The Delhi high court recently dismissed Asian News International’s (ANI) application for an interim injunction against OpenAI over its use of the news agency’s copyright-protected content to train ChatGPT.
This is India’s first judicial answer to a question confronting courts worldwide: may a Generative AI company train its models on copyrighted content without permission? At this interim stage, the answer is an unequivocal yes.
ANI argued that OpenAI copied and stored its articles to train ChatGPT, which then reproduced its content in generated responses. The court accepted that storing copyrighted works for training is ‘reproduction’ under Section 14 of the Copyright Act of 1957, but held that it falls within the “fair dealing” exception in Section 52(1)(a) and is not infringement. It further held that ChatGPT’s responses were not substantially similar to ANI’s articles.
The court thereby substantially expands the fair dealing exception, an interpretation that sits uneasily with the text of the Act and the principles that underpin copyright law.
Re-assessment of the purpose: The court held that the storage of copyrighted works in this case is protected by the fair dealing exception since it permits use for “private or personal use, including research.” It reached this conclusion by giving both ‘private use’ and ‘research’ an expansive interpretation.
Relying on the Copyright (Amendment) Act of 2012, which inserted the word ‘personal’ alongside the existing phrase ‘private use,’ it held that Parliament did not intend to restrict the exception to individuals. Instead, ‘private’ was interpreted broadly to include a private corporate entity.
It also rejected the argument that ‘research’ is confined to academic and non-commercial activity, holding that commercial use does not, by itself, go against OpenAI’s fair-dealing defence.
This marks a significant shift in the understanding of Section 52(1)(a)(i). Traditionally, ‘private use’ has been understood as use by individuals for their own purposes, while ‘research’ has generally been associated with academic or other non-commercial inquiry.
The judgement extends both concepts to cover the large-scale commercial training of AI models. Whether Parliament intended such an expansion is, however, open to debate. The 2012 amendment was enacted to bring India’s copyright framework closer to international standards.
Yet, comparable jurisdictions have a far narrower approach. The EU limits the exception to private, non-commercial use by a natural person, the UK to non-commercial purposes and Australia to “private and domestic use.”
It also sits uneasily with the Delhi high court’s decision on Rupendra Kashyap vs Jiwan Publishing House, which held that Section 52(1)(a)(i) is unavailable as a resort where the use amounts to commercial exploitation.
The court’s understanding of ‘research’ is equally significant. The UK limits the exception to non-commercial research. Australia and New Zealand assess fairness by reference to factors such as the purpose of the copying, its market effect, the amount copied and the availability of licences.
In India, the Delhi high court adopted a broader interpretation that includes commercial AI training.
Re-assessment of fairness: How the court assessed fairness is equally significant, as a central measure of it is the effect on the market for the work that was used. The court reasoned that OpenAI’s use of ANI’s literary works caused neither actual nor potential harm to the agency.
Yet, OpenAI’s own commercial practices suggest otherwise. It licenses content from publishers such as Axios, Vox Media and the Financial Times. It could have worked out a similar deal for ANI, which alleged differential treatment that places it at a commercial disadvantage.
The court observed that ANI did not demonstrate any loss of market share or subscription revenue. However, those are not the only forms of economic harm that the copyright law seeks to prevent. It also protects an owner’s ability to commercially license its works.
OpenAI itself has recognized that market by licensing works from several publishers for AI training. If it denied ANI similar terms, then it arguably deprived ANI of a commercially valuable opportunity.
This market is far from hypothetical. In Bartz vs Anthropic, the plaintiffs valued AI training rights at about $3,113 per work. This illustrates that such rights have measurable commercial value.
Legislative reform, please: The transformative potential of AI and the public benefits it promises might influence how this issue is viewed, but even so, this presumptive gain must be balanced against the rights of copyright owners.
The Copyright Act predates GenAI and was never designed to address AI training, which may well warrant legislative reform. Jurisdictions that have accommodated AI training have typically legislated rather than expanded existing exceptions. Singapore has introduced a dedicated exception for computational data analysis, while the EU enacted text and data mining exceptions.
India is exploring legislative solutions too, with a proposed framework that would permit AI training on lawfully accessed works while ensuring fair remuneration through collective licensing.
Balancing innovation with the rights of creators is a legislative exercise. It cannot be done through extensions of a fair-dealing exception beyond what the Copyright Act originally intended to cover.
The authors are, respectively, associate, Nishith Desai Associates, and associate, Resolüt Partners.
