The Department for Promotion of Industry and Internal Trade’s (DPIIT) working paper on regulating use of copyrighted works in AI training comes at a time when the issue can no longer be deferred. Litigation around AI training has already surfaced in India and is well underway in other jurisdictions. Given that generative models depend on ingesting large volumes of copyrighted material, uncertainty around what constitutes lawful training use has become untenable. In that sense, the proposal for a mandatory blanket licence is not a speculative regulatory exercise but a response to a problem that has already emerged and is likely to intensify.
The paper’s central move is to reject both polar positions: on the one hand, broad free use or text-and-data-mining (TDM) exceptions that risk hollowing out incentives for human creators; on the other, case-by-case licensing that is structurally incompatible with the scale and architecture of AI systems. Training data sets today cannot be disaggregated into neatly identifiable units linked to individual rights holders. Any framework built on granular permissions would thus be slow, fragmented, and ultimately unworkable. More significantly, such a system would reward those with the resources to negotiate expensive licensing arrangements. Large AI firms could internalise such costs, unlike start-ups and smaller developers. It would create disproportionate compliance burdens and a market dynamic that inadvertently favours incumbents.
Hybrid Rationale
This offers the rationale for the DPIIT’s hybrid approach of a statutory blanket licence for all lawfully accessed works, paired with a remuneration right activated only upon commercialisation. The immediate benefit is legal certainty. AI developers would train models without navigating thousands of individual permissions, while rights holders would have a predictable pathway to compensation. The absence of upfront fees and a linkage to commercial deployment rather than experimentation also reduces early-stage barriers for smaller firms. In a domain where the alternative is prolonged litigation or reliance on legally ambiguous practices, this clarity is not trivial.
Challenges in Implementation
A related issue the paper raises implicitly is retrospective applicability. In principle, bringing past training activities within the ambit of the new framework has a certain logic as many developers have trained on copyrighted material without clear legal cover, and a uniform regime would place all actors on one regulatory footing. But implementing this retrospectively may produce significant challenges. Determining the extent of past training, assessing if models can meaningfully be unwound, and ensuring retrospective royalty liabilities are not arbitrary or prohibitive will require careful calibration.
The mandatory nature of the licence also raises legitimate concerns. Removing the ability of rights holders to refuse inclusion of their works eliminates the element of individual consent central to copyright as a property right. Some analysts have questioned if this effectively converts private creative output into a compulsory public-use resource, even if compensated—a serious constitutional issue that can’t be sidestepped. Also, the administrative feasibility of building a central royalty-collection and distribution entity is uncertain. The challenge of attributing inputs is unresolved, and the risk of arbitrary or distortive rate-setting by a government-appointed body is real.
Industry responses have understandably diverged. Nasscom’s preference for a broad TDM exception with opt-outs reflects a belief that compulsory royalties function as a form of innovation tax. Content creators see the proposal as a long-overdue mechanism to ensure their works are not absorbed into AI systems without compensation. What matters is how the regulatory architecture balances these competing claims.
By placing the paper in the public domain and inviting comments, the DPIIT has allowed room for harmonising conflicting concerns rather than treating the majority view of the committee as the final answer. The consultative stage will be decisive. Royalty-setting formulas, governance and accountability mechanisms for the centralised entity, data-traceability expectations, and compliance thresholds will determine whether the system functions as intended or just adds to regulatory friction.
What is clear is that there’s a need for a coherent legal framework. Thus, the blanket licence model is best understood not as an ideal solution but as a pragmatic response to the structural characteristics of AI training. Its success will depend on how effectively the eventual legislation translates this logic into implementable institutions.