India's First AI Training Ruling Went OpenAI's Way

On July 24, Justice Amit Bansal of the Delhi High Court refused to grant the news agency ANI an interim injunction against OpenAI. It is the first substantive judicial finding in India on whether training a large language model on copyrighted news requires a licence, and the answer, at this stage, is no.

The result reads like a clean win for OpenAI. The reasoning is more interesting than that, because the court gave ANI the thing it needed to be in court at all — and then took away everything it came for.

Jurisdiction first, and ANI won that part

OpenAI's threshold argument in this case has always been geographic. Its servers are not in India. Its training was not performed in India. Therefore, it argued, Indian copyright law has nothing to attach to.

The court did not accept that. It held that the temporary electronic storage of ANI's works during the training process constitutes reproduction under Section 14 of the Copyright Act, 1957 — and reproduction is an act the statute reaches. That finding is what kept the case alive and what makes this ruling matter beyond its outcome. An Indian court has now said that ingesting Indian copyrighted material into a model is a copyright-relevant act under Indian law, regardless of which country the GPUs sit in.

Every lab operating in India should read that sentence twice. It is the durable half of this judgment. The territorial defence — we trained elsewhere, so you cannot touch us — did not survive.

Then the exception swallowed the claim

Having established that reproduction occurred, the court then held it was excused. Justice Bansal found that OpenAI's storage of ANI's literary works for the purpose of training the models underlying ChatGPT falls prima facie within Section 52(1)(a) of the Copyright Act — the fair-dealing carve-out covering "private or personal use, including research."

This is where India's statute diverges sharply from American law, and where the ruling will be argued about hardest. US fair use is an open-ended, four-factor balancing test; a court weighs purpose, nature, amount, and market effect, and can reach almost any result. India's Section 52 is a closed list of enumerated exceptions. Fitting commercial model training into "private or personal use, including research" is a construction that stretches a provision written for scholars and photocopiers across a data pipeline feeding a product with hundreds of millions of users.

The court's framing is that training is a research-type act on the input side, distinct from the commercial character of the output product. Whether that distinction holds up at final hearing — with a full evidentiary record rather than an interim application — is genuinely open. Bansal himself flagged the limit: the observations are confined to the interim application and carry no weight on the final outcome. The main suit is still pending.

The RAG finding is the sharpest part

ANI's second theory was about outputs, not inputs. It pointed to ChatGPT responses reproducing or closely tracking its reporting, which would be infringement independent of how the model was trained.

The court rejected it on two grounds, and the second one is unusually well-reasoned for a technical question in a copyright case.

First, the ordinary test: the outputs were not substantially similar to ANI's works. Similarity of facts is not similarity of expression, and copyright protects only the latter.

Second — and this is the part worth studying — the court reasoned about where the text came from. It noted that GPT-4's training data ended in April 2022 and GPT-4o's in April 2024, while the ANI articles at issue were published in August and September 2024. Content that postdates a model's training cutoff cannot have been memorized during pre-training. The court therefore inferred that the responses were produced by live retrieval — retrieval-augmented generation — rather than recalled from model weights, and held that RAG-generated outputs did not infringe.

That is a court using a training cutoff as forensic evidence, and it establishes a distinction most litigation has been muddling: memorization and retrieval are different mechanisms with different liability profiles. A model reciting from weights is a copying question. A model fetching a live page and summarizing it is closer to what a browser does — and the court treated it that way.

The implication cuts both ways for labs. RAG looks safer here. But the same logic invites a plaintiff with pre-cutoff material and evidence of verbatim regurgitation to build a much stronger case than ANI could. ANI's timing worked against it: it sued over articles its adversary could not plausibly have memorized.

What this changes

ANI filed in November 2024, the first Indian media organization to take OpenAI to court, and the case has been watched as a bellwether for how the world's largest internet market by users would treat model training. India matters here in a way that is easy to underrate — it is one of ChatGPT's biggest user bases and a jurisdiction with an aggressive, well-organized news industry.

For now, the score reads: Indian courts can hear these cases, and training is provisionally excused. Publishers negotiating licensing deals in India just lost leverage — the alternative to a deal is a lawsuit that has now failed at its first serious test. Labs gained a template for arguing training as research under a statute that, on its face, seemed less accommodating than US fair use.

None of it is final. Interim relief is a low-resolution instrument, decided on prima facie views and the balance of convenience, and the substantive suit will be litigated on a full record with expert evidence about what these systems actually retain. Parallel cases are running in the United States and Canada, and the more national courts land on the same question, the more the fault lines show.

But the first data point from India is in, and it favors the labs.