Delhi judge rejects ANI's injunction, says OpenAI's ChatGPT training looks like fair use

A Delhi High Court handed OpenAI a significant preliminary win on Thursday, rejecting news agency ANI’s request for an injunction that would have blocked the company from using its copyrighted content. Justice Amit Bansal didn’t mince words about the weakness of ANI’s case, noting the agency couldn’t produce a single example of ChatGPT memorizing and reproducing its articles verbatim. The evidence ANI submitted actually backfired: OpenAI demonstrated that the articles cited, published in late 2024, post-dated the training data cutoffs for the GPT-4 and GPT-4o models. The judge’s preliminary view is that any similarities likely came from retrieval augmented generation, a real-time search-like function, rather than memorization during training.

What makes this ruling particularly notable is its treatment of the training process itself. Justice Bansal tentatively found that ingesting copyrighted works for AI training falls under India’s private use and research exception. He set a crucial condition, however: the training copies must come from lawful sources, not pirated repositories or paywalled sites. Legal scholar Andres Guadamuz pointed out that this appears to be the first time a court has explicitly accepted a private use defense for AI training. The judge applied a three-part fairness test and sided with OpenAI on all counts, finding no economic harm to ANI because the two operate in totally different sectors.

ANI’s case unraveled in other ways too. The news agency had resorted to adversarial prompting, explicitly instructing ChatGPT to reproduce articles exactly, and still came up empty-handed. Justice Bansal underscored that mere facts in news articles aren’t copyrightable, and reproducing topics or headlines doesn’t amount to direct competition. You can’t claim economic damage from someone who isn’t your competitor, and you definitely can’t claim it when your own evidence fails to show the model spitting out your stories.

This ruling lands in a global landscape that is, frankly, a mess of contradictions. U.S. courts have tossed similar cases from Raw Story and AlterNet for lack of concrete harm, and the GitHub Copilot plaintiffs couldn’t find a single identical code snippet. Yet The Intercept has seen partial success through DMCA claims. The Delhi decision doesn’t settle the core copyright questions permanently—the court will revisit memorization and RAG in the main proceedings. But for now, it’s a forceful signal that courts won’t be impressed by lawsuits built on speculation rather than actual evidence of copying.