Fair Dealing at the Frontier: Generative Artificial Intelligence, Section 52(1)(a) of the Copyright Act, 1957, and the Case for a Calibrated Text and Data Mining Regime in India

Abstract

The training of generative artificial intelligence systems on copyright-protected material has become the defining doctrinal controversy of contemporary intellectual property law. Large language models are built by copying, tokenising and statistically encoding vast corpora of protected expression, an activity that engages the reproduction right in Section 14 of the Copyright Act, 1957, yet fits uneasily within any exception the Indian legislature has enacted. This paper examines how Indian copyright law has begun to answer that question and argues that the answer presently emerging is doctrinally fragile and institutionally misplaced. The analysis proceeds from the Delhi High Court’s decision of 24 July 2026 in ANI Media Pvt. Ltd. v. Open AI OpCo LLC, the first substantive judicial engagement in India with model training, in which the Court held on a prima facie view that the storage of scraped news content for training falls within the fair dealing exception for private or personal use including research under Section 52(1)(a)(i), and refused an interim injunction. The paper reconstructs the two-limb purpose-and-fairness test applied in that ruling and evaluates it against the closed-list architecture of Section 52, the three-step test in Article 9(2) of the Berne Convention and Article 13 of TRIPS, and the Indian jurisprudence on fair dealing running from Civic Chandran through the Delhi University photocopying litigation.

It then situates the Indian position within a comparative field that has moved sharply in the opposite direction: the Munich Regional Court in GEMA v. OpenAI treating memorisation within model parameters as an act of reproduction outside the European text and data mining exceptions, the United Kingdom abandoning its proposed opt-out exception in March 2026, and the United States generating contradictory first-instance fair use outcomes with no appellate resolution. Against that background the paper assesses the Department for Promotion of Industry and Internal Trade Working Paper of December 2025, which proposes a statutory blanket licence administered through a Copyright Royalties Collective for AI Training. The paper concludes that neither an expansive judicial reading of Section 52(1)(a) nor an unconditional statutory licence is defensible, and proposes a calibrated legislative model built on a lawful-access-conditioned text and data mining exception, an enforceable machine-readable reservation of rights, mandatory training-data disclosure, and a separately governed remuneration mechanism for commercially deployed models.