The Collision Point
India's Copyright Act, 1957 — drafted before the internet, before software, and generations before generative AI — is now at the centre of the country's most consequential tech-policy debate of 2026. Two developments have collided to force a reckoning. First, the Delhi High Court's reserved verdict in ANI Media Pvt Ltd v OpenAI (Justice Amit Bansal), India's first AI copyright case, which concluded arguments in March 2026 after more than a year of hearings. Second, the Department for Promotion of Industry and Internal Trade's (DPIIT) December 2025 working paper proposing a mandatory licensing regime for AI training data — an approach more structurally ambitious than anything attempted by the EU, US, or UK. Together, they mark India's most serious attempt to answer a question no major jurisdiction has cleanly resolved: what do AI developers owe creators whose work trains their models?
What DPIIT Is Proposing
The working paper, released December 8, 2025 and opened for public consultation through February 6, 2026, proposes a blanket compulsory licence backed by a new collecting body: the Copyright Royalties Collective for AI Training (CRCAT). Under the model, AI developers receive a legal right to train on any "lawfully accessed" copyrighted work — no individual permission negotiations required. In exchange, they pay a royalty calculated as a flat percentage of the AI system's total global gross revenue, overseen by a government-appointed Rate-Setting Committee that reviews rates every three years. Critically, no upfront payment applies: royalties only trigger upon commercialisation, meaning pre-revenue startups face no immediate burden. A Welfare Fund would hold royalties for creators without formal sector organisation.
Before criticising the model, it is worth steelmanning it. The strongest argument for compulsory licensing here is a market-failure argument. When one side of a negotiation controls an AI system trained on billions of documents and the other side consists of individual journalists, photographers, and publishers in a developing economy, voluntary frameworks reliably produce near-zero compensation. Opt-out mechanisms, as the EU's experience demonstrates, are burdensome and underused. India's creative and publishing industry cannot individually negotiate against OpenAI or Google DeepMind. A state-backed collective mechanism at least guarantees compensation exists rather than making it contingent on a creator's ability to litigate for a decade.
Where the Model Breaks
The design has significant gaps that could undermine both of the goals it claims to serve.
Rate-setting without creator voices. The Rate-Setting Committee draws from government appointees and industry associations. Unregistered individual authors, freelance journalists, and small publishers — who together likely account for the majority of India's original creative output consumed in AI training — have no formal representation. The DPIIT model does not specify what share of collected royalties flows to specific creator categories versus administrative overhead, leaving distribution methodology opaque. The result is a mechanism that may collect royalties from AI companies without reliably delivering them to the people whose work was used.
The undefined commercialisation trigger. Royalties activate at "commercialisation," a term the working paper leaves operationally undefined. This is particularly problematic because dominant AI companies often operate at substantial losses funded by venture capital. An AI system charging subscription fees while reporting net losses could credibly contest whether it has "commercialised" in any legally meaningful sense. Without a revenue-based definition — not profit-based — enforcement becomes adversarial, costly, and unpredictable.
The paywall loop. The compulsory licence applies only to "lawfully accessed" content, excluding circumvented paywalls. But an AI trained on legitimately obtained premium journalism can reproduce paywalled insights freely to users, destroying the subscription revenue model that funds professional reporting. The working paper does not address this downstream harm — which is, arguably, more economically damaging to publishers than the training itself.
Competitive asymmetry with the EU and US. The EU AI Act allows a text and data mining (TDM) exception for research and non-commercial use, with a rights-holder opt-out for commercial training. The US operates under case-by-case fair use litigation. India's retroactively applied mandatory licensing — without any research exemption — could impose costs on domestic AI startups that their US and EU counterparts do not bear, precisely as India tries to establish itself as a tier-one AI hub.
The Legal Backdrop
Running in parallel, ANI Media v OpenAI has become a proxy for all of these unresolved questions. ANI, a major Indian news agency, alleges OpenAI trained ChatGPT on its copyrighted news archive without permission or payment. OpenAI contends the use qualifies as fair dealing under Section 52 of the Copyright Act, 1957.
The challenge for OpenAI is structural. India's copyright framework uses an exhaustive closed list of permitted acts — research, private study, criticism, review, news reporting — not the flexible purpose-based analysis available under US fair use doctrine. AI training does not appear on the list. Court-appointed technical advisors reportedly disagreed on key questions, meaning Justice Amit Bansal must establish new legal precedent rather than apply settled doctrine. With India representing approximately 100 million weekly ChatGPT users and OpenAI committed to major Indian data centre infrastructure via a Tata Group partnership, the commercial stakes of an adverse ruling are considerable.
The Proportionate Path Forward
India's instinct to move beyond ad hoc litigation toward a legislative framework is correct. A judge-made rule in ANI — in either direction — risks unintended consequences for the entire sector. But the DPIIT model needs targeted revision before it becomes law.
A narrow TDM exception for non-commercial research and education, mirroring the EU's approach, would protect India's academic AI sector without undermining creator compensation for commercial deployments. The Rate-Setting Committee should include independent creator representatives elected by their sector bodies, not government-appointed surrogates. "Commercialisation" must be defined by reference to revenue, not profit, with a de minimis threshold for startups below a specified floor. And DPIIT's forthcoming Part II of the working paper — addressing copyright in AI outputs — should address the paywall circumvention gap explicitly.
India has the rare opportunity to design an AI copyright regime from scratch rather than inheriting a framework ill-suited to the technology. Getting this right matters not just for Indian creators and developers, but for dozens of Global South jurisdictions that will calibrate their own rules based on what India's experiment produces.