Open Access Research Article
White Black Legal – International Law Journal · ISSN 2581-8503
One Licence, Invisible Authors? Copyright Royalty Distribution under India’s Proposed AI-Training Regime
Read the Full Research Paper
Access the complete open-access article in PDF format. No login is required.
Abstract
ONE
LICENCE, INVISIBLE AUTHORS? COPYRIGHT ROYALTY DISTRIBUTION UNDER INDIA’S
PROPOSED AI-TRAINING REGIME
AUTHORED
BY - DHRUV GANDHI
Abstract
India’s emerging debate on generative artificial
intelligence and copyright has largely concentrated on whether training
requires permission and whether creators should receive compensation. A less
examined question is how compensation reaches individual creators after it
enters a central royalty pool. The Department for Promotion of Industry and
Internal Trade’s proposed blanket-licensing framework makes this question
particularly urgent. Simplifying access for developers does not necessarily
produce equitable payment for authors.
This paper examines the distributional consequences of
replacing individual permission with a centrally administered remuneration
right. Through doctrinal analysis of the Copyright Act 1957 and a limited
comparison with the European Union’s AI-training transparency requirements, it
identifies three risks: dependence on developer-controlled training records,
exclusion through registration-based payment systems, and concentration of
receipts among established rights holders. It argues that a training-data summary
adequate for public transparency may remain inadequate for allocating
royalties.
The paper proposes a two-tier accountability framework
combining public dataset summaries with confidential, auditable provenance
records. It also recommends accessible claims procedures, protection for
unclaimed royalties, representative collective governance, and a carefully
delimited author remuneration entitlement. The central argument is that a
blanket licence must be accompanied by a corresponding duty to preserve
distribution-relevant evidence. Without that duty, India could create an
efficient mechanism for purchasing access while leaving the creators whose
works finance the system economically invisible.
Keywords: Copyright, generative artificial intelligence, AI training, collective
licensing, royalty distribution and data provenance.
I.
Introduction
The copyright controversy surrounding generative
artificial intelligence usually begins with a permission question: may
developers reproduce protected works to train their models? India’s recent
policy discussion adds a payment question. DPIIT’s Working Paper on Generative
AI and Copyright proposes a hybrid model allowing developers to use lawfully
accessed protected content under a blanket licence, with royalties payable upon
commercialisation and administered through a central mechanism.[1]
However, collecting money and compensating creators
are different institutional tasks. A system may efficiently collect royalties
from developers while distributing them through incomplete records,
inaccessible claims procedures, or rules favouring large catalogues. The
unresolved question is therefore not simply whether payment is required, but
who can establish an entitlement to it.
Consider a hypothetical Hindi essayist whose work is
included in a training collection alongside English-language newspaper
archives. The developer pays the prescribed royalty. Yet the essayist does not
know that the work was used, does not belong to a collecting organisation, and
cannot locate it within a dataset summary. The payment system has operated, but
the author remains unpaid.
This paper examines that distributional gap. Its
argument is deliberately narrower than the general debate over AI authorship or
fair dealing: if legislation replaces individual permission with collective
remuneration, it must also preserve the evidence necessary to distribute that
remuneration.
The paper uses doctrinal analysis and institutional
comparison. It does not claim that the proposed regime is already law or that
any distribution method can measure a work’s exact contribution to model
performance. Instead, it proposes minimum safeguards for identifying eligible
claimants and allocating receipts transparently.
II. From Exclusive Rights to Collective Remuneration
The existing statutory starting point is reproduction.
Section 14 of the Copyright Act 1957 includes electronic storage within the
reproduction right for literary, dramatic and musical works, with corresponding
provisions for other categories. Training arrangements involving downloads and
retained copies therefore require analysis of the acts performed, applicable
permissions and statutory exceptions. The claim that a model ultimately learns
patterns does not, by itself, answer whether earlier copying was authorised.[2]
Section 52 contains specific exceptions, including
fair dealing for identified purposes and certain forms of transient or
incidental storage. These provisions should not be treated as an unrestricted
commercial text-and-data-mining exception. Equally, it would be premature to
declare every training operation infringing without examining its facts and the
defence invoked.[3]
Existing compulsory and statutory licences are
structured around particular circumstances or uses. Sections 31, 31A, 31B, 31C
and 31D do not themselves establish a general licence for generative-AI
training. A new blanket regime would therefore require a clear legislative
foundation rather than administrative reliance on an existing licensing label.[4]
The DPIIT proposal would change the relationship
between permission and remuneration. Its proposed central body, the Copyright
Royalties Collective for AI Training, would collect payments and allocate them
through participating rights-holder organisations. The proposed architecture
includes developer training-data disclosures and work-registration mechanisms
for royalty distribution.[5]
Such a framework can reduce negotiation costs.
Nevertheless, it also shifts practical control. An individual rights holder’s
ability to negotiate access would be replaced, within the licence’s scope, by a
claim against a collective payment system.
That substitution requires justification beyond the
promise that royalties will exist. The system must explain how a claimant
becomes visible, how ownership is verified, how use is established, and how
allocation decisions can be challenged. Otherwise, the administrative
simplicity enjoyed by developers may be purchased through evidentiary burdens
transferred to creators.
III. The Distributional Blind Spot
A. Copyright Owners Are Not Always Authors
An AI royalty scheme must distinguish the creator from
the person presently owning the relevant right. Section 17 begins with author
ownership but contains important exceptions. Assignments under sections 18 and
19 can further separate authorship from ownership. Payment to a publisher,
employer or assignee therefore does not necessarily produce payment to the
individual who created the work.[6]
This distinction matters when a policy is justified in
the language of supporting creativity. If its operative provisions compensate
only current owners, its benefits may legitimately reach owners but fail to
fulfil the broader promise made to authors.
The solution is not to disregard valid assignments. It
is to decide expressly whether the new regime creates an owner entitlement, an
author entitlement, or both. That choice should appear in legislation rather
than emerge accidentally from collective distribution rules.
B. Registration Can Become a Payment Barrier
A work database can assist ownership verification,
prevent duplicate claims and simplify payment. Yet registration-based
distribution may disadvantage creators who lack information, documentation or
institutional connections.
The distinction between copyright subsistence and
payment administration is crucial. The Berne Convention protects enjoyment and
exercise of copyright against formalities within its applicable framework. This
does not establish that every administrative step for collecting money is
prohibited. It does, however, require careful scrutiny where failure to enter a
database makes a statutory remuneration entitlement practically unusable.[7]
Registration should therefore facilitate payment, not
become the only route to proving a claim. A creator should be able to submit
publication records, contractual documents or other reliable evidence after
learning of inclusion.
The same principle requires multilingual interfaces
and offline assistance. These are proposed safeguards, not assumptions about
the present accessibility of any particular collecting organisation. Their
purpose is to prevent administrative design from deciding whose copyright has
economic value.
C. A Summary Is Not an Allocation Ledger
The proposed training-data disclosure mechanism
connects dataset reporting to royalty distribution. Available descriptions of
the proposal indicate that collective receipts would be divided among classes
of works according to disclosed use, followed by distribution under
participating organisations’ policies.[8]
That connection creates a difficult evidentiary
question. A summary identifying broad datasets may show that books, newspapers
or images were used. It may not identify which protected works entered the
training collection, whether duplicates were removed, or who owns the relevant
rights.
Allocation can consequently rest on assumptions that
creators cannot inspect. Counting files, pages or tokens may produce materially
different results. None should be presented as an objective measure of creative
value merely because it is computationally convenient.
The hypothetical essayist illustrates the problem. A
dataset description such as “Indian-language web content” supplies general
transparency but little assistance in proving an individual payment claim.
IV. What European Transparency Can and Cannot Teach
India
Article 53 of the European Union AI Act separates
copyright compliance from training-content transparency. Providers of
general-purpose AI models must maintain a copyright-compliance policy and
publish a sufficiently detailed summary of training content. Its specified
open-source exemptions concern certain documentation obligations; they do not
simply remove the copyright-policy and public-summary duties.[9]
This separation offers an important lesson: neither
payment nor disclosure automatically establishes lawful copying. A provider may
publish a summary while still facing questions about rights compliance.
However, the European provision should not be mistaken
for a ready-made royalty-allocation system. A public summary serves
transparency and oversight. A collective payment mechanism additionally needs
evidence connecting works, uses and eligible claimants.
India should therefore adopt a two-tier approach.
First, public disclosures should identify major
datasets, source categories, collection periods and broad classes of works.
These disclosures should help creators recognise possible inclusion without
exposing complete datasets.
Second, confidential records should preserve more
granular provenance information for verification by an authorised auditor or
adjudicator. Appropriate records could include source identifiers, acquisition
dates, available licence information, dataset versions and deduplication
methods.
The distinction protects legitimate confidentiality
while resisting a blanket claim of secrecy. Developers need not publish every
commercially sensitive record to the world. They should nevertheless preserve
enough information to justify allocations made from a statutory royalty pool.
V. A Distribution-First Legislative Model
A. Make Provenance a Condition of Licensing
The proposed licence should carry a prospective
record-preservation obligation. Developers relying on it should maintain
distribution-relevant records from collection through commercialisation.
The obligation should track what the developer can
reasonably document. It should not demand proof of each work’s precise causal
contribution to a model’s capabilities. Eligibility and use are more
administrable questions than the economic influence of an individual sentence
or image.
Nor should a training licence confer unrestricted
authority over later uses. Copyright Act provisions on technological protection
measures and rights-management information demonstrate that access, copying and
treatment of identifying information involve distinct statutory questions. The
new framework should expressly address how those provisions interact with
licensed training.[10]
A developer’s payment should not excuse failure to
keep records. Where records are materially deficient, the administrator should
be able to require corrective reporting and independent audit, subject to
proportionate procedures and appeal.
B. Separate Eligibility from Valuation
The system should first determine whether a protected
work was used within the licensed activity and who is entitled to claim. Only
then should it determine the payment amount.
Combining these stages risks making an unclear
valuation method a barrier to eligibility. A creator should not need to prove
that a work improved model performance before establishing that it entered the
relevant training collection.
Valuation rules should be published and periodically
reviewed. Possible methods include work-category pools, deduplicated use
measures and sampling where complete identification is disproportionate. Each
method involves normative choices.
For example, allocation by volume can favour long
works; allocation by catalogue size can favour large owners; allocation by
asserted market value can favour established markets. The legislation should
require explanation of these consequences instead of allowing technical metrics
to conceal policy decisions.
C. Create a Delimited Author Remuneration Right
The Copyright Act already protects non-waivable
royalty entitlements in specified circumstances involving literary and musical
works incorporated into films and sound recordings. Those protections do not
automatically establish an AI-training royalty right for every author. They
nevertheless provide a statutory example of preserving author remuneration
despite transfers of exploitation rights.[11]
A future AI provision could create a defined author
share that cannot be assigned away, while leaving the remaining owner
entitlement intact. Its scope would need to specify eligible works,
beneficiaries, contractual treatment and arrangements for heirs.
This is preferable to assuming that payment to a
corporate rights holder inevitably supports authors. It also avoids the
opposite error of ignoring ownership arrangements that financed production and
publication.
D. Protect Unclaimed Royalties
Unclaimed money should not automatically become an
additional payment to the largest registered catalogues. That approach would
convert information failures into advantages for already visible claimants.
Instead, the administrator should maintain a
separately accounted reserve, publish search and claim procedures, and permit
later applications within a clearly defined period. Rules should address
disputed ownership, duplicate claims and amounts too small for immediate
individual payment.
Any eventual use of residual funds should be
legislatively authorised and publicly reported. A collective should not
determine the destination of unclaimed statutory entitlements solely through
internal policy.
E. Build Representative Governance
India’s copyright-society provisions already address
administration, member control, tariff schemes and reporting. They provide a
starting point for governance safeguards, though an AI-wide collective would
require rules adapted to its broader role.[12]
The proposed institution should include meaningful
representation for individual authors, independent publishers and different
work categories. Representation should not depend exclusively on catalogue size
or receipts.
Annual reports should disclose collections,
administrative deductions, allocation methods, reserves and complaint outcomes.
External audits should examine distribution accuracy as well as financial
accounts.
A specialised claims process should offer reasoned
decisions and independent review. A claimant challenging an allocation should
not need to begin expensive infringement litigation merely to obtain an
explanation of collective accounting.
VI. Boundaries, Treaty Compatibility and
Counterarguments
A. Training Permission Must Remain Purpose-Limited
The scope of any licence should distinguish training
copies from later reproduction, communication or distribution. A training
payment should not become a defence to every infringing output.
Section 51 governs infringement, while section 57
protects specified author interests independently of ownership. These
provisions reinforce the need to avoid treating a training licence as
comprehensive immunity. Whether a particular output infringes or violates an
author’s special rights remains a separate, fact-sensitive question.[13]
The proposal should also distinguish protected
expression from unprotected material. Copyright payments should not attach
merely because a dataset contains information.
B. Remuneration Does Not Settle Treaty Compliance
A compulsory remuneration framework must address
international copyright obligations. Berne article 9(2) and TRIPS article 13
impose constraints on limitations affecting exclusive rights. Providing payment
may bear on prejudice to rights holders, but it does not automatically
establish compliance with every part of the applicable test.[14]
The legislature should therefore examine the defined
activity, eligible users, commercial scope and effect on normal exploitation. A
broad licence may require stronger justification than a narrowly bounded
research exception.
This paper does not conclude that the proposal
necessarily violates treaty obligations. It argues that administrative
efficiency and royalty collection are not substitutes for that analysis.\
C. The Compliance-Cost Objection
The strongest practical objection is cost. Detailed
records and audits can make training more expensive, particularly for small
developers.
The answer should be proportionality, not complete
exemption from accountability. Standard reporting formats, shared verification
infrastructure and risk-based audits could reduce burdens. Smaller developers
could receive simplified procedures while preserving basic source and dataset
records.
Sampling may be appropriate where item-level
identification is genuinely impracticable. However, its limits should be
disclosed, and affected claimants should retain an opportunity to supply
additional evidence.
D. The Creator-Autonomy Objection
Stakeholder criticism of the DPIIT proposal has also
questioned compulsory participation and whether revenue-linked payment
meaningfully compensates creators.[15]
Better distribution cannot fully answer the autonomy
objection. An author may oppose training even where payment is accurate. The
proposed safeguards therefore should not be presented as curing every concern
about mandatory licensing.
Their narrower justification is conditional: if
Parliament chooses a blanket regime, it should not combine reduced control over
access with weak access to remuneration. Legislative review should assess both
payment outcomes and the continuing necessity of compulsory participation.
VII. Conclusion
India’s proposed AI-training framework raises a
question that copyright debate too often postpones: what happens after
royalties are collected?
A central licence may simplify access, but it cannot
fairly compensate creators without reliable evidence, accessible claims
procedures and accountable distribution rules. Dataset summaries alone cannot
perform every function required of a payment system.
The appropriate response is a distribution-first
architecture. Public summaries should be paired with confidential provenance
records; eligibility should be separated from valuation; authors and owners
should be distinguished; and unclaimed money should remain protected.[16]
The central principle is straightforward: legislation
replacing individual permission with collective remuneration must preserve a
credible route from the licensed use to the person entitled to payment.
Otherwise, the system risks making copyright easier to purchase while making
its creators harder to see.
[1] Department for Promotion of Industry and Internal
Trade, Working Paper on Generative AI and Copyright, Part I: One Nation One
License One Payment—Balancing AI Innovation and Copyright (2025); Press
Information Bureau, ‘DPIIT Publishes First Part of Working Paper on
AI–Copyright Interface’
[2] Copyright Act 1957, ss 14(a)(i),
14(c)(i), 14(d)(i) and 14(e)(i).
[4] Copyright Act 1957, ss 31-31D.
[5] Ikigai Law, ‘Exploring the DPIIT’s
Working Paper on Generative AI and Copyright’ (discussion of the proposed
hybrid licensing model and CRCAT).
[6] Copyright Act 1957, ss 17-19.
[7] Berne Convention for the
Protection of Literary and Artistic Works (Paris Act, amended 28 September
1979), art 5(2).
[8] Ikigai Law, ‘Exploring the DPIIT’s
Working Paper on Generative AI and Copyright’ (discussion of the AI Training
Data Disclosure Form, work databases and royalty distribution).
[9] Regulation (EU) 2024/1689 laying
down harmonised rules on artificial intelligence, art 53(1)(c)-(d), (2) and
(7).
[10] Copyright Act 1957, ss 2(xa), 65A
and 65B.
[11] Copyright Act 1957, ss 18(1),
provisos concerning literary and musical works incorporated into cinematograph
films and sound recordings, and 19(9)-(10).
[12] Copyright Act 1957, ss 33-36.
[13] Copyright Act 1957, ss 51 and 57.
[14] Berne Convention, art 9(2);
Agreement on Trade-Related Aspects of Intellectual Property Rights, art 13;
World Trade Organization, WTO Analytical Index: TRIPS Agreement-Article 13.
[15] Software Freedom Law Center,
India, ‘Comments on DPIIT Working Paper on Generative AI and Copyright-Part 1’
(3 February 2026).
[16] Ikigai Law (n 8).
How to Cite This Article
DHRUV GANDHI, One Licence, Invisible Authors? Copyright Royalty Distribution under India’s Proposed AI-Training Regime., White Black Legal – International Law Journal, ISSN: 2581-8503, Vol. Volume 4, Issue VOLUME 4 ISSUE 2, October 2026, pp. 1093-1105. Available at: https://www.whiteblacklegal.co.in/details/one-licence-invisible-authors-copyright-royalty-distribution-under-indias-proposed-ai-training-regime
Author & Publication Record
Authors: DHRUV GANDHI
Registration ID: 107192 | Published Paper ID: WBL7192
Year: Oct- 2026 | Volume: 4 | Issue: 2
Approved ISSN: 2581-8503 | Country: Delhi, India
Page No.: 1093-1105
Full Text Preview
Open in New Tab
Copied