Summary: This article examines the Delhi High Court’s judgement in ANI v. Open AI as more than a copyright ruling, arguing that its reasoning on what constitutes “research” offers a persuasive analytical framework for one of the central uncertainties under the DPDPA: whether commercial AI model training can qualify for the Act’s research exemption. It contends that the Court’s purpose, fairness and public interest test, while developed under Section 52(1)(a) of the Copyright Act, translates naturally to Section 17(2)(b) of the DPDPA, and removes a significant conceptual obstacle to treating AI training as research under data protection law. It will be relevant to readers tracking the evolving Indian jurisprudence on AI and its intersection with copyright and data protection law.
Much has been, and will continue to be, written about the Delhi High Court’s judgement in the interim application in ANI Media Pvt. Ltd. v. Open AI OpCo LLC[1] (“ANI v. Open AI”), delivered on July 24, 2026, after nearly four months in reserve. The judgement dismisses ANI’s injunction application in a suit that continues to trial.
As it takes its place beside Bartz v. Anthropic[2] and Kadrey v. Meta[3] in common law jurisprudence on the viability of the fair use and fair dealing defence for AI training, this article argues that the decision offers more to India’s fledgling AI ecosystem than its copyright findings alone.
Innovation often outpaces regulation in India. But innovation built on an unsteady regulatory foundation carries risks. As is evident from the DPIIT committee’s ‘Working Paper on Generative AI and Copyright’, the committee has, in categorical terms, concluded that using copyright-protected material to train AI models without royalty payment would be an infringing act.[4]
Viewed against that backdrop, a judicial finding on the scope of fair dealing under Section 52(1)(a) of the Copyright Act, 1957 (“Copyright Act”), settles, at least for now, whether private storage of copies for training, and generation of inferences from that data, infringes copyright, and it is valuable well beyond the four corners of the dispute itself.
The court’s findings do not merely clear the air surrounding development of AI models and systems in India under copyright law, its reasoning also offers an analytical framework for one of the central uncertainties under the Digital Personal Data Protection Act, 2023 (“DPDPA”) – whether commercial AI model training may qualify for a ‘research’ exception.
The Copyright Act and the DPDPA protect distinct legal interests – creative expression on the one hand, and privacy on the other, with Copyright Act protecting an economic interest and DPDPA operationalising a fundamental right. Yet, both confront the same foundational question in the context of generative AI: whether AI model training constitutes “research” deserving of an exception, despite its incidental use of protected material.
The training of LLMs uses data that often includes personal data: names, quotes, images and other identifiers woven into news reports, social media posts and other material, which forms the “oil” for AI training, and are hard to exclude.
The DPDPA’s consent-focussed architecture provides very few non-consent bases, unlike the GDPR. The narrowly conditioned exemption for “publicly available data[5]”, and the limited research exemption, which exempts processing “necessary for research, archiving or statistical purposes” under standards to be prescribed by the Central Government[6], offer a much less clear “safe harbour” for AI training, particularly when read with the broad prescriptions in the Second Schedule of the Digital Personal Data Protection Rules, 2025 (“DPDP Rules”). This has led to questions on whether the research exemption covers AI development at all, or whether it excludes commercial or automated processing altogether.
Indeed, when Singapore – a framework from which India’s DPDPA draws considerable inspiration – built a regulatory scaffold for AI development, it made two complementary moves: amending its copyright legislation[7] to permit computational data analysis and issuing much clearer guidance on legitimate-purpose exceptions under its data protection framework despite monetisation paywalls[8].
India has made no equivalent legislative move on either front. It is against this backdrop that ANI v. Open AI offers a much-needed boost for India’s AI ecosystem, by reasoning what “research” means and how a research-linked exception ought to be tested. In doing so, it provides a principled analytical framework that may guide the future interpretation of the DPDPA’s research exemption.
Updating “Research”
Like the DPDPA, the Copyright Act does not define “research”.
The Delhi High Court, drawing on dictionary meaning and prior case laws, held that the process of training LLMs, through extraction and reorganisation of stored data to generate statistical output, qualifies as research.[9] More significantly, the Court rejected the assumption that “research” is confined to human intellectual activity, and extended the ambit of research to also cover machine learning, provided it is carried out at the behest, and for the benefit, of humans.[10] In doing so, the Court accepted that the commercial deployment of AI does not, by itself, strip the underlying training activity of its research character.[11]
If AI-driven research undertaken at human behest and for human benefit satisfies the copyright exception, there is little principled reason for a privacy exception, worded in substantially wider terms, to be read any more restrictively. The parallel to Section 17(2)(b) of the DPDPA appears too close to be coincidental. Both provisions bundle a stated purpose, research, together with a limiting condition on how that purpose may be pursued. That said, the nature of the limitations differs. Section 17(2)(b) requires that the data not be used to take any decision specific to a Data Principal, which is a safeguard against individual profiling and decision-making. However, AI training is likely to satisfy the DPDPA’s research exception more readily than it cleared the copyright test.
The Three-Fold Test
Having found that the purpose test was satisfied, the Court set out three factors relevant for assessing fairness:
- Purpose Limitation: Whether research is limited to the stated purpose of research (i.e., training the LLM) rather than some other business or functional end. This closely resembles the purpose limitation in Section 17(2) and entries (b) and (c) of the Second Schedule of the DPDP Rules.
- Fairness: Whether the use prejudices the legitimate interests of the rights holder (or the data principal) and serves broader public interest. Although the DPDPA does not adopt an express “fairness” inquiry, its safeguards relating to minimisation, accuracy, and reasonable security reflect a comparable concern with limiting prejudice to data principals.
- Public Interest: Whether the use serves the broader public interest, including scientific research, technological innovation, and the dissemination of knowledge.[12] The Court found that LLM training advances research, education, accessibility, and the development of AI more broadly.[13] This public interest is strengthened further where the training also improves products and services made available to users.
Zoomed out far enough, the Court’s balancing exercise bears some resemblance to the role played by legitimate interests under the EU GDPR, namely socially beneficial processing with appropriate safeguards for affected rights holders.[14]
The Road Ahead
For AI developers building or fine-tuning models in India, ANI v. Open AI is unlikely to be the last word on either copyright or privacy, but it provides the clearest judicial guidance yet on how Indian courts may approach AI model training as “research”. While the judgement does not finally resolve how the DPDPA will apply to AI training, what it does is remove a significant conceptual obstacle by recognising that commercial AI model training is capable of constituting research.
Whether AI developers can ultimately rely on the DPDPA’s research exemption will still depend on satisfying its statutory safeguards independently, but ANI v. Open AI provides a persuasive analytical foundation for that enquiry, and is likely to influence how future courts or the Data Protection Board approach it.
[1] ANI Media Pvt. Ltd. v. Open AI OpCo LLC, I.A. 45300/2024 in CS(COMM) 1028/2024, Delhi High Court, judgment dated July 24, 2026.
[2] Bartz v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal. filed August 19, 2024).
[3] Kadrey v. Meta Platforms, Inc., Case No. 3:23-cv-03417 (N.D. Cal. filed July 7, 2023).
[5] The DPDPA involves an element of causation, and only excludes consent for data which has been caused to be made public by the data principal, or under applicable law.
[6]DPDPA, Section 17(2)(b): the provisions of the Act (other than Sections 5 and 8) shall not apply to processing of personal data “necessary for research, archiving or statistical purposes if the personal data is not to be used to take any decision specific to a Data Principal and such processing is carried on in accordance with such standards as may be prescribed.”
[7]Singapore’s Copyright Act 2021 was amended to include Section 244, permitting text and data mining for AI training, alongside amendments to the Personal Data Protection Act 2012 introducing a ‘legitimate interests’ exception and clarifying the treatment of research and business improvement purposes.
[8] Paragraph 3.4 of the Advisory Guidelines on the Use of Personal Data in Generative AI, accessible at 143cb9d4-532e-4cca-9a77-bcc0415ca294.pdf.
[9] ANI v. Open AI, at paras 5, 202–219.
[10]ANI v. Open AI, at paras 212–219.
[11] ANI v. Open AI, at paras 173-180.
[12]ANI v. Open AI, at paras 236–240.
[13]ANI v. Open AI, at paras 241–254.
[14]General Data Protection Regulation (EU) 2016/679, Article 6(1)(f) and Recital 47, which requires a balancing of the controller’s legitimate interests against the interests and fundamental rights of the data subject.