ANI Media v. OpenAI — Delhi High Court Treats AI Training as Prima Facie Fair Dealing

Background

Indian news agency ANI sued OpenAI, alleging that the company copied and stored ANI’s news reports to train the large language models behind ChatGPT and that ChatGPT could reproduce protected portions of ANI’s work. ANI sought an interim injunction that would restrict OpenAI’s use of its content while the lawsuit continued.

The 135-page judgment addressed whether Indian courts had jurisdiction even though OpenAI said its training servers were outside India, whether model training and storage implicated ANI’s reproduction rights, whether ChatGPT outputs copied protected expression, and whether the use qualified as fair dealing under Section 52(1)(a) of India’s Copyright Act. The ruling concerns interim relief only; the underlying suit remains pending.

The Court’s Holding

Justice Amit Bansal denied the injunction. At this preliminary stage, he concluded that storing ANI’s publicly available literary works to train an LLM fell within Section 52(1)(a), which protects fair dealing for private or personal use, including research. The court treated model training as research even though OpenAI operates commercially, reasoning that commercial status does not automatically defeat the statutory exception.

The court found both the purpose and fairness components satisfied. Training uses works to extract patterns and relationships rather than to offer substitutes for the original reports. The judgment also emphasized that ANI had not shown economic competition or lost subscribers attributable to ChatGPT, while ANI’s own licensing proposal demonstrated that any eventual injury could be measured in money.

On outputs, the court found no sufficient proof that ChatGPT had memorized and regurgitated ANI’s protected expression. News facts themselves—such as government decisions, judicial rulings, sports results, and public statements—are not monopolized merely because ANI reported them first. The relevant question was whether an output reproduced a substantial part of ANI’s original expression. The examples before the court, including retrieval-augmented generation results, were not substantially similar enough to establish a prima facie infringement case.

The balance of hardships and public interest also favored OpenAI. The court noted that ANI and its subscribers could block web crawlers, that OpenAI represented it had blocked ANI’s site for training and search retrieval, and that a broad injunction could disrupt LLM development and millions of users. Requiring licenses from every source of publicly available training data, the court reasoned, could make development economically impractical, including for Indian AI companies.

The judgment expressly cautioned that all findings are preliminary and will not control the final outcome after trial.

Key Takeaways

• The Delhi High Court became one of the first courts to hold that copying publicly available works for generative-AI training can prima facie qualify as fair dealing under Indian copyright law.
• A commercial AI service may still invoke the research branch of Section 52(1)(a); commercial purpose is relevant but not automatically disqualifying.
• Copyright protects a news publisher’s original expression, not the underlying facts, and infringement by model output requires substantial similarity rather than merely answering the same factual question.
• Evidence of actual memorization, regurgitation, market substitution, or lost licensing revenue will remain important as the case proceeds.
• The decision denies only interim relief and does not finally resolve liability.

Why It Matters

The ruling is a major early signal for AI developers and publishers in India. It adopts a relatively broad reading of fair dealing for model training and places significant weight on transformation, public access to data, proof of market harm, and the social cost of restricting LLM development. That approach may influence other Indian cases and provides an important comparative-law counterpoint to still-developing U.S. and European AI-copyright doctrine.

Publishers should not read the decision as eliminating copyright risk. The court left room for claims supported by evidence that a model stores and reproduces protected expression, competes with the source, or causes measurable market damage. Developers, meanwhile, gain a strong interim precedent but not a final safe harbor.