opt-out mechanics run on their own track. No single ruling closes this.

How publishers counter

Shift to output evidence. The lesson here is unambiguous: asserting "you trained on our work" isn't enough; you must show "this model emits our work." Publishers preparing suits will move resources from crawl logs toward systematic prompt experiments.

Block by contract and by technology. As case law hardens unfavorably, extra-judicial tools matter more — robots.txt and crawler blocking, CDN-level bot filtering, harder paywalls. Cloudflare defaulting to blocking AI crawlers is part of this shift. It does nothing about models already trained, though.

Lobby for legislation. When courts don't deliver, legislatures are the next stop. Press bodies in several countries are pushing for remuneration rights or compulsory licensing for AI training, with active debate in the EU, Australia and Canada. Slow, but capable of more fundamental change than case law.

Go to the table. Which is what most publishers are actually doing. If the odds of winning in court are low, early licensing can be the rational choice — though it's hard to deny this ruling improves the AI side's negotiating position.

India-specific variable: appeals are common and frequently run all the way to the Supreme Court. If ANI appeals, a Delhi High Court division bench and potentially the Supreme Court will revisit this. With no prior application of Section 52 to foundation-model training, the odds of a different interpretation upstairs are real.

What actually changes for you

If you create content, nothing shifts immediately, but the direction is clear. Blocking training itself is becoming legally harder, and the defensible lines are moving toward access control (keeping crawlers off your site) and output monitoring (checking whether your material comes back out verbatim).

If you build AI products, the operational lesson from this line of cases isn't about training — it's about acquisition path. What produced Anthropic's $1.5 billion settlement wasn't training; it was pirated source material. Use only legitimately accessible data and keep records of how you got it. That's the highest-value risk control available right now.

If you're a media executive, this is a signal to revisit licensing strategy. If the "sue and win" scenario has gotten less probable, your negotiating card has to shift from legal threat to the actual value of your data — recency, exclusivity, structure. AI companies pay because they need the data, not ultimately because they fear the lawsuit.

If you work in law or policy, watch the convergence of reasoning, not the outcome. Structurally different regimes — US fair use, Indian fair dealing — are arriving at similar conclusions. That changes the starting point for legislative debate everywhere.

From a Korean perspective, the notable fact is the absence of comparable case law. Korea's Copyright Act has an ongoing debate around a text-and-data-mining exception, and the Ministry of Culture has issued AI-copyright guidance, but judicial findings haven't accumulated. When domestic disputes between Korean publishers and AI companies arrive in earnest, Indian and American rulings will show up as reference material.

If you're a user, nothing to feel today. But as these rulings accumulate, the odds of a service like ChatGPT being abruptly restricted in a given country go down. Two years ago that risk was real. It's steadily shrinking.

🥄 Three Things You're Probably Wondering

— So AI can train on news articles freely now? It's not that simple. This is an Indian court, at the interim stage, making a prima facie finding. It can be reversed at trial and an appeal is still available. And "training is permitted" is a completely different claim from "reproducing the original in outputs is permitted." AI companies still lose cases where reproduction is proven.

— Does an Indian ruling affect the US or Korea? It has no binding force whatsoever outside India. What does happen in practice is that courts elsewhere cite foreign decisions as comparative material when facing similar questions. And the fact that AI companies can now point to a major jurisdiction having reasoned this way is a genuine change in negotiations and litigation posture.

— What happens to ANI now? The main suit continues, an appeal is available, and one looks likely. But losing at the interim stage means weaker leverage, so resolving through a renegotiated licensing deal rather than litigating to judgment can't be ruled out. Either path takes a long time to reach an ending.

References

Numbers and criteria are as of announcement and may change.