Who Pays When AI Learns?

The Department for Promotion of Industry and Internal Trade ("DPIIT") on 08.12.2025 published Part I of the working paper by the Committee on Generative Artificial Intelligence and Copyright ("the Committee") – focused on the "One Nation One License One Payment" Model. The Committee was tasked with answering the following questions:

1. How does Generative AI ("AI") interact with copyright law in India?

2. Whether the current legal framework is adequate to address challenges emerging from these interactions? Are any amendments needed?

The Committee therefore identified two major areas of the intersection of AI and copyright law: (a) Use of copyrighted works as training data, usually without permission—dealt with by the present working paper, and; (b) copyright in AI-generated outputs, which will be addressed in Part II of the working paper. Interestingly, in Part I, the Committee does not draw a formal or doctrinal distinction between "training" and "pre-training." It deliberately treats them as part of a single, continuous process of AI training for copyright analysis.

India's Current AI-Copyright Status Quo

The paper discusses in some detail what India's approach to AI regulation is – first examining the India's National AI Strategy & IndiaAI Mission. The following are the critical points raised:

1. The working paper notes that the National Strategy on AI, released by NITI Aayog in 2018 under the theme #AIforAll, envisioned AI as a transformational tool across sectors and highlighted the need to strengthen India's IP ecosystem to support AI innovation.

2. Building on this, the IndiaAI Mission, approved in 2024 with an investment of over ₹10,300 crore, aims to create a unified, sovereign AI ecosystem, and as a result of such large-scale investment, thousands of GPUs are being deployed at subsidised rates, multiple ministries are being integrated into a national dataset-sharing framework, new labs and fellowships are being rolled out across educational institutions, and indigenous LLM efforts (such as those by Sarvam AI, Gnani AI, and others) are already being funded.

3. Parallelly, India's creative economy amounts to ~4.3% of GDP, employs millions (including informal folk sectors), and is projected to grow at a CAGR of 13% between 2021 and 2026 making it $80 billion USD economy.

4. Further, a report by the International Council for Research on International Economic Relations (ICRIER) estimates that India's informal music industry alone employs over 1.4 crore people.

The working paper, therefore, acknowledges the importance of a balanced approach, and looks to understand various models, and how stakeholders from both the AI developers' side and the creators' side have reacted to the same.

Stakeholder Consultations

The Committee consulted a variety of stakeholders (listed out in Annexure A), specifically:

• AI developers, tech companies

• Publishers, music & film industries

• News associations

• Collecting societies

• Legal experts

Unsurprisingly, the Committee received polarised views from the Tech Industry and Content Industry, as is represented in the table below:

What the Law Says

The Committee further analysed the existing legal framework, and how it addresses copyright concerns with respect to AI. Broadly, it observes that there is a definite vacuum that needs rectification, and subsequently analyses various models to address these issues.

1. The Copyright Act 1957 ("the Act") was designed for conventional modes of use and dissemination of works rather than large-scale computational ingestion of content. That is, the Act never envisioned the nature of "usage" of copyrighted materials that forms the basis of AI training.

2. Section 14 of the Act provides a comprehensive set of rights that are exclusive to the copyright owner with respect to such copyrighted works, which broadly include the right to storage, adaptation, copying, and reproduction. Section 52 of the Copyright Act provides for situations where explicit authorization is not provided, but the manner of work does not constitute infringement, referred to as "fair use" or "fair dealing." Fundamentally, if the rights under Section 14 are used by an individual not authorised to do so (whether under Section 14 or 52), it is considered as infringement of copyright.

3. The Committee observed that the training lifecycle breached Section 14 in various ways – as may be understood from the figure below:

4. The Committee further observed that there existed "fair use" provision under Section 52 either for Text and Data Mining ("TDM"), which is the method used by AI developers to procure massive datasets. Legislations across the world have recognised TDM as a valid exception to copyright laws, forming a part of the fair use/fair dealing doctrine. In India, however, Section 52(1)(a) is extremely narrow in scope, and allows fair dealing only for:

   a. private/personal use (including research)

   b. criticism/review

   c. reporting of current events

But AI training is commercial, large-scale, profit-oriented, so the exceptions so provided are unlikely to fit.

5. AI developers have argued, nonetheless, that AI training merely extracts "information" or "statistical patterns" rather than expressive content, which would put it outside the ambit of copyright infringement. However, the Working Paper treats such claims with caution, noting that courts and commentators increasingly recognise that the training process may internalise expressive elements of works, particularly where models are capable of memorising and reproducing substantial portions of training data.

6. The legal concerns regarding copying (see Table 2) cannot be cleanly enforced either, since temporary and intermediate copies are made during AI training, and Indian law has no guidance regarding enforcement. Developers are unclear if they are infringing rights, and rightsholders are unclear about their rights.

7. ***ANI v. OpenAI **--* Delhi HC is considering whether the storage and use of copyrighted news content for AI training amounts to infringement and whether fair dealing can apply.

Models Analysed

The Committee picked out five global models, and analysed their applicability to India.

Voluntary Licensing Model (USA)

Here, AI developers negotiate license with copyright holders before using copyrighted works for AI training. Scope of these licenses is determined through private contractual arrangements, with no statutory intervention beyond ordinary copyright enforcement.

The Committee noted that while this is extremely sound in theory, it poses a few problems:

1. Huge operating costs for licensing;

2. Asymmetric bargaining power between negotiating parties;

3. Impracticality of licensing millions of works individually.

*Quick note: *Apart from this, note that the licensing model of the USA is supplemented by an extremely flexible fair use model, which allows TDM so long as the work is "sufficiently transformative." Essentially, an extremely open, consumer centric market protects bargaining power, but at the same time, as long as the works are "transformed," the AI developer need not seek a separate license. Neither is India's market is as open or consumer-centric as the US, nor is the fair-use exceptions sufficiently wide, making this model impractical.

Collective and Extended Collective Licensing Model [CL/ECL] (Nordic Belt)

ECL allows a recognised collective management organisation (CMO)—the European equivalent of copyright societies—to license not only its members' works, but also non-members' works within the same category, unless those rightsholders explicitly opt out. Royalties are collected centrally and distributed to both members and non-members.

The Committee noted that while this is a rights-cum-economic centric approach, it requires an institutional set-up that is absent in India:

1. Mature CMOs/copyright societies;

2. Close to universal representation of rightsholders in such CMOs;

3. Transparent governance models;

4. CMOs are required to actively search for any author or owners who cannot be identified or

5. located at the time when the licence for use is granted.

*Quick note: *The Committee also observes that CL and ECL solve two major issues, i.e., consent and remuneration from the copyright owners' perspective, and reduce transaction costs from the AI Developer perspective. However, it determined that a major challenge is that the collective bargaining position may be leveraged for an excessive royalty rate. This feels like a rather poor argument to make, since one of the core objectives of the CMOs is to ensure collective bargaining, since there is already an imbalance between rightsholders and AI developers. However, the associated challenge of incumbent developers benefiting from this system very much holds good.

Statutory Licensing (India)

Here, instead of work-specific or owner-specific licenses, users are required to pay a pre-determined royalty calculated on government or court fixed rates, as compensation to the rightsholders. Statutory licensing also reduces transaction costs and creates a predictable environment for licensees of works. While this model takes away the power of the copyright owners to refuse licensing or negotiate a fee, it guarantees them fair compensation.

The Committee notes that this model reduces transaction costs, improves predictability, and ensures that copyright owners receive a fair royalty as a trade-off for the power of the copyright owners to refuse licensing or negotiate a fee. However, it does have a few issues in the context of Indian law:

1. Statutory licensing under Indian copyright law is not automatic. Existing statutory and compulsory licences (for example under Sections 31, 31A, and 31D of the Copyright Act) require active procedural engagement with regulatory authorities, since AI training is not a one-time act but a continuous, iterative activity, repeated procedural compliance is impractical;

2. Determining licensing costs assumes that the user can identify the categories of the work, the volume and nature of use can be reasonably specified, and compliance can be monitored against declared parameters. However, since AI training uses constantly evolving technologies, this would mean constant reporting and updates to the authorities—incompatible with the dynamic and automated nature of AI development.

Blanket TDM Exceptions

This refers to unrestricted, royalty-free use of copyrighted material for TDM and AI training purposes, without requiring individual licenses or permissions, so long as the user has lawful access to the material. This specific model was proposed by the tech community and was strongly rejected by the Committee, citing the following grounds:

1. Copyright is built on the idea that creators are entitled to control and economically exploit their works. A blanket TDM exception removes both elements at once. By allowing AI developers to freely use copyrighted works for training, the exception effectively converts creative labour into a free input for a commercial industry, and effectively reducing the incentive to create. Further, if creators cannot control use or remuneration, they have lost the power to exploit their works.

2. India's informal creative economy lacks formal representation, legal knowledge or even membership in CMOs/societies. A blanket TDM exception would expose these creators' works to large-scale commercial use without any possibility of compensation or redress, disproportionately harming those already least protected.

3. Once the data is used for training, it is irreversible. The consequences are functionally permanent, and the creators no longer have the capacity to exercise practical control over the exploitation of their works. This is exacerbated by the capacity of AI models to memorise or reproduce works.

TDM Exceptions with Opt-Out (EU)

Here, AI developers may use copyrighted material only if rightsholder(s) have not explicitly denied (opted-out of) such use. In practice, the European Union's** Copyright in the Digital Single Market Directive** ("CDSMD") allows two exceptions: (i) for scientific research (by universities, research bodies, etc.), or; (ii) for commercial purposes where the rightsholders have not opted out from data collection.

The Committee notes that this is quite burdensome to the copyright owner, and lacks mechanisms to prevent large-scale or permanent harms:

1. The Committee criticises the reversal of burden: instead of AI developers proving permission, creators must police the use of their own works. This is especially challenging for smaller creators, who do not know when or how their works are being scraped, nor do they have the technical means to signal opt-out effectively. As a result, opt-out protections would primarily benefit large publishers, not individual creators.

2. Further, opt-out cannot help control downstream processing, reuse of derived datasets or training on retained data – i.e., once content enters the training pipeline, opt-out operates too late. Metadata stripping and data aggregation further weaken the effectiveness of such reservations.

3. Further, if opt-outs are used excessively, then datasets may become smaller, less diverse or low quality, causing degrading model quality, increase bias and hallucinations, and ultimately undermine the stated innovation benefits of the TDM exception itself. Further, to check whether opt-outs have been followed, the developer should be subject to transparency requirements, which the Committee frames as being burdensome, and anti-innovation.

*Quick note: *The issues cited by the Committee are valid, but lack nuance. Please note:

1. The Committee treats this burden-shifting as a necessary feature of opt-out, rather than a design choice. It does not distinguish between:

   a. decentralised opt-out (robots.txt, metadata flags, contractual notices), and

   b. centralised, state-backed opt-out mechanisms with upstream enforcement.

As a result, the critique collapses all opt-out models into their weakest existing form, without considering whether institutional mediation between the two proposed models could rebalance burdens back onto developers.

2. The Committee implicitly assumes that all opt-out operates ex post, after data ingestion. It therefore treats irreversibility as a reason to reject opt-out entirely, rather than as an argument for ex-ante enforcement.

3. The threat of data shrinkage is inevitable if the Committee wishes to respect the consent of copyright owners. In fact, the premise of this paper itself is how copyright owners can better exercise their rights while allowing for AI development – which should make the primary consideration how developers/government can incentivise copyright owners to give permission (sufficient remuneration or granular consent to usage could prove to be good incentives).

4. Fundamentally, the Committee frames transparency almost exclusively as an innovation tax, rather than as a governance necessity, especially given the power imbalance and market dominance at play. Other highly innovative sectors (finance, telecom, pharmaceuticals) routinely accept transparency and auditability as baseline conditions. The Committee does not fully justify why AI training should be treated as uniquely fragile to compliance obligations.

Proposed Model

Summarily, the Committee recognises the following elements as necessary for regulating the domain:

1. Availability of lawfully accessed content for training to AI Developers as a matter of right without the need for negotiations, ensuring high-quality AI development with mitigated risk of AI bias

2. Fair compensation to copyright holders

3. Rate setting via a quick, transparent process

4. Preserving the ability of affected parties to challenge/ask for review of rates established before a judicial forum

5. Facilitates both big players, start-ups and individuals in the AI tech as well as creative industry, creating no barriers to entry

6. Upholds the basic principles of copyright and rewards human creativity

7. Mitigates risk of litigation and disputes.

Core Premise: Mandatory Access with Compensatory Protection

The foundational element of the Committee's proposed model is the separation of access from control. The model proceeds on the assumption that large-scale AI training requires unrestricted access to diverse datasets and that individual veto rights over training data would make such access unworkable. Accordingly, the Committee proposes that AI developers should be legally permitted to use all lawfully accessible copyrighted works for training purposes as a matter of right. In exchange, copyright owners are no longer granted the ability to refuse such use, but are instead protected through a guaranteed statutory right to remuneration.

Removal of Consent and Opt-Out Mechanisms

The Committee takes the view that consent-based or opt-out mechanisms would fragment training datasets, increase compliance complexity, and undermine innovation, particularly for startups and domestic AI developers. So, according to the model proposed, once a work is lawfully accessible, it is automatically available for AI training under the proposed framework, without any requirement for express permission or reservation of rights by the copyright owner. By replacing negotiated licences with a uniform statutory mechanism, the Committee looks to prevent a scenario where only well-resourced, established AI developers can afford to comply with complex licensing regimes**.**

The Committee also looks to reduce dataset fragmentation, bias and hallucination (which were argued to be consequences of an opt-out mechanism reducing the quantity and quality of datasets). That is, by ensuring universal access to lawfully available content, the model seeks to preserve dataset diversity and prevent AI systems from being trained on narrow or low-quality data.

Statutory Remuneration

Since the copyright owner has lost a great degree of control over the works, the Committee introduces a statutory remuneration right for copyright holders. This right operates independently of consent and ensures that creators receive compensation whenever their works are used for AI training. Unlike voluntary licensing, remuneration is not determined through private negotiation but through government-determined royalty rates, which are intended to reflect the scale and economic value of AI training activities.

Calculation: The model proposes the creation of a government-appointed rate-setting body to determine applicable royalty rates. This mechanism is expected to consider factors such as the category of works used, the scale of AI training, and the commercial benefits derived by AI developers.

Distribution: The model purports a "one licence, one payment" approach through a single, central, non-profit body—proposed as the Copyright Royalties Collective for AI Training ("CRCAT"). AI developers would make payments under a single blanket licence, while CRCAT would be responsible for collecting, managing, and distributing royalties to copyright societies, collective management organisations, and individual rightsholders, including non-members who register their works. Royalties are pooled by the CRCAT into different categories (A/V, text, music, etc.), and all rightsholders within a category share in the pool, regardless of whether their specific work can be shown to have been used in training.

*Membership: *The Committee proposes that the CRCAT should distribute royalties not only to members of recognised copyright societies, but also to non-member rightsholders, provided they register their works with the collective.

Reduced Transparency Obligations

To ensure that innovation is protected, the Committee proposed reduced transparency obligations on AI developers. Rather than requiring disclosure of training datasets, verification of individual permissions, or continuous reporting of data sources, compliance is largely limited to participation in the collective licensing system and payment of royalties. That is, AI developers are expected to:

• obtain a blanket licence,

• pay royalties to a central collective body, and

• comply with general regulatory conditions,

without having to reveal dataset composition or individual works used.

Critical Analysis

1. Through this model, it is clear that the Committee has taken a pro-innovation approach, appearing to be at the cost of the rights of the copyright holder. The objective of Section 14 of the Act is to allow the owner to control how their original work is used, reproduced, and monetized, and to prevent unauthorized use by others – forming the basis of "commercial exploitation." However, when the copyright owner has no control over if and how they wish to be remunerated, would the essence of the rights granted to them remain? The Committee does not convincingly establish why AI training warrants such a broad and permanent exception, particularly when the use is commercial and extractive rather than incidental or public-serving in nature.

Additional concerns emerge from the nature of the license itself. The DPIIT does not clarify if the license in question exists because AI Training constitutes a breach of copyright. Put simply, why should the license exist? Is AI training without a license a breach of S.14 of the Act? If so, what is the nature of such breach, and why can it not be enforced under the existing licensing regime?

2. On the matter of remuneration, the Committee at no point provides clarity on:

• How will copyright owners know whether their works have been used?

• How will the government/relevant authorities police the usage of these works and the payment of royalties?

These questions do not arise from the need of granularity. They are important to clarify, however, since the Committee claimed that transparency is an excessive burden, and yet requires the payment of royalties, which are to be distributed on the basis of datasets used. How will these be identified without transparency requirements is to be clarified. Furthermore, reduced transparency requirements have been proposed, not requiring any dataset disclosures, making the questions all the more pertinent.

3. A broader concern, actually, is that the Committee repeatedly frames creator-facing safeguards, like opt-outs, dataset transparency, or attribution of rights (moral rights) as "burdensome" or "anti-innovation," while treating developer convenience as a neutral baseline, even going so far as to critique collective bargaining as anti-innovative owing to the possibility of leveraging a higher royalty pay. By treating innovation arising from AI training as "uniquely fragile" to compliance obligations, the model effectively insulates developers from accountability mechanisms that are routine in other highly innovative sectors such as finance or telecommunications.

<!-- -->

4. The opt-out model was criticised for requiring awareness on part of the copyright owner, especially when the copyright owner is part of India's massive informal economy. The proposed hybrid model, however, does not clarify how the problem will be addressed post-implementation. That is, owners will still require knowledge regarding registration and the royalties they will have the right to. If awareness building is possible here, why could the opt-out model, which is more creator-centric, have been implemented along with such awareness programs?

Summarily, the proposed model prioritises innovation through compulsory access and pooled remuneration, but in doing so, it sacrifices creator control, transparency, and accountability in a manner that risks entrenching market power and hollowing out the core values of copyright law. While the model may facilitate rapid AI development, it does so by redefining copyright in ways that are insufficiently justified, weakly accountable, and likely to deepen existing power asymmetries.