Introduction
On 24 July 2026, the Delhi High Court refused ANI an interim injunction against OpenAI, holding prima facie that storing ANI’s news content to train ChatGPT falls within the fair dealing exception in Section 52(1)(a) of the Copyright Act, 1957 (judgment). Read broadly, the order looks like clearance for AI developers to train on copyrighted works. This post argues that it is narrower than it appears. The order concerned news articles and a text model, and it rested partly on the absence of shown market harm. Visual art differs on exactly those points, so whether training image generators on artists’ work is fair dealing remains open in India.
What the Court decided?
ANI sued OpenAI in 2024, alleging that it scraped and stored ANI’s articles to train ChatGPT and that ChatGPT reproduced them in its responses. Justice Amit Bansal accepted that storing the works during training engages the reproduction right, but held, at the prima facie stage, that the storage was protected by Section 52(1)(a). The Court treated training as an internal research process that neither substituted ANI’s works nor demonstrably affected its market, and it gave weight to the public interest in AI development. On outputs, it found that ANI had not shown ChatGPT regurgitating its work.
Two cautions follow. First, the order is interim and the main suit is still pending. Second, the result was tied to its record: the articles ANI pointed to as examples of copying were published after training had ended, and no substitution was demonstrated. A plaintiff with a different market and a different record can be expected to fare differently. Visual artists are such plaintiffs.
Copying is still reproduction
The threshold is not controversial. Section 14(c)(i) gives the owner of an artistic work the exclusive right to reproduce it in any material form, expressly including storing it in any medium by electronic or other means, and Section 51 treats doing so without a licence as infringement. Assembling a training dataset means copying whole images, often at very large scale, and the ANI order accepts that storage during training engages the right. The real contest is over the exception.
Where Fairness parts company with ANI?
The Court asked two questions: whether the purpose fell within Section 52(1)(a), and whether the use was fair. Indian courts have not settled on a single fairness test. The factors familiar from US law (purpose, nature of the work, amount taken and market effect) offer a workable frame, and the ANI order is reported to have drawn on US decisions. On each factor, visual art fares differently from the ANI facts.
Purpose. The Court read “research” as extending to machine learning and weighed a public-interest consideration that critics note does not appear in the text of Section 52(1)(a). Image generators are typically sold by subscription, and their outputs compete with commissioned illustration and design. The further a use travels from a developer’s laboratory to a product sold into the artist’s own market, the less work the word “research” can do.
Nature of the work. OpenAI argued that ANI’s articles were factual content. Whatever the merit of that argument for news, it has no purchase on paintings, illustrations and photographs, whose value lies in the author’s choices of composition, colour and line, which are precisely what generative models learn to reproduce.
Amount taken. Training ingests works whole. Copying the whole is not automatically unfair, but Indian courts also look at what was taken, not only how much: in Super Cassettes v. Hamar Television the Delhi High Court treated the quality of what was taken as important, not merely its quantity. Where outputs can be traced to an artist’s distinctive expression, the taking is both total and substantial.
Market effect. Here ANI and art part company most sharply. ANI failed in part because substitution was not demonstrated. For artists, substitution is the product: an image generator sells, in seconds and at low cost, what illustrators, designers and photographers sell by commission. In Blackwood & Sons v. Parasuraman, decided under the earlier 1911 Act, the Madras High Court accepted that an intention to compete and profit from the original makes a dealing unfair. An artist who can show lost commissions, or outputs marketed as replacements for commissioned work, removes the finding that carried the day in ANI. Evidence will decide these cases: artists should document the prompts that name them, the outputs that resemble their works and the commissions they have lost.
The Limits of the Argument
Copyright protects expression, not ideas,[1] and an image made “in the manner of” a studio does not, without more, copy any particular protected work. The Ghibli-style image trend shows why the public perceives a harm the Act may not recognise; it also shows the evidential gap, since whether the model was trained on Ghibli’s work was left open. Artists’ claims are therefore strongest at the stages the Act does reach: copying into training datasets, and outputs that reproduce identifiable expression. The grievance about style is better answered by legislation than by Section 51.
Conclusion
ANI is a first word, not the last. It shows that Indian courts are open to reading Section 52(1)(a) to cover AI training where no market harm is shown and public interest is invoked. Whether that reasoning survives a plaintiff whose market is being substituted has not been tested. Two things follow. First, courts should treat commercial deployment and market substitution as central to any fair dealing finding for generative models, not as incidental. Second, the Court’s concern that licensing every source would be unworkable is a reason to design collective licensing or remuneration mechanisms, not to leave artists to litigate one image at a time. If “research” is read broadly enough to cover commercial training, Parliament should pair that breadth with a route to compensation.
[1]R.G. Anand v. Delux Films, AIR 1978 SC 1613; (1978) 4 SCC 118.
