Training is the engine room of generative artificial intelligence (GenAI). As these models become more scalable, they will require even larger volumes of content, including syndicated and non-syndicated online data.
This raises an unavoidable question: Can AI developers train on such content without obtaining consent or licences from rights holders?
That question reached the Delhi high court in the ANI Media vs OpenAI case. ANI alleged that OpenAI had stored and used its news reports to train the models underlying ChatGPT and that the chatbot could reproduce its protected content.
ANI sought an immediate restraint in the form of an interim injunction, including deletion of its works allegedly stored by OpenAI.
While the suit remains pending, the court denied ANI an interim injunction.
What the court said: First, the court held that storing publicly available works to train large language models (LLMs) could qualify as ‘fair dealing’ for private use, including research; meaning one could use such copyrighted works without the express permission of the copyright holder.
The court reasoned that training data which is retained in a closed environment and not disclosed to the public amounts to private use. Also, applying an updated interpretation to an older law, it held that machine learning could qualify as research.
Second, it refused to treat OpenAI’s commercial character as an automatic disqualification. Research does not necessarily cease to be research merely because a company undertakes it or because the resulting product is monetized.
This raises an interesting question about what ‘commercial’ use really means. Professional journalism is routinely funded through advertising, subscriptions and syndication. Yet the legal protection for fair reporting of current events is not confined to non-profit journalism. The commercial setting in which an activity occurs cannot, therefore, operate as an unstated threshold bar to fair dealing.
This does not mean that every commercial use is fair. It only means that an activity otherwise covered by fair dealing does not lose that protection merely because it is undertaken commercially. Market substitution and economic harm remain relevant to determine whether the use is fair.
Third, the court found insufficient evidence that ChatGPT had memorized or substantially reproduced ANI’s protected expression. It distinguished the material used to train a model from the responses later generated for users. Infringement at the output stage must be established independently.
These findings are preliminary. The ruling neither creates blanket immunity for AI training nor prevents claims against outputs that reproduce protected works.
Who should bear the interim risk? The order’s strongest policy signal lies in its assessment of the competing consequences of an injunction. The court noted that ANI had offered OpenAI a licence for $7.5 million, demonstrating that its asserted injury was capable of monetary valuation. If ANI ultimately succeeds, its loss may be monetarily compensated.
The possible consequences of an injunction were harder to contain. The court observed that the development and success of LLMs depend on the availability of data, and that requiring licences from multiple sources could make LLM development economically unviable. The court concluded that an injunction could prejudice not only OpenAI, but millions of users, technology advancements and the public at large.
The court did not discard ANI’s rights in favour of an abstract promise of innovation. It made a narrower interim choice: a potentially compensable licensing loss did not justify restraining a technology where the wider consequences might be more difficult to reverse.
Innovation cannot mean immunity: Publishers’ concerns nonetheless remain legitimate. Journalism requires sustained investment in gathering, verifying and presenting information. If AI products reproduce or substitute for publishers’ content, they could erode the traffic, subscriptions and licensing revenues that fund its creation.
The court’s separation of training from outputs is therefore critical. Protection for training should not immunize memorization, regurgitation or real-time retrieval that substitutes the original work. The next disputes may focus less on what a model learned and more on what an AI product retrieves, presents and monetizes.
Going forward: The judicial and regulatory outlooks may now be moving in different directions.
The court found an interim answer within the existing fair dealing framework. However, the Department for Promotion of Industry and Internal Trade working paper on generative AI and copyright, by contrast, proposes a mandatory licensing framework under which developers can train on content while rights holders receive recurring remuneration through a centralized mechanism.
While both approaches seek to enhance AI development, they differ fundamentally on whether rights holders must be paid for enabling such development.
It will be interesting to see how this divergence plays out. Parliament may endorse the judicial approach, introduce a compensation mechanism or redraw the balance altogether.
The real question after ANI vs OpenAI is how India can develop a clear and workable framework that supports competitive AI development while balancing the interests of rights holders.
While the interim order provides the first judicial answer, it is unlikely to be the last word. As the technology evolves and the litigation progresses, the legal position may continue to develop through further judicial and legislative engagement.
The authors are, respectively, partner and head of digital disruption, and associate, Cyril Amarchand Mangaldas