Open Access Research Article
White Black Legal – International Law Journal · ISSN 2581-8503
Expanding the Exception: Generative AI Training and the Architecture of Indian Copyright Law
Read the Full Research Paper
Access the complete open-access article in PDF format. No login is required.
Abstract
EXPANDING THE EXCEPTION: GENERATIVE
AI TRAINING AND THE ARCHITECTURE OF INDIAN COPYRIGHT LAW
AUTHORED
BY - MR. PRATYUSH PANDEY
Ph.D.
Scholar, Faculty of Law, University of Allahabad
CO-AUTHOR
- DR. ANSHUMAN MISHRA
Professor,
Faculty of Law, University of Allahabad
Abstract
On
24 July 2026 the Delhi High Court, in ANI
Media v OpenAI, declined to restrain the use of a news agency's reports in
training large language models, holding prima facie that such use falls within
section 52(1)(a)(i) of the Copyright Act, 1957. This article contends that,
although the refusal of interim relief was defensible on the record before the
Court, the interpretive route by which it was reached sits uneasily with the
closed-list architecture of Indian fair dealing. The Court's reading rests on
three propositions: that 'private' use extends to a closed corporate
environment, that machine training is 'research', and that commerciality is
immaterial at the threshold. Taken together, these propositions risk converting
a purpose-bound exception into something approaching open-ended fair use
without legislative sanction. Drawing on developments in the European Union,
the United States, the United Kingdom, Japan, Singapore and Australia, the
article demonstrates that every jurisdiction which has permitted commercial
training at scale has done so through a conditioned statutory
text-and-data-mining provision rather than by enlarging a private-use
exception. It concludes by proposing a tiered legislative framework that
distinguishes non-commercial research, remunerated commercial training and
retrieval-time reuse of news, with particular regard to the position of
Indian-language publishers.
Keywords:
fair dealing; section 52(1)(a); text and data mining; generative artificial
intelligence; news publishers; Indian languages.
I. Introduction
Indian
copyright law now has its first judicial answer to the question of generative
ingestion, and that answer favours the developer. In ANI Media v OpenAI, Bansal J refused an interim injunction, holding
prima facie that the storage of ANI's news reports for the purpose of training
the models underlying ChatGPT fell within section 52(1)(a)(i) of the Copyright
Act, 1957.¹ The order is interlocutory and is expressly stated to be without
prejudice to the final adjudication of the suit. It is, nonetheless, the only
Indian authority on the point, and it has already begun to inform licensing
negotiations between publishers and developers while ANI's appeal awaits
determination by a Division Bench.²
The
ruling arrives at a moment of evident divergence within the Indian state. Some
seven months earlier, a committee constituted by the Department for Promotion
of Industry and Internal Trade (DPIIT) had concluded that fair dealing is too
narrow and too uncertain to sustain AI training, and had recommended a
mandatory blanket licence coupled with a statutory right of remuneration.³ The
judiciary and the executive have thus proceeded from opposing premises: the
former treats the existing exception as already sufficient, while the latter
regards it as structurally incapable of the task and proposes new institutional
machinery in its place.
This
article addresses three questions. First, can a purpose-bound fair-dealing
provision, drafted with individual users and scholars in view, lawfully
accommodate industrial-scale commercial ingestion? Secondly, what did the Delhi
High Court's reasoning secure, and what did it concede, in answering that
question affirmatively? Thirdly, what may India learn from jurisdictions that
have legislated specifically for text and data mining, and from the litigation
now testing those provisions? It is argued that the Court reached a pragmatic
outcome by an interpretive route that India's closed-list design does not
comfortably support, and that the allocation of value generated by news and
Indian-language corpora is properly a matter for Parliament rather than for an
expansive reading of section 52(1)(a).
The
argument proceeds as follows. Part II sets out the statutory framework and the
Indian jurisprudence on fairness. Part III tests commercial ingestion against
the purpose and fairness limbs of section 52(1)(a). Part IV examines ANI Media v OpenAI and its subsequent
appellate history. Part V undertakes a comparative analysis of six
jurisdictions. Part VI considers the particular position of news and
Indian-language material, and Part VII advances a tiered proposal for reform.
Part VIII concludes.
II. The Statutory Framework
A. The Right Engaged
Ingestion
engages the reproduction right before any question of exception arises. Section
14(a)(i) confers upon the owner of a literary work the exclusive right to
reproduce it in any material form, a formulation that expressly includes
storage in any medium by electronic means.⁴ The ordinary sequence of dataset
construction, in which a website is crawled, its text extracted and cleaned,
duplicates removed and the resulting material retained in a training corpus,
produces stored copies at each stage. The statute draws no distinction between
transient and enduring storage, and under section 51 the defendant's motive is
immaterial to the question whether infringement has occurred.
B. The Available Exceptions
Section
52(1)(a) permits fair dealing with any work other than a computer programme for
three enumerated purposes: private or personal use, including research;
criticism or review; and the reporting of current events and current affairs.
An Explanation inserted in 2012 provides that the storage of a work in any
electronic medium for those purposes does not constitute infringement.⁵ Clauses
(b) and (c), introduced by the same amendment, exempt transient or incidental
storage in the technical process of electronic transmission, or for the purpose
of providing links, access or integration.⁶ These latter provisions were
evidently framed with intermediaries and caching in contemplation. A curated
corpus retained over the months that a training run may require is neither
transient nor incidental in any ordinary sense, with the consequence that
clause (a) must bear the entire weight of any defence.
Two
structural features of the provision merit emphasis. India follows the
closed-list model inherited from the United Kingdom, under which a dealing must
first be brought within an enumerated purpose before its fairness falls to be
assessed. Further, the Act contains no text-and-data-mining exception,
notwithstanding that the 2012 amendments presented a natural occasion on which
one might have been considered.
C. The Judicial Approach to Fairness
Indian
courts have converged upon a limited set of considerations in assessing
fairness. In Civic Chandran v Ammini Amma,
the Kerala High Court examined the quantity and value of the matter taken, the
purpose of the taking, and whether the later work was liable to compete with
the original.⁷ The Delhi High Court in Super
Cassettes v Hamar Television accepted that commercial use does not of
itself defeat a plea of fair dealing, while treating the commercial character
of the use as a weighty consideration.⁸ The Division Bench in the Rameshwari Photocopy litigation
subsequently construed the educational exception by reference to its purpose
rather than the profit motive of the defendant.⁹ Underlying each of these
decisions is the three-step test of the Berne Convention, which confines
exceptions to certain special cases that neither conflict with the normal
exploitation of the work nor unreasonably prejudice the legitimate interests of
the author.¹⁰
The
jurisprudence thus displays a consistent orientation. Indian courts ask what
the user was in fact doing, and whether the result competes with the work.
Neither inquiry was formulated with a machine reader in contemplation, and it
is to that difficulty that the article now turns.
III. Commercial Ingestion and the
Limits of Fair Dealing
A. The Purpose Limb
It
is at the threshold of purpose that commercial training is most vulnerable. The
phrase 'private or personal use' sits naturally with an individual who copies
for study, or with a small circle not addressed to the public, and the
conjunction of the two adjectives suggests that the clause is concerned with
the sphere in which a work is enjoyed. A developer's servers are, admittedly,
private in the sense of being closed to outsiders. The purpose of the copying,
however, is the construction of a product offered to the public at very
considerable scale. To treat the privacy of the process as though it were the
privacy of the use is, it is submitted, to conflate the location of the copy
with the end it serves.
The
term 'research' offers developers firmer ground. Unlike section 29(1) of the
United Kingdom's Copyright, Designs and Patents Act 1988, the Indian provision
does not confine research to non-commercial ends.¹¹ A court adopting an
updating construction may therefore characterise statistical learning as a
species of inquiry. The difficulty lies in the breadth of that
characterisation. If the development of a commercial model constitutes
research, then so does almost every product-development activity that involves
reading, and the exception ceases to be purpose-bound in any meaningful sense.
B. Commerciality and Extent
The
textual argument that commerciality is not disqualifying is a strong one. Where
Parliament intended to exclude commercial use it did so expressly, as in clause
(ad), which permits the copying of computer programmes only for non-commercial
personal use.¹² It does not follow, however, that commerciality is irrelevant;
it is simply relocated to the fairness inquiry, where it weighs heavily. So too
does the extent of the taking. Training reproduces whole works, and frequently
whole archives, rather than extracts. The customary response is that the use is
non-expressive, in that the model extracts statistical regularities rather than
the expression that readers value. That argument carries genuine force as a
description of what training produces, but it describes the output of the
process rather than the act of reproduction which the statute in fact
regulates.
C. The Market-Harm Paradox
The
most revealing difficulty concerns market harm. An unlicensed use is less
likely to be fair where a market exists for licensing it, and such a market
plainly now exists: news publishers are concluding arrangements with AI
developers, and ANI itself quoted OpenAI a licence fee of USD 7.5 million in
October 2024.¹³ At the interlocutory stage that quotation assisted OpenAI,
since it indicated that ANI's loss was capable of being compensated in damages.
At the fairness stage, however, the same fact ought logically to operate in the
opposite direction, for it evidences a market that unlicensed ingestion
displaces. A mode of analysis that permits a single item of evidence to defeat
the claimant twice over merits close scrutiny on appeal.
IV. ANI Media v OpenAI: A Case Study
A. The Dispute
ANI
instituted proceedings in November 2024, alleging two distinct wrongs: first,
that OpenAI had trained its models on ANI's reports without a licence; and
secondly, that ChatGPT reproduced ANI's content and attributed to it statements
that ANI had never published.¹⁴ Judgment on the application for interim relief
was reserved on 27 March 2026 and pronounced on 24 July 2026.
B. The Reasoning of the Single Judge
The
Court's reasoning may be traced through five stages. On jurisdiction, storage
on servers located in the United States was treated as the terminal link in a
chain of acts commencing in India, with the result that offshore training did
not oust the Court's jurisdiction. On the right engaged, electronic storage
during training was held to constitute reproduction under section 14(a)(i),
irrespective of whether the copies were subsequently deleted. On purpose, the
Court held that section 52(1)(a) requires two distinct inquiries, one as to
purpose and one as to fairness; it read 'private' as extending beyond
individuals to a closed group or company, read 'research' as encompassing
machine learning, and held that commerciality alone did not disqualify the defence.
On fairness, the Court found that the use was confined to training, that no
economic loss had been demonstrated, and that the public benefits of language
models were considerable. Finally, on the output claim, ANI's examples
post-dated the models' training cut-off dates, bore the hallmarks of retrieval
at the time of the query rather than memorisation, and were not substantially
similar to ANI's expression.¹⁵ In assessing the balance of convenience, the
Court further observed that ANI had never blocked OpenAI's crawlers and that
its own licence quotation demonstrated that any harm was compensable.
C. The Durable Contributions of the Judgment
Three
aspects of the judgment are likely to endure. The chain-of-acts approach to
jurisdiction forecloses the possibility of evading Indian law simply by
locating servers abroad. The insistence that storage amounts to reproduction
preserves the analytical integrity of the inquiry, declining the invitation to
treat training as though it involved no copying at all. And the separation of
training claims from claims founded on retrieval-based outputs reflects a
discipline that courts in other jurisdictions have found difficult to maintain.
D. Points of Vulnerability
The
judgment is, however, open to criticism on three grounds. As regards purpose,
the construction placed upon 'private' and 'research' performs work that the
closed list in section 52 was designed to prevent, for the reasons developed in
Part III. As regards the equities, the significance attached to ANI's failure
to block crawlers imports, by implication, an opt-out regime that Parliament
has never enacted. Copyright in India, consistently with the Berne Convention,
subsists without formalities.¹⁶ A rule under which silence is treated as
consent transfers the burden to every publisher, including smaller publishers
who may be unaware of the technical means by which crawlers are excluded.
Finally, the observation that training would be economically unviable if
licences were required is, in substance, a judgment of policy. It may well be
correct, but it is precisely the concern to which the DPIIT committee responded
by proposing a statutory licence rather than an enlarged exception.
E. Subsequent Developments
ANI
has appealed. After the matter could not be taken up on 8 September 2026, a
Division Bench comprising Jhingan and Arora JJ issued notice to OpenAI in
mid-September and listed the appeal for hearing on 5 December 2026.¹⁷ ANI's
reported grounds are that Indian law adopts fair dealing rather than fair use,
and that the private-use clause cannot authorise commercial exploitation at
scale. The Broadband India Forum has applied to intervene in support of the
order under appeal.¹⁸ Whatever the outcome, the Division Bench will be deciding
an interlocutory question on a limited evidential record, a circumstance that
itself strengthens the case for legislative clarity.
V. The Comparative Landscape
None
of the jurisdictions examined in this Part has permitted commercial training by
enlarging a private-use or research exception. Those that permit it have done
so by statute and subject to conditions; those that resist it are moving, with
varying degrees of urgency, towards licensing. The position described is that
obtaining in September 2026.
A. The European Union
The
European Union draws a deliberate distinction between research and commerce.
Article 3 of the Directive on Copyright in the Digital Single Market permits
text and data mining by research organisations and cultural heritage
institutions without any possibility of opt-out, whereas article 4 permits
mining by any person unless the rightsholder has reserved its rights in an
appropriate, machine-readable manner.¹⁹ The Artificial Intelligence Act now
requires providers of general-purpose models to observe such reservations.²⁰
German courts have tested both limbs. The Hanseatic Higher Regional Court in Kneschke v LAION upheld a non-profit
dataset under the research exception and treated a natural-language notice on a
website as an ineffective reservation.²¹ The Munich Regional Court in GEMA v OpenAI reached the opposite
conclusion in respect of memorised song lyrics, holding that the mining
exceptions cover acts of analysis but not the retention of works within a
model.²² The first reference to the Court of Justice concerning generative AI, Like Company v Google, asks whether
training constitutes reproduction and, if so, whether article 4 applies; the
Grand Chamber heard argument in March 2026.²³
B. The United States
American
courts, applying the open-ended fair-use standard in 17 USC § 107, have tended
to divide along the line between the manner in which data was acquired and the
use to which it was put. In Bartz v
Anthropic, training on lawfully purchased books was held to be fair use,
whereas the assembly of a library from pirated copies was not; the litigation
subsequently settled for USD 1.5 billion.²⁴ In Kadrey v Meta, the defendant prevailed because the plaintiffs had
failed to establish market harm, although the Court cautioned that market
dilution might defeat fair use in other circumstances.²⁵ Thomson Reuters v Ross, the only decision to reject fair use
outright, concerned a non-generative competing product and is now before the
Third Circuit, which heard argument in June 2026.²⁶ The case most directly
concerned with news is New York Times v
Microsoft and OpenAI, in which motions for summary judgment remain
pending.²⁷
C. The United Kingdom
The
United Kingdom's text-and-data-mining exception is confined to non-commercial
research.²⁸ In December 2024 the government proposed a commercial exception
subject to an opt-out, but its statutory report of 18 March 2026 abandoned that
preference in favour of a more cautious, evidence-gathering posture. In the
meantime, Getty Images v Stability AI
held that model weights which do not store the training images are not
infringing copies, and that training conducted abroad fell outside the reach of
the English court.²⁹ The consequence is a rule that is stringent in principle
at home but difficult to enforce against models trained elsewhere.
D. Japan and Singapore
Japan
and Singapore have adopted permissive statutory solutions. Article 30-4 of the
Japanese Copyright Act permits the exploitation of works for purposes that do
not involve the enjoyment of their expression, subject to a proviso excluding
uses that unreasonably prejudice the interests of the copyright owner.³⁰
Official guidance treats training directed at reproducing a work's expression
as falling outside the provision, and major newspaper publishers have sued
Perplexity in respect of retrieval-based answers rather than training alone.³¹
Singapore's section 244 permits copying for the purpose of computational data
analysis where the user has lawful access to the material, and the Act renders
void any contractual term purporting to exclude that permission.³²
E. Australia
Australia
has moved furthest in the opposite direction. The Attorney-General formally
ruled out a text-and-data-mining exception in October 2025, and on 15 July 2026
the Prime Minister announced that the use of Australian works for AI training
would remain subject to creators' control, with implementing legislation
anticipated in early 2027.³³ Statutory, collective and voluntary licensing
models remain under consultation.
F. Lessons for India
Three
lessons emerge from this survey. First, permission for commercial training has
everywhere been a legislative choice, and it has invariably been accompanied by
conditions, whether lawful access, a reservation mechanism or a test of
non-enjoyment. Secondly, courts are converging upon a distinction between
training, memorisation and retrieval. Thirdly, provenance matters: copies
obtained from pirated or access-controlled sources fare poorly in almost every
jurisdiction. The Delhi judgment is consistent with the second and third of
these lessons, but not with the first.
VI. News and Indian-Language Material
A. The Distinctive Character of News
News
presents the weakest case for publishers at the stage of training and the
strongest at the stage of retrieval. Copyright protects expression, not the
facts that a reporter gathers. A model that learns from thousands of wire
reports concerning an election absorbs facts and patterns of phrasing, and
seldom retains any single report. The more serious commercial threat
materialises later, when a conversational system retrieves the day's reporting
at the moment of the query and answers the reader's question in full. The Delhi
High Court was alive to this distinction, observing that the outputs before it
resembled live retrieval rather than memorised text.³⁴ It follows that Indian
law should treat training and retrieval as distinct activities, since the risk
of substitution is concentrated in the latter.
B. The Particular Stakes for Indian Languages
Regional-language
journalism presents a more acute difficulty still. Digitised text in most of
the twenty-two languages listed in the Eighth Schedule to the Constitution is
scarce in comparison with English. Scarcity enhances the value of each article
to a developer seeking capability in Odia, Maithili or Bhojpuri, while the
publishers who produce that text are, for the most part, small and modestly
resourced. Three consequences follow. The first is that any opt-out burden
falls unevenly: a rule that treats an unblocked crawler as consent protects
large English-language publishers with technical staff while exposing
district-level newspapers that have none. The second is that the translation
right is engaged, since section 14(a)(iv) reserves to the owner the right to
make any translation of a literary work.³⁵ A system that ingests a Tamil report
and conveys its substance in English, or the converse, may implicate that right
in ways that a largely English-centred debate has overlooked. The third is that
remuneration may fail to reach those who generate the material. The DPIIT model
pays only in respect of registered works and distributes through collecting
societies,³⁶ yet many regional publishers belong to no such society, with the
result that unclaimed royalties would pass to a welfare fund rather than to
those whose text built the capability in question.
A
regime appropriate to Indian conditions must therefore do more than transplant
a foreign text-and-data-mining clause. It must take account of linguistic
scarcity, of uneven technical capacity and of the translation right.
VII. Redefining Fair Dealing: A
Tiered Framework
It
is submitted that Parliament should confine section 52(1)(a) to the function
for which it was designed and legislate separately for machine ingestion. The
framework proposed below distinguishes three activities that the present debate
tends to conflate.
A. Non-Commercial Research Mining
A
new clause should be inserted in section 52(1) permitting reproduction for the
purposes of text and data mining by universities, research institutions and
cultural heritage bodies, where undertaken for non-commercial research. The
permission should be conditioned on lawful access, should admit of no opt-out,
and should be protected against contractual override. Such a provision would
mirror article 3 of the Digital Single Market Directive and section 29A of the
United Kingdom statute.³⁷
B. Commercial Training on Lawfully Accessed Works
Commercial
training should be authorised by statute, rather than by an extended reading of
'private use', and it should be remunerated. The DPIIT committee's proposal of
a statutory licence with a right to remuneration supplies an appropriate
foundation,³⁸ but three refinements would render it more equitable. First,
rather than relying on crawler directives, rightsholders should be able to
reserve their works, without charge, through a publicly maintained register;
reservation would convert the statutory licence into an obligation to
negotiate, an arrangement suited to large news archives that does not burden
smaller publishers. Secondly, a fixed proportion of collected royalties should
be ring-fenced for works in Eighth Schedule languages other than English, to be
distributed through publishers' associations where no collecting society
exists. Thirdly, the licence should not extend to copies obtained by
circumventing paywalls or technological protection measures in contravention of
section 65A, or to copies derived from known pirate sources.³⁹
C. Retrieval and Output
The
retrieval of current news at the time of a query should fall outside the
training licence altogether. A conversational system that fetches and
summarises the day's reports functions more as a news aggregator than as a
learner. It should be free to reproduce very short extracts with attribution
and a link, consistently with the reporting limb in section 52(1)(a)(iii),
while fuller summaries should require a licence.⁴⁰ Outputs that reproduce a
substantial part of a work should, in any event, remain infringing irrespective
of the regime governing inputs.
D. Transparency
None
of these measures can operate effectively without disclosure. Developers
placing models on the Indian market should be required to file a sufficiently
detailed summary of their training sources, as both the DPIIT committee and the
EU Artificial Intelligence Act contemplate.⁴¹
E. The Interim Judicial Approach
Pending
legislative intervention, courts should construe section 52(1)(a)(i) narrowly,
treat evidence of a licensing market as relevant to the assessment of fairness,
and decline to infer consent from the absence of a crawler block. The Division
Bench hearing ANI's appeal has an opportunity to articulate precisely these
principles.
VIII. Conclusion
The
Delhi High Court has shown that section 52(1)(a) can be made to accommodate
commercial generative ingestion. The more pressing question is whether it ought
to be. The refusal of interim relief was defensible on a limited record: there
was no proof that ANI's pleaded articles formed part of the training corpus, no
evidence of lost subscribers, and a claimant that had never excluded the
crawler. The reasoning, however, carries consequences that extend well beyond
the dispute. To read 'private' as including a corporate server, and 'research'
as including product development, is to move Indian law towards an open-ended
fair-use standard without Parliament having resolved to do so.
The
comparative record points in the other direction. Those jurisdictions that
permit commercial mining have done so by statute and subject to conditions,
while those that have declined to do so are turning to licensing.
India
should make this choice deliberately. A tiered statute that frees
non-commercial research, subjects commercial training to a remunerated licence,
and treats the retrieval of current news as a distinct market would afford
developers the certainty they seek while securing for publishers, and
especially for those writing in Indian languages, a share of the value that
their work creates. Fair dealing could then return to its original office: the
protection of the student, the critic and the reporter.
Notes
1. ANI Media Pvt Ltd v Open AI OpCo LLC
2026:DHC:5900 (Del), IA 45300/2024 in CS(COMM) 1028/2024 (24 July 2026) (Amit
Bansal J); see the summary of the judgment in Intepat, 'ANI vs OpenAI Judgment:
What the Delhi High Court Actually Held' (6 August 2026) <https://www.intepat.com/blog/ani-vs-openai-judgment-delhi-high-court>
accessed 27 September 2026.
2. MediaNama,
'ANI appeals the ruling that made AI training on Indian news fair dealing'
(September 2026) <https://www.medianama.com/2026/09/223-ani-ai-training-indian-news/>
accessed 27 September 2026.
3. DPIIT
Committee on Generative AI and Copyright, Working
Paper on Generative AI and Copyright, Part 1: One Nation One Licence One
Payment (DPIIT, 8 December 2025) <https://www.dpiit.gov.in/static/uploads/2025/12/ff266bbeed10c48e3479c941484f3525.pdf>
accessed 27 September 2026.
4. Copyright
Act 1957, ss 14(a)(i), 51.
5. ibid
s 52(1)(a) and Explanation, inserted by the Copyright (Amendment) Act 2012.
6. ibid
s 52(1)(b)–(c).
7.
Civic Chandran v Ammini Amma 1996 PTR 142 (Ker).
8. Super Cassettes Industries Ltd v
Hamar Television Network Pvt Ltd (2011) 45 PTC 70 (Del).
9. Chancellor, Masters and Scholars of
the University of Oxford v Rameshwari Photocopy Services
2016 SCC OnLine Del 6229 (DB).
10. Berne
Convention for the Protection of Literary and Artistic Works (as amended 28
September 1979) art 9(2); Agreement on Trade-Related Aspects of Intellectual
Property Rights (15 April 1994) art 13.
11. Copyright,
Designs and Patents Act 1988, s 29(1).
12. Copyright
Act 1957, s 52(1)(ad).
13. ANI (n 1); MediaNama,
'Delhi HC seeks OpenAI response in ANI case: Here's the timeline' (September
2026) <https://www.medianama.com/2026/09/223-delhi-hc-openai-ani-case/>
accessed 27 September 2026 (reporting discoverability partnerships between
OpenAI and Indian newspaper groups).
14. ANI (n 1).
15. ibid,
relying on Academy of General Education,
Manipal v B Malini Mallya (2009) 4 SCC 256 and R G Anand v Delux Films (1978) 4 SCC 118, and distinguishing Tips Industries Ltd v Wynk Ltd 2019 SCC
OnLine Bom 13087.
16. Berne
Convention (n 10) art 5(2).
17. 'Delhi
High Court seeks OpenAI reply on ANI challenge to copyright ruling' Business Standard (New Delhi, 15
September 2026) <https://www.business-standard.com/india-news/delhi-high-court-seeks-openai-reply-on-ani-challenge-to-copyright-ruling-126091501292_1.html>
accessed 27 September 2026.
18. MediaNama
(n 13).
19. Directive
(EU) 2019/790 on copyright and related rights in the Digital Single Market
[2019] OJ L130/92, arts 3–4.
20. Regulation
(EU) 2024/1689 laying down harmonised rules on artificial intelligence [2024]
OJ L 2024/1689, art 53(1)(c).
21. Kneschke v LAION eV
(OLG Hamburg, 5 U 104/24, 10 December 2025), affirming LG Hamburg, 310 O 227/23
(27 September 2024).
22. GEMA
v OpenAI (LG München I, 42
O 14139/24, 11 November 2025).
23. Case
C-250/25 Like Company v Google Ireland
Ltd (pending); Bird & Bird, 'Like Company v Google: CJEU Holds
First-Ever Hearing on Generative AI and Copyright' (10 March 2026) <https://www.twobirds.com/en/insights/2026/like-company-v-google-cjeu-holds-first-ever-hearing-on-generative-ai-and-copyright-on-10-march-2026>
accessed 27 September 2026.
24. Bartz v Anthropic PBC
No C 24-05417 WHA (ND Cal, 23 June 2025).
25. Kadrey v Meta Platforms Inc
No 23-cv-03417-VC (ND Cal, 25 June 2025).
26. Thomson Reuters Enterprise Centre
GmbH v Ross Intelligence Inc 765 F Supp 3d 382 (D Del
2025), appeal pending No 25-2153 (3d Cir, argued 11 June 2026).
27. New York Times Co v Microsoft Corp
No 1:23-cv-11195 (SDNY); 'Historic NYT v. OpenAI copyright battle heats up' Axios (8 September 2026) <https://www.axios.com/2026/09/08/nyt-openai-microsoft-copyright-lawsuit>
accessed 27 September 2026.
28. Copyright,
Designs and Patents Act 1988, s 29A; Data (Use and Access) Act 2025, ss
135–137; Department for Science, Innovation and Technology, Intellectual
Property Office and Department for Culture, Media and Sport, Report on Copyright and Artificial
Intelligence (18 March 2026); Law Library of Congress, 'United Kingdom:
Government Adopts Wait-and-See Approach to Regulating AI and Copyright' (8
April 2026) <https://www.loc.gov/item/global-legal-monitor/2026-04-08/united-kingdom-government-adopts-wait-and-see-approach-to-regulating-ai-and-copyright>
accessed 27 September 2026.
29. Getty Images (US) Inc v Stability AI
Ltd
[2025] EWHC 2863 (Ch).
30. Copyright
Act (Japan) (Act No 48 of 1970) art 30-4.
31. Legal
Subcommittee under the Copyright Subdivision of the Cultural Council, General Understanding on AI and Copyright in
Japan (Agency for Cultural Affairs, 2024); 'Nikkei and Asahi Shimbun sue
Perplexity AI over alleged copyright violations' The Japan Times (Tokyo, 26 August 2025) <https://www.japantimes.co.jp/news/2025/08/26/japan/crime-legal/japan-newspapers-sue-ai-startup/>
accessed 27 September 2026.
32. Copyright
Act 2021 (Singapore), ss 187, 243–244.
33. MinterEllison,
'Copyright and AI: Consultation on licensing for AI training' <https://www.minterellison.com/articles/copyright-and-ai-consultation-on-licensing-for-ai-training>
accessed 27 September 2026; Prime Minister of Australia, 'AI in Australia's
Interest' (Speech, University of Sydney, 15 July 2026) <https://www.pm.gov.au/media/ai-australias-interests-0>;
Hamilton Locke, 'No free pass for AI' (23 July 2026) <https://hamiltonlocke.com.au/no-free-pass-for-ai-australia-confirms-copyright-will-be-protected-under-new-mandatory-framework/>
accessed 27 September 2026.
34. ANI (n 1).
35. Copyright
Act 1957, s 14(a)(iv).
36. DPIIT
Working Paper (n 3); Ikigai Law, 'Exploring the DPIIT's working paper on
Generative AI and Copyright' (10 December 2025) <https://www.ikigailaw.com/article/655/exploring-the-dpiits-working-paper-on-generative-ai-and-copyright>
accessed 27 September 2026.
37. Directive (EU) 2019/790 (n 19) art 3; CDPA 1988 (n 28)
s 29A.
38. DPIIT
Working Paper (n 3).
39. Copyright
Act 1957, s 65A; cf Bartz (n 24).
40. Copyright
Act 1957, s 52(1)(a)(iii); cf Directive (EU) 2019/790 (n 19) art 15(1).
41. DPIIT
Working Paper (n 3); Regulation (EU) 2024/1689 (n 20) art 53(1)(d).
Bibliography
Cases
•
Academy
of General Education, Manipal v B Malini Mallya
(2009) 4 SCC 256
•
ANI
Media Pvt Ltd v Open AI OpCo LLC 2026:DHC:5900 (Del)
•
Bartz
v Anthropic PBC No C 24-05417 WHA (ND Cal, 23 June 2025)
•
Chancellor,
Masters and Scholars of the University of Oxford v Rameshwari Photocopy
Services 2016 SCC OnLine Del 6229 (DB)
•
Civic Chandran v Ammini Amma 1996 PTR 142 (Ker)
•
GEMA v OpenAI (LG München I, 42 O 14139/24, 11 November 2025)
•
Getty
Images (US) Inc v Stability AI Ltd [2025] EWHC 2863 (Ch)
•
Kadrey
v Meta Platforms Inc No 23-cv-03417-VC (ND Cal, 25 June 2025)
•
Kneschke
v LAION eV (OLG Hamburg, 5 U 104/24, 10 December 2025)
•
Case C-250/25 Like Company v Google Ireland Ltd (pending)
•
New
York Times Co v Microsoft Corp No 1:23-cv-11195 (SDNY)
•
R
G Anand v Delux Films (1978) 4 SCC 118
•
Super
Cassettes Industries Ltd v Hamar Television Network Pvt Ltd
(2011) 45 PTC 70 (Del)
•
Thomson
Reuters Enterprise Centre GmbH v Ross Intelligence Inc
765 F Supp 3d 382 (D Del 2025)
•
Tips
Industries Ltd v Wynk Ltd 2019 SCC OnLine Bom 13087
Legislation and treaties
•
Berne Convention for the Protection of
Literary and Artistic Works (as amended 28 September 1979)
•
Copyright Act 1957 (India)
•
Copyright Act (Japan) (Act No 48 of 1970)
•
Copyright Act 2021 (Singapore)
•
Copyright, Designs and Patents Act 1988
(UK)
•
Data (Use and Access) Act 2025 (UK)
•
Directive (EU) 2019/790 on copyright and
related rights in the Digital Single Market [2019] OJ L130/92
•
Regulation (EU) 2024/1689 laying down
harmonised rules on artificial intelligence [2024] OJ L 2024/1689
•
Agreement on Trade-Related Aspects of
Intellectual Property Rights (15 April 1994)
Official reports
•
DPIIT Committee on Generative AI and
Copyright, Working Paper on Generative AI
and Copyright, Part 1 (DPIIT, 8 December 2025)
•
Department for Science, Innovation and
Technology, Intellectual Property Office and Department for Culture, Media and
Sport, Report on Copyright and Artificial
Intelligence (18 March 2026)
•
Legal Subcommittee under the Copyright
Subdivision of the Cultural Council, General
Understanding on AI and Copyright in Japan (Agency for Cultural Affairs,
2024)
Secondary sources
•
Bird & Bird, 'Like Company v Google:
CJEU Holds First-Ever Hearing on Generative AI and Copyright' (10 March 2026)
•
Hamilton Locke, 'No free pass for AI:
Australia confirms copyright will be protected under new mandatory framework'
(23 July 2026)
•
Ikigai Law, 'Exploring the DPIIT's working
paper on Generative AI and Copyright' (10 December 2025)
•
Intepat, 'ANI vs OpenAI Judgment: What the
Delhi High Court Actually Held' (6 August 2026)
•
Law Library of Congress, 'United Kingdom:
Government Adopts Wait-and-See Approach to Regulating AI and Copyright' (8
April 2026)
•
MediaNama, 'ANI appeals the ruling that
made AI training on Indian news fair dealing' (September 2026)
•
MinterEllison, 'Copyright and AI:
Consultation on licensing for AI training'
How to Cite This Article
MR. PRATYUSH PANDEY, DR. ANSHUMAN MISHRA, Expanding the Exception: Generative AI Training and the Architecture of Indian Copyright Law., White Black Legal – International Law Journal, ISSN: 2581-8503, Vol. Volume 4, Issue VOLUME 4 ISSUE 2, October 2026, pp. 1259-1277. Available at: https://www.whiteblacklegal.co.in/details/expanding-the-exception-generative-ai-training-and-the-architecture-of-indian-copyright-law
Author & Publication Record
Authors: MR. PRATYUSH PANDEY & DR. ANSHUMAN MISHRA
Registration ID: 107206 | Published Paper ID: WBL7206 & WBL7207
Year: Oct- 2026 | Volume: 4 | Issue: 2
Approved ISSN: 2581-8503 | Country: Delhi, India
Page No.: 1259-1277
Full Text Preview
Open in New Tab
Copied