eNews
India proposes licensing fee for AI companies that train on copyrighted content
A government proposal to let AI developers train on protected content, for a fee, reflects India’s preference for legislative intervention over courtroom improvisation
MUMBAI: India’s government is pushing for a mandatory licensing regime that would allow artificial intelligence companies to train their models on copyright-protected works, bypassing the thorny question of whether such training constitutes infringement in the first place. The proposal, contained in a working paper by the Department for Promotion of Industry and Internal Trade, would require AI developers to pay rightsholders through a government-designated collective body in exchange for access to lawfully obtained content.
The working paper, known by its acronym DPIIT, frames the mechanism as one of access and compensation rather than consent. That framing is telling. It sidesteps, rather than resolves, a foundational legal question: does AI training on copyrighted material actually engage exclusive rights under the Copyright Act, 1957?
The answer, in Indian law, is far from settled. Courts have consistently held that copyright protects expression, not information. Rulings such as Eastern Book Company v. D.B. Modak and Barbara Taylor Bradford v. Sahara Media underline that point. If AI training is understood as extracting statistical relationships rather than reproducing expression, there may be no infringement to speak of and nothing to licence in the first place.
The contrast with the United States is structural. American law begins with a broad conception of rights and carves out space for fair and transformative use. Recent cases including Bartz v. Anthropic and Kadrey v. Meta are playing out within that framework, where the central question is not whether copyright is engaged, but whether the use is sufficiently transformative to escape liability. India has no equivalent open-ended doctrine. The fair dealing exemptions under Section 52 of the Copyright Act operate within a closed, exhaustive list, and Indian courts have repeatedly declined to import the American approach.
That does not make Indian law purely restrictive. The 2012 amendment to the Copyright Act and a consistent line of judicial and legislative thinking have favoured access where markets fail. A compulsory licensing framework is already embedded in the Act. The DPIIT proposal fits squarely within that tradition.
The European model offers a third point of comparison. The text and data mining provisions under the EU’s Directive on Copyright in the Digital Single Market allow use subject to opt-out, but have struggled to generate meaningful compensation for rightsholders in practice.
The more pressing question is whether a statutory licensing layer is even necessary, given that voluntary markets for AI training data are already forming. Deals to licence curated datasets are multiplying, allowing rightsholders to set terms, price their content and refuse access if they choose. A blanket licence risks flattening that flexibility. It also raises practical difficulties: what counts as lawful access, how value is attributed across large mixed datasets, and whether premium content simply migrates to parallel private arrangements.
A pending case, ANI Media Pvt. Ltd. v. OpenAI, may shed some light on whether the ingestion and storage of content amounts to infringement under existing Indian doctrine. It will not, however, answer the broader policy question of how, or whether, AI training should be priced.
The working paper was authored by Kaushik Moitra, partner; Karnika Vallabh, counsel; and Nandini Tyagi, associate, all at law firm Bharucha & Partners.
India, on current trajectory, is unlikely to follow the American path of leaving the question to judicial improvisation. Legislative intervention of some kind looks probable. The harder task is designing a mechanism precise enough to handle doctrinal uncertainty without snuffing out the voluntary licensing markets that are, slowly, beginning to work.
