Published on: 11th June 2026
Authored by: H.Priya
Saveetha School of Law, Chennai
INTRODUCTION
The rise of generative artificial intelligence has upended one of the oldest bargains in intellectual property law: that creators control how their work is used, and that such control is the engine of creative incentive. Large language models, image generators, and multimodal AI systems are trained on vast corpora of human-authored text, music, imagery, and audiovisual material — often scraped from the internet or assembled through intermediaries, frequently without the knowledge or consent of the original creators. The scale of this consumption is unprecedented. Where a traditional broadcaster might use thousands of copyrighted songs in a year, a single AI training run may invest billions of tokens drawn from millions of protected works.
This has produced a global legal reckoning. Litigation has erupted across the United States, Europe, and Asia as authors, publishers, and news agencies challenge AI developers for what they describe as systematic copyright infringement. At the same time, AI companies argue that training on publicly available data constitutes transformative use or falls within fair dealing exceptions, and that restricting such use would chill technological innovation at a critical moment of global competition.
India has now entered this debate with a proposal that is both pragmatic and ambitious. In December 2025, the Department for Promotion of Industry and Internal Trade (DPIIT) released Part I of a Working Paper recommending a mandatory blanket licensing regime for AI training on copyrighted content. It is among the first national-level proposals anywhere in the world to articulate a structured, rights-compensating framework for this purpose — and it deserves careful examination.
THE BLANKET LICENSING MODEL: WHAT IT IS AND HOW IT APPLIES TO AI
Blanket licensing is not a novel legal concept. It has been the backbone of the performing rights industry for over a century. The basic idea is straightforward: rather than requiring a user to negotiate individual licenses with each rights holder whose work they wish to use, a collective management organization (CMO) aggregates rights from large numbers of creators and issues a single comprehensive license covering the entire repertoire. The licensee pays a fixed or formula-based fee; the CMO distributes royalties to members based on usage data. In the United States, ASCAP and BMI have operated this model for music performance rights since the early twentieth century, surviving antitrust scrutiny largely because courts recognized — as the Supreme Court made clear in BMI v. CBS (1979) — that the blanket license creates an entirely new product: instant, frictionless, legally secure access to an enormous catalogue that could not otherwise be assembled through individual negotiation at any viable cost.
The DPIIT proposal adapts this logic to the AI context. Its central recommendation, reflecting the majority view of an eight-member expert committee, is that India should adopt a mandatory blanket licence allowing AI developers to use any copyrighted work to which they have lawful access — purchased books, subscribed databases, authorised online content, or publicly available material — for training AI systems. The compulsory nature of the licence is its defining feature: rights holders cannot opt out of having their works used for AI training. In exchange, they receive a statutory remuneration right, administered by a government-designated “umbrella organisation” formed by rights holders themselves. AI developers, meanwhile, gain a single-window compliance mechanism that eliminates the need for thousands of individual negotiations and the legal uncertainty that currently surrounds every training dataset.
Under Indian copyright law as it currently stands, no such framework exists. The Copyright Act of 1957 (as amended) does not address AI training. The existing text and data mining exception under Section 52 is narrow and contested in its application to commercial AI development. The result has been a legal vacuum in which developers either proceed under assumptions about fair dealing that may not hold, or they operate in legal grey zones that expose them to significant litigation risk. The DPIIT proposal acknowledges this plainly, noting that prolonged negotiations and high transactional costs under the status quo could significantly hinder innovation, particularly for start-ups and MSMEs.
LEGAL AND POLICY IMPLICATIONS: BENEFITS, TENSIONS, AND UNRESOLVED QUESTIONS
The proposal’s primary virtue is clarity. By establishing a statutory licensing regime with judicial oversight of royalty rates and a centralized collection mechanism, it replaces legal ambiguity with a defined compliance pathway. For AI developers — especially smaller companies and academic researchers who cannot afford the legal overhead of negotiating with individual rights holders — this is genuinely transformative. It also, crucially, guarantees creators a revenue stream from AI training, something that current fair dealing arguments under litigation would deny them entirely if courts side with developers.
There are, however, significant tensions embedded in the proposal that will require careful resolution in the final framework.
The most philosophically significant is the compulsory nature of the licence. Unlike voluntary collective licensing, where rights holders choose to affiliate with a CMO and retain the right to withdraw, mandatory blanket licensing overrides the creator’s individual control entirely. An author who objects on moral or commercial grounds to having their novel used to train a generative system cannot refuse. This raises genuine questions about moral rights — a concept with constitutional resonance in the Indian context — and about the adequacy of monetary compensation as a substitute for control. The Berne Convention’s three-step test, which governs permissible limitations on copyright, requires that exceptions apply only in certain special cases, do not conflict with normal exploitation of the work, and do not unreasonably prejudice the legitimate interests of the author. Whether a mandatory blanket licence for commercial AI training survives this test will be a central question in international IP forums.
There is also the difficult problem of royalty rate-setting. The proposal envisions judicial supervision of rates to ensure they are fair, reasonable, and non-discriminatory — language borrowed directly from the ASCAP consent decree framework. But determining what a fair royalty looks like in this context is genuinely hard. Unlike music performance, where usage can be tracked (which song was played, how many times, on which platform), the contribution of any individual work to the emergent capabilities of a trained model is statistically diffuse and practically unmeasurable. How do you price the contribution of a single novel to a language model trained on a hundred billion tokens? The umbrella organisation and the courts will need economic methodologies that do not yet exist.
A further complication concerns the scope of “lawful access.” The proposal limits the blanket licence to works that developers have accessed lawfully. But the boundaries of that concept are contested. Is scraping publicly available web content lawful access? What about cached material, user-uploaded content on platforms, or datasets assembled through intermediaries whose own acquisition methods are opaque? Without precise statutory definitions, the lawful access threshold may generate as much litigation as it resolves.
Finally, there is the question of output protection — the risk that AI systems trained on copyrighted works reproduce substantial portions of that content in their outputs, effectively distributing the protected work without authorization. The DPIIT working paper notes that mechanisms may be required to prevent such reproduction, but leaves the design of those mechanisms to future consultation. This is a significant gap, because training-phase and output-phase risks are distinct and require different regulatory tools.
A COMPARATIVE GLANCE: HOW OTHER JURISDICTIONS ARE RESPONDING
India is not developing its approach in isolation. The global landscape reflects a spectrum of responses, none of which has yet produced a settled framework.
The European Union’s AI Act and the Copyright in the Digital Single Market Directive together establish a text and data mining (TDM) exception that permits mining for research purposes and, more broadly, for commercial uses unless rights holders opt out through machine-readable reservations. This opt-out model preserves creator agency in principle, but its practical effectiveness depends on whether rights holders can implement reservations at scale — a significant technical and organizational challenge for individual authors. The EU framework has been criticized by both sides: publishers argue the opt-out is unworkable, while AI companies argue that legitimate reservations will create patchwork datasets that disadvantage European AI development.
The United States has no AI-specific copyright legislation. The question of whether training on copyrighted works constitutes fair use is being litigated in multiple federal courts simultaneously — the New York Times v. OpenAI case being the most prominent — and the outcomes will likely diverge across circuits before the Supreme Court provides a definitive answer. The U.S. Copyright Office has been conducting a study on AI and copyright and has signaled that existing law may be inadequate, but legislative action remains distant.
Japan has arguably gone furthest in the pro-development direction, having clarified through administrative guidance that its copyright law’s TDM exception is broad enough to permit AI training on copyrighted works without compensation or opt-out rights, provided the use is for information analysis rather than direct reproduction. This approach has been welcomed by AI developers but criticized by creators’ organizations.
India’s proposed mandatory blanket licence sits in a different position from all of these. Unlike the EU’s opt-out model, it provides no individual withdrawal right. Unlike the U.S. litigated approach, it offers legal certainty up front. And unlike Japan’s broad exception, it guarantees compensation. In this sense, it attempts a genuine third path: structured, compensated, compulsory access.
FUTURE OUTLOOK: SHAPING IP LAW AND AI INNOVATION
If the DPIIT framework is enacted following stakeholder consultation, its effects will extend well beyond Indian borders. India is a significant AI development hub, home to a large and growing ecosystem of AI companies, a massive English-language content industry, and a population whose creative output flows into global datasets. A statutory licensing framework that legitimizes and compensates AI training could attract AI development investment, provide a template for developing nations navigating similar tensions, and create pressure on other jurisdictions to move from litigation-based uncertainty toward negotiated regulatory solutions.
For IP law more broadly, the proposal represents a significant doctrinal evolution: the extension of collective licensing from use-by-use consumption (a broadcast playing a song) to capability-building consumption (a model learning from a corpus). The distinction matters because the traditional justification for blanket licensing — eliminating transaction costs for uses that would otherwise be individually negotiable — applies with even greater force to AI training, where individual negotiation is not merely costly but functionally impossible. Viewed this way, the DPIIT proposal is less a radical departure than a logical extension of principles that courts and legislatures have repeatedly endorsed in the performing rights context.
The unresolved questions — royalty rate methodology, output protections, the three-step test, the scope of lawful access — will require the kind of deliberate, technically informed policy-making that India’s copyright institutions have not always demonstrated. The Copyright Office will need to develop expertise in AI systems. The umbrella organisation, once established, must have the administrative capacity and independence to distribute royalties fairly in the absence of granular usage data. And the courts, when called upon to set rates, will need economic frameworks purpose-built for this task.
None of this is simple. But the direction is right. The alternative i.e., continuing to permit AI training on copyrighted works without compensation while creators absorb the economic disruption of AI-generated competition — is neither legally sustainable nor politically stable. India’s proposal, for all its unresolved edges, offers a serious attempt to make the AI era work for both innovation and the creative economy that AI depends upon. How the stakeholder consultation shapes the final framework will be one of the most consequential IP policy decisions of this decade.
REFERENCES:
- Clark, G. A. (1979). Blanket licensing: The clash between copyright protection and the Sherman Act. Notre Dame Law., 55, 729.
- Bayat, M. An Analysis of the Blanket License System in Copyright Law: A Comparative Study and the Feasibility of Its Implementation in the Iranian Legal System.
- Ayushi Shukla (2025). Centre Proposes Mandatory Blanket Licence for AI Training on Copyrighted Works, Royalties for Creators, Live Law. https://www.livelaw.in/ipr/centre-mandatory-license-for-ai-training-ai-training-copyrighted-works-royalties-for-creators-513425
- Rahul Dahiya (2025) Mandatory Blanket Licence For AI Training? Understanding DPIIT’s Landmark Proposal On Copyright And GenAI, Mondaq. https://www.mondaq.com/india/licensing-syndication/1719238/mandatory-blanket-licence-for-ai-training-understanding-dpiits-landmark-proposal-on-copyright-and-genai
