Authored By: Vishwa Jyoti
Campus Law Centre University of Delhi
INTRODUCTION
The Delhi High Court’s interim order of 24 July 2026 in ANI Media Pvt Ltd v Open AI OpCo LLC has made copyright law central to India’s debate on generative artificial intelligence. The Court declined ANI’s application for an interim injunction against OpenAI. On a prima facie view, it held that OpenAI’s storage of ANI’s works for training the large language models underlying ChatGPT could fall within the fair-dealing exception in section 52(1)(a) of the Copyright Act 1957. It also found that the outputs relied upon by ANI did not establish substantial similarity, memorisation or regurgitation of ANI’s works.[1]
The order is significant, but its limits are equally important. It is an interim determination, its observations do not decide the pending suit, and it does not establish an unlimited right to use copyrighted works for all forms of commercial AI development. The dispute nevertheless exposes a structural mismatch between copyright law’s conventional focus on identifiable authors and discrete acts of copying, and models trained through large-scale computational processing of text, images, sound and code.[2]
This article argues that India should neither prohibit AI training categorically nor grant unrestricted immunity. It should distinguish training from output infringement, preserve a limited space for genuine research, require proportionate transparency for commercial providers, recognise copyright where meaningful human creativity is reflected in the final expression, and develop licensing, collective licensing or effective opt-out mechanisms for large-scale commercial uses. The article first examines the statutory framework, then analyses ANI, identifies structural gaps, considers the European Union model, and proposes targeted reforms.
COPYRIGHT LAW – GENERATIVE AI
Protected Expression and Reproduction
The Copyright Act 1957 protects original literary, dramatic, musical and artistic works, as well as cinematograph films and sound recordings. Section 13 identifies the protected subject matter. Section 14 gives copyright owners exclusive rights, while section 51 identifies circumstances in which unauthorised exercise of those rights may amount to infringement.
For literary works, section 14 expressly includes storing a work in any medium by electronic means within the right of reproduction. AI training can therefore involve legally relevant copying when source works are collected, stored, processed or converted into machine-readable inputs. The central question is not whether the technology is novel, but whether the particular copying is authorised or falls within an applicable statutory exception. he existence of a reproduction is therefore only the starting point. The decisive question is whether the copying is authorised, licensed or protected by a statutory exception.[3]
The distinction between ideas, facts and protected expression also remains important. Copyright does not confer a monopoly over news events or facts, but it protects original expression. In R G Anand v Deluxe Films, the Supreme Court treated substantial similarity in protected expression as central to infringement analysis. The principle is directly relevant to AI outputs: a response may discuss the same event as a protected article without infringing, but reproduction of protected expression may give rise to infringement.[4]
AI-Assisted Authorship
Section 2(d)(vi) of the Copyright Act provides that, in relation to a computer-generated literary, dramatic, musical or artistic work, the author is the person who causes the work to be created. The provision is a useful starting point, but generative systems complicate the identification of that person. Developers build models, providers operate them, users prompt them, and editors may select, alter and arrange outputs.[5]
The better approach is to distinguish causation from meaningful authorship. Eastern Book Company v D B Modak confirms that originality requires independent creation and at least a minimal degree of creativity: trivial or purely mechanical input is insufficient. Applied cautiously to generative AI, a user who makes iterative creative choices, selects among outputs and substantially edits the final expression has a stronger claim than a user who merely enters a short prompt and accepts the unedited response. The inquiry should focus on the claimant’s contribution to the final expression, rather than merely on the claimant’s role in initiating the software process.[6]
This approach protects AI as a tool without treating every machine-generated response as a new private right. It also recognises that section 2(d)(vi) should not be read mechanically: the fact that a person initiated a computational process does not necessarily establish that the resulting expression embodies sufficient original human contribution.
Fair Dealing and Section 52
Section 52(1)(a) provides that fair dealing with a work, other than a computer programme, for private or personal use including research, criticism or review, and certain reporting, does not constitute infringement. The provision must be read together with the statutory treatment of electronic storage and the specific purposes identified in section 52(1)(a). It does not state a blanket rule that every commercial use is unfair or that every use described as research is fair.[7]
Indian fair-dealing jurisprudence treats fairness as contextual rather than mechanical. Civic Chandran v Ammini Amma and Super Cassettes Industries Ltd v Hamar Television Network Pvt Ltd demonstrate that the statutory defence depends upon the purpose and character of the dealing and the circumstances of the particular use. The analysis should therefore consider the nature and amount of copying, the purpose of the use, market substitution and the legitimate interests affected.[8]
AI training makes this inquiry unusually difficult. A research institution using a limited, lawfully accessed corpus for scientific investigation is materially different from a commercial provider ingesting millions of expressive works to develop a general-purpose model. The distinction should not depend only on whether both activities can be described as “research”; scale, safeguards, access, market effects and available licensing routes also matter. These difficulties explain why the ANI order is important but cannot substitute for a general legislative framework.
THE ANI INTERIM ORDER
The Interim Decision
The relevant proceeding is ANI Media Pvt Ltd v Open AI OpCo LLC, CS(COMM) 1028/2024, before the Delhi High Court. Justice Amit Bansal issued the relevant interim order on 24 July 2026. Court separated the training claim from the alleged reproduction of ANI’s works in outputs. The Court identified a two-step examination under section 52(1)(a): first, whether the use satisfied one of the statutory purposes; and secondly, whether the dealing was fair.
On the purpose question, the Court held prima facie that the training of the LLMs could constitute research and that the Court treated the relevant storage and use as falling within the statutory framework of research, subject to its prima facie assessment. It applied an updating construction to the statutory term “research”, reasoning that contemporary technological developments should be accommodated without treating AI-assisted research as outside the statutory language.[9]
On fairness, the Court considered the purpose and character of the use, the absence of demonstrated market substitution, and the public benefits associated with LLMs. It concluded, at the interim stage, that both the purpose and fairness tests were satisfied. The Court therefore held prima facie that OpenAI’s storage of ANI’s works for training fell within section 52(1)(a).[10]
The Court separately found that the outputs relied upon by ANI were not substantially similar to its original literary works and that ANI had not established memorisation and regurgitation. The decision therefore illustrates why training legality and output infringement must be analysed as distinct questions. The order should therefore be treated as fact-specific interim guidance rather than as a universal exemption for model training.[11]
What the Decision Does Not Decide
ANI should not be treated as a general safe harbour for AI developers. The Court expressly limited its observations to the interim application and stated that they would not affect the final outcome of the suit. The evidentiary record in a final trial may differ, particularly regarding datasets, memorisation, output behaviour, market harm, licensing and the technical operation of a model.[12]
The decision also does not establish that commercial AI training will always be fair. Its reasoning was tied to the statutory text and the evidence before the Court, including the absence of demonstrated market substitution and the particular use of ANI material. A provider that reproduces protected expression, ignores effective rights reservations, or uses training data in a manner that substitutes for the original market may present a materially different case.
STRUCTURAL GAPS IN INDIAN LAW
Information Asymmetry
The most serious practical difficulty is informational asymmetry. Model providers control evidence needed to assess a copyright claim, including dataset records, collection methods, crawler logs, licences, filtering systems, deduplication processes and technical material concerning memorisation. Individual authors and smaller publishers usually cannot determine whether their works were included in training data without unusual output evidence or access to provider-held information. This imbalance affects not only the merits of a claim but also the claimant’s ability to plead a legally sustainable case.[13]
India should require proportionate record-keeping for commercial general-purpose model providers. Providers should retain information about source categories, acquisition methods, relevant dates, filtering and copyright-risk controls, while publishing a high-level summary. More detailed information could be disclosed confidentially to a court or regulator when a claimant establishes a credible basis for investigation, protecting trade secrets while reducing the information imbalance.
Market Effects and Collective Licensing
Mass training creates a collective-action problem. Individual authors may lack the bargaining power or practical ability to negotiate with a provider, while one-to-one licensing of enormous datasets may be costly and administratively unrealistic. The Copyright Act already recognises copyright societies as collective licensing institutions capable of issuing licences and collecting and distributing fees.[14]
India could therefore consider a sector-specific collective licensing route for defined categories of commercial training. Such a mechanism should include representative governance, transparent remuneration, meaningful notice and an effective opt-out where appropriate. It should not automatically apply to public-domain material, open licences or genuinely non-commercial research, and it should not deprive individual rightsholders of ordinary licensing where individual negotiation is feasible. Any collective mechanism must remain voluntary or carefully defined by statute, preserve adequate notice and provide a fair method for distributing remuneration.
Personal Data and Copyright
The Digital Personal Data Protection Act 2023 regulates the processing of digital personal data and defines obligations around lawful processing and data fiduciaries. It addresses interests in personal data, not ownership of literary, artistic or other protected expression. A dataset may therefore raise both privacy and copyright questions, but compliance with one regime does not answer the other.[15]
AI governance should keep these interests analytically separate. Privacy law concerns individuals and personal information: copyright law concerns protected expression and the rights of authors and owners. Legislative coordination is desirable, but privacy compliance should not be treated as a substitute for copyright analysis.
LESSONS FROM THE EUROPEAN UNION
The European Union offers a useful comparative model because it combines text-and-data-mining rules with transparency obligations for general-purpose AI models. Directive (EU) 2019/790 creates a mandatory scientific-research text-and-data-mining exception for research organisations and cultural heritage institutions with lawful access, while Article 4 establishes a broader exception for lawfully accessible works where rightsholders have not appropriately reserved the relevant rights.[16]
The AI Act adds a separate transparency layer. Article 53 requires providers of general-purpose AI models to maintain a copyright-compliance policy, including respect for rights reservations under the Directive, and to make publicly available a sufficiently detailed summary of training content. The EU model therefore does not simply declare all training lawful: it combines defined exceptions with traceability and rights-management obligations.[17]
India should adapt this architecture rather than copy it. A proportionate Indian framework could distinguish non-commercial scientific research from commercial general-purpose training, require commercial providers to publish standardised high-level training summaries, and permit regulators or courts to require more specific confidential disclosure when justified. This would improve enforceability without requiring public disclosure of trade secrets. The EU model is persuasive rather than binding in India. Its value lies in the principles of transparency, traceability and differentiated treatment of research and commercial model development.
A PROPORTIONATE REFORM AGENDA
Define Computational Training
India should enact targeted amendments to the Copyright Act 1957 or a complementary AI-and-copyright framework. First, the law should define computational training and distinguish non-commercial scientific research from commercial general-purpose model development. A limited research exception could apply to lawfully accessed material subject to safeguards against redistribution and misuse. Commercial training should have a clearer statutory route through licensing, effective opt-out mechanisms or another legislatively defined framework.[18]
Require Records and Disclosure
Secondly, providers above a defined scale should maintain training-data records and publish standardised summaries. Thirdly, the law should adopt a human-contribution standard for AI-assisted outputs, assessing creative instructions, iterative selection, arrangement, editing and transformation rather than merely asking who pressed the button. Fourthly, remedies should be proportionate: injunctions should remain available for deliberate memorisation and repeated reproduction, while collective licensing, remuneration and technical safeguards may better address large-scale training disputes.
Adopt Proportionate Remedies
Reform should also recognise countervailing interests. Excessively burdensome licensing could entrench large incumbents by making model development unaffordable for smaller Indian firms and research institutions. Conversely, an unlimited training privilege could shift the costs of AI development entirely onto creators. The appropriate objective is therefore a framework that allocates those costs transparently while preserving competition, research and access to technology.
CONCLUSION
Generative AI exposes a mismatch between India’s existing copyright framework and contemporary methods of computational learning. The Copyright Act can address many disputes through its provisions on reproduction, fair dealing and computer-generated works, but it does not provide sufficient certainty about commercial training, evidence disclosure, collective remuneration or the human contribution required for AI-assisted authorship.
ANI is an important judicial development because it separates training from output infringement and recognises that LLM training may, on the evidence before the Court, fall within section 52(1)(a). It is nevertheless an interim decision, not a comprehensive legislative settlement. India should respond with targeted rules on training-data transparency, research exceptions, commercial licensing or opt-out mechanisms, human authorship and proportionate remedies.
India should adopt targeted rules on data transparency, record preservation, research exceptions, commercial licensing or effective opt-out mechanisms, human authorship and proportionate remedies. Such a framework would preserve genuine technological research while ensuring that commercial AI development does not make creators invisible within systems built from their work. The objective should be accountability without technological paralysis.
REFERENCE(S):
Table of Cases
ANI Media Pvt Ltd v Open AI OpCo LLC 2026:DHC:5900 (Delhi High Court, 2026)
Civic Chandran v Ammini Amma 16 PTC 329 (Kerala High Court, 1996)
Eastern Book Company v D B Modak (2008) 1 SCC 1
R G Anand v Deluxe Films (1978) 4 SCC 118
Super Cassettes Industries Ltd v Hamar Television Network Pvt Ltd 2010 SCC OnLine Del 2086
Table of Legislation
Copyright Act 1957
Digital Personal Data Protection Act 2023
Directive (EU) 2019/790 of the European Parliament and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market [2019] OJ L 130/92
Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence [2024] OJ L, 2024/1689, 12 July 2024
Bibliography
Ahmed Aaquil and Mahmood Faisal, ‘Crafting the Future: AI Governance, IP, and Privacy in India’s Digital Age’ (2025) 1 HILSR Law Review 143.
Edwards Lilian and Veale Michael, ‘Slave to the Algorithm? Why a Right to an Explanation is Probably Not the Remedy You are Looking For’ (2017) 16 Duke Law and Technology Review 18.
European Parliament and Council, Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence.
Ministry of Law and Justice, The Digital Personal Data Protection Act 2023 (Act 22 of 2023).
[1] ANI Media Pvt Ltd v Open AI OpCo LLC 2026:DHC:5900 (Delhi High Court, 2026) ¶¶ 271–274.
[2] Copyright Act 1957, ss 2(d)(vi), 14, 51 and 52.
[3] Copyright Act 1957, s 14(a)(i); ANI Media (n 1) ¶¶ 128–135.
[4] R G Anand v Deluxe Films (1978) 4 SCC 118.
[5] Copyright Act 1957, s 2(d)(vi).
[6] Eastern Book Company v D B Modak (2008) 1 SCC 1.
[7] Copyright Act 1957, s 52(1)(a).
[8] Civic Chandran v Ammini Amma 16 PTC 329 (Kerala High Court, 1996); Super Cassettes Industries Ltd v Hamar Television Network Pvt Ltd 2010 SCC OnLine Del 2086.
[9] ANI Media (n 1) ¶¶ 211–219.
[10] ANI Media (n 1) ¶¶ 220–256.
[11] ANI Media (n 1) ¶¶ 241–256, 271.
[12] ANI Media (n 1) ¶¶ 272–274.
[13] Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 [2024] OJ L, 2024/1689, 12 July 2024, art 53(1)(a), (c)–(d).
[14] Copyright Act 1957, ss 33–35.
[15] Digital Personal Data Protection Act 2023, ss 2(i), 4, 8.
[16] Directive (EU) 2019/790 of the European Parliament and of the Council of 17 April 2019 [2019] OJ L 130/92, arts 3–4.
[17] Regulation (EU) 2024/1689 (n 15) art 53(1)(c)–(d).
[18] Directive (EU) 2019/790 (n 18) art 4; Copyright Act 1957, s 33.
