Fair Dealing In The Age Of AI: Lessons From ANI v OpenAI

Introduction 

In July 2026, the Delhi High Court declined to grant ANI Media an interim injunction against OpenAI, being the first Indian Court to substantively engage on whether training a large language model with copyrighted content is unlawful.¹ At the centre of this finding is whether OpenAI’s use of ANI’s articles to train its models, fall under the ambit of the ‘fair dealing’ exception under Section 52(1)(a) of the Copyright Act, 1957 (“Act”).

Section 52(1)(a) has never been amended or been the subject of a judicial proceeding, in the context of AI training before. A prima facie interpretation of this provision meant the Court had to resolve three distinct issues:

1. Commercial use: Merely because OpenAI is a commercial enterprise that monetises its tool ‘ChatGPT’, does their activity of training the AI model using copyrighted material, take it outside the purview of “private or personal use”?
2. “Private”: Can an activity carried out within a company’s internal systems, rather than by a single individual, still be “private” within the meaning of the section?
3. “Research”: Does extracting patterns and relationships across a dataset to train models on text, fall within the term “research”?

This article examines the Court’s prima facie reasoning and what this means going forward for AI developers training models using content sourced or published by various entities in India.

Question before the Court

The Act in India follows a “fair dealing” model, which is narrower than the open-ended “fair use” standard used in the United States. Rather than asking Courts to weigh a set of factors in any given case, the law in India sets out specific purposes for which copyrighted work can be used without permission, which are listed exhaustively in Section 52 of the Act. If a use does not fall within one of these listed purposes, the fair dealing defence is not available at all, regardless of how reasonable the use might otherwise seem. Section 52(1)(a) is one such listed purpose.

Section 52(1)(a) of the Act provides that a fair dealing of a literary work for the purpose of “private or personal use, including research” does not constitute infringement. The provision itself is brief. It names the purpose, sets no further conditions and leaves “private”, “personal”, and “research” undefined.

This provision was drafted long before AI existed. The question in this case, therefore, was whether it could still apply to something the Parliament never anticipated i.e., a commercial company using huge amounts of copyrighted text to train a large AI model. ANI argued that this provision was meant for an individual engaged in personal study or inquiry and not a company operating a global AI product.

Each question was decided in OpenAI’s favour and rested on different interpretations, including statutory comparisons, case law extending to interpretation of various terms including “private”, “research” etc.

The core reasoning 

The Court held that “Commercial use” is not a disqualification by itself. ANI’s main objection was that “private or personal use” is a language built for an individual and not for a company built to profit from the very use in question. The Court tested this against the backdrop of the remaining clauses under Section 52(1) which include translations, adaptations and educational use, several of which expressly exclude commercial use in their text. Section 52(1)(a) contains no such exclusion. Read this against the other sub-clauses, the Court treated that silence as deliberate stating that where Parliament wanted to bar commercial use, it said so and where it did not, the omission carries meaning. On this basis, OpenAI’s commercial nature did not by itself remove it from the exception.

The Court then held that “Private” extends beyond a single individual. The Court declined to treat “private” and “personal” as synonyms, reasoning that each word in the provision should carry independent meaning. It relied on the case of Academy of General Education, Manipal v. B. Malini Mallya², where a dance performance staged by an educational institution, not an individual, was held to be private use. Applying that principle, the Court found that OpenAI’s training data, held within internal systems and never disclosed or displayed to any person, was “purely private” in nature, notwithstanding that the entity holding is a large organisation rather than a single person.

On “Research”, the Court held that is extends to machine training. “Research” is undefined in the Act, and the Court drew reference from Blackwood v. Parasuraman³, as a form of diligent investigation whose process remains internal and output alone becomes public. It found that LLM training with its extraction of statistical patterns and relationships from text, refined through repeated correction, fits that description in substance, even though the “investigator” is a machine rather than a person. To bridge the gap between a word drafted decades before this technology existed and its application here, the Court relied on State (Through CBI/ New Delhi) v. S.J. Choudhary⁴, to invoke the doctrine of updating construction. In this case, the Hon’ble Supreme Court held that a typewritten document could satisfy requirements drafted for handwritten ones, since a statute should be read as capable of extending to new technology performing the same underlying function. Applying that reasoning here, the Court held that machine training carried out at the direction of and for the benefit of humans, can qualify as “research” for the purposes of Section 52(1)(a).

Together, the three findings meant that OpenAI’s use of ANI’s articles for training purposes fell within the fair dealing exception.

Implications for AI developers

For AI companies operating in or serving the Indian market, the most immediate practical effect is on the sourcing of Indian content without a license. This takes pressure off from such companies, from procuring any form of licenses for content. Companies training models at scale are likely to treat this as a green light to continue ingesting content published in India, without necessarily prioritising licensing, since the legal risk of doing so has visibly dropped. OpenAI has struck paid deals with outlets like Associated Press, the Financial Times and Conde Nast elsewhere, but this ruling will reduce the urgency to do the same in India, at least for now.

At the same time, some of what worked in OpenAI’s favour, is worth building into practice. Being able to show on demand, exactly what was trained on what data and when, training cutoffs, sourcing trails, etc., was central to defeating the reproduction claim. Companies that cannot produce this documentation may be in a materially weaker position in a similar dispute. Sourcing discipline also matters just as much. The Court drew a comparison with Bartz v. Anthropic,⁵ where part of the training data came from unrecognised and illegal sources, and noted that ANI’s content, by contrast, was not obtained that way. OpenAI had scraped it from publicly accessible parts of the internet. This distinction was significant because it kept the “research” finding tied to lawfully accessed content. Therefore, companies using scraped or aggregated datasets should check and be able to show where that data comes from. The reasoning that protected OpenAI rested specifically on the fact that its content was lawfully accessible - if a company is unable to demonstrate this, regardless of how the training process itself is characterised, this defence may not be available. In essence, calling the activity “research” does not, on its own, excuse how the underlying data was sourced. The two are treated as separate questions.

Implications for publishers and copyright holders 

For Indian publishers, this order weakens their negotiating position with AI companies. Until this case, publishers could reasonably argue that unlicensed training using their content sat in a legal grey zone, which gave them some leverage to push for licensing deals. That grey zone has narrowed. Wire services like ANI are likely to bear this the first and hardest since their business depends specifically on licensing content to other outlets.

There is also a coalition angle worth looking at. Bodies like the Digital News Publishers Association, Indian Music Industry and the Federation of Indian Publishers had intervened on ANI’s side arguing that ChatGPT’s outputs amount to substantial reproduction of copyrighted works and this ruling is a setback in that sense. It is likely to shift their energy to building a share licensing pool so publishers can negotiate with AI companies as a group rather than individually and, push for a legislative fix that changes the underlying law.

On the practical side, publishers who have not used basic technical protections are likely to start treating them as standard practice. OpenAI, like most AI companies, allows website owners to opt out of crawling and the Court found that this option was available to ANI throughout, but that ANI had simply not used it. This counted against ANI when the Court decided whether urgent relief was warranted. A publisher that could have protected itself technically, but chose not to, has a weaker claim. Going forward, publishers considering their claims will need to come in with real numbers of traffic data, subscriber loss and revenue tied to AI substitution. They will also need clearer evidence of actual reproduction, not just training. The Court found that ChatGPT does not store or retrieve full copies of articles to answer questions. It generates a response based on patterns learned during training, applied to what the user asks. This means it does not necessarily reproduce any article word for word, even when it has been trained on that article. Showing that an AI tool was trained on a publisher’s content is therefore not the same as showing that it reproduces that content and future claims will need to establish the latter directly.

Conclusion 

By holding that AI training can qualify as “research”, that commercial use does not rule out the exception and that a company’s internal systems can count as “private”, the court has given AI developers a workable position and narrowed the ground for publishers.

For AI developers, the advantage gained here should be reinforced through good documentation and clean data sourcing. For publishers, the better response is likely not to revisit this particular loss, but to prepare for what comes next, which is, stronger evidence of harm, more consistent use of technical safeguards and greater reliance on collective action alongside legislative change.

Footnotes

^{1 https://www.thehindu.com/news/national/delhi-hc-declines-interim-injunction-against-openai-in-ani-copyright-suit/article71261655.ece}

^{2 Academy of General Education, Manipal v. B. Malini Mallya, (2009) 4 SCC 256.}

^{3 Blackwood v. Parasuraman, 1958 SCC OnLine Mad 62.}

^{4 State (Through CBI/ New Delhi) v. S.J. Choudhary, 1990 SCC OnLine SC 174.}

^{5 Bartz v. Anthropic, 3:24-cv-05417 (N.D. Cal.).}

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.