Delhi Court Says Training on News Archives Is Fair Dealing

Consider the position of a wire service editor watching this case. Your archive is the asset. Your reporters filed every story in it, your subscribers pay for access to it, and a model you had no agreement with has read all of it. That was the argument Asian News International brought to the Delhi High Court, and on July 24 it lost the first round.

Justice Amit Bansal denied ANI's request for an interim injunction against OpenAI in a 135-page order following 32 hearings. The court held that storing ANI's content to train a large language model falls within the fair dealing exception in Section 52(1)(a) of India's Copyright Act, which covers private or personal use including research. It is the first explicit judicial finding in India that model training fits inside that exception.

Inside the 135-page order

The court applied a three-part fairness assessment and found for OpenAI on each: the training use was limited in scope, ANI had not demonstrated economic harm given that the two parties operate in different markets, and outputs consisting of topics and headlines did not directly compete with ANI's product.

The reasoning matters more than the outcome. A ruling that training is permitted because a plaintiff failed to show harm leaves room for a different plaintiff to succeed by showing it. This is a narrow win read closely, even if the headlines make it sound sweeping.

The reproduction claim collapsed on dates

ANI's second argument was that ChatGPT reproduced its articles, and it submitted outputs as evidence. OpenAI answered with timestamps: its training data cutoffs of April 2022 and April 2024 predated the ANI articles cited, which ran in August and September 2024. The similarities, the judge concluded, came from retrieval augmented generation pulling live web content rather than from memorized training data.

The court also noted that even under adversarial prompting explicitly designed to extract exact text, ANI could not produce a single verbatim copy. For anyone assembling a copyright case against a model developer, that is the practical bar being set: demonstrated verbatim reproduction, not resemblance.

For publishers outside India too

The main suit is still pending, and the court signaled it will revisit whether RAG constitutes communication to the public, along with the memorization questions, in the full proceedings. So this is an interim position rather than settled Indian law.

Read alongside other jurisdictions, the picture is genuinely split. This ruling lines up with US decisions that have dismissed copyright claims against language model developers, and diverges from European decisions treating training as copyright-relevant reproduction. Any organization licensing content, or relying on content it does not own, is now operating under rules that differ by market rather than converging.

The operational takeaway is unglamorous: if your business depends on either side of this question, the contract is doing more work than the case law. Licensing terms, indemnities and data provenance clauses are what will hold when the jurisdictions disagree, and they disagree right now.

Sources: The Decoder, Business Standard, India Legal