This article is originally published in Mondaq at Gen-AI and Copyright – A Legal Tug Of War Impacting the World and Touching the Shores of India.

Introduction

Generative artificial intelligence (“Gen-AI”) is an idea whose time has come. It is no more the future of technology but its present. Gen-AI is today being used in every aspect of our professional, business, and personal lives. As with the advent of every technological development, the world is finding itself in a vacuum on existing regulations to streamline and regulate the scope of such technology. Gen-AI is no different. One such legal hurdle that the world is now identifying, is the co-existence and relevance of copyright protection, when novel work-products are used to develop and grow Gen-AI. While most jurisdictions are still rappelling around regulations on use of Gen-AI, some jurisdictions have been able to make more headway than others.

Copyright is a legal recognition of an author’s exclusive right over his creative work, including the right to display, distribute, perform, and gain financial benefit from such work. With the advent of technology, Gen-AI poses an unprecedented challenge to preservation of copyright. Although Gen-AI increases efficiency by generating new content, these outputs may use and/ or closely resemble copyrighted works, without the author’s permission. This raises an important question of preserving original work, and whether outputs generated by Gen-AI constitutes a copyright infringement.

An exception to copyright infringement under Indian law is ‘fair dealing’, being the legally permissible extent to which copyrighted work can be used without consent of the owner. The [Indian] Copyright Act, 1957 (“Copyright Act”) lays down an exhaustive list of categories which would not constitute infringement,[1] including ‘fair dealing’, as compared to other jurisdictions where broader discretion is vested upon courts.

The jurisprudence pertaining to the interplay between copyright protection and Gen-AI is at a nascent stage in India and is still evolving. This article aims to identify the evolution of this law across various jurisdictions, the relevance of the same in the Indian context and the direction being taken by the legislature and the judiciary in India alike. We also explore certain directions, orders and judgments passed by High Courts across the country, which holds the field until a consistent and confirmed law is laid down.

DHC speaks on ‘fair dealing’ and Gen-AI

On 24 July 2026, the Hon’ble Delhi High Court spoke its ‘prima facie’ mind on the test of ‘fair dealing’ by Gen-AI i.e., OpenAI while learning from copyrighted material. In a judgment on an interim application for injunction in ANI Media Pvt. Ltd. v. Open AI Inc.,[2] the Hon’ble Delhi High Court refused to grant protective relief to ANI rendering first of its kind prima facie findings which substantively examine whether training a Gen-AI system on copyrighted content falls within ‘fair dealing’ under Section 52 of the [Indian] Copyright Act. Although rendered at interim stage, the decision provides much-needed judicial recognition of content storage and reproduction by AI under the [Indian] Copyright Act.

The Court examined whether the storage of ANI’s copyrighted works for training would itself constitute infringement, and rendered the decision in favour of Open AI, holding that such storage would be subject to a ‘fair dealing’ test under Section 52(1)(a)(i). In doing so, the Court did not apply the US four-factor test, holding it derives from the US statute itself. Instead, referring to Article 9 of the Berne Convention for the Protection of Literary and Artistic Works (1886), the Court laid down its own three-factor test.

The test now includes a two-step examination: (1) Purpose Test, i.e., whether the purpose of storage is for private or personal use, including research; and (2) Fairness Test, i.e., whether the content produced by Gen-AI prejudiced ANI’s legitimate commercial interests.

The three factors to determine ‘fair dealing’ under Section 52(1)(a) of the [Indian] Copyright Act with respect to Gen-AI would include:

1. Whether the AI’s use of the copyrighted content is limited to training its large language models.
2. Whether the usage of the copyrighted work by the AI would result in economic competition and would prejudice the legitimate interests of the copyright holder, thereby causing actual or potential damage?
3. Whether the functions performed by the AI serve the overall public interest?

This approach by the Hon’ble Delhi High Court may act as a landing light for similar issues being dealt with by courts across the country. It is well known in similar copyright disputes that protective/injunctive reliefs at an interim stage are essential for the claiming party. This is especially because final adjudication of such disputes may take relatively longer by when great damage may be caused by purported misuse of the party’s copyright. By carrying out the in-depth analysis at the interim stage itself, the Hon’ble Delhi High Court has avoided any hinderance to the growth of Gen-AI in India.

First Sign of Judicial Resistance: The US Test

The US statute has expressly set out the four-fold test of “fair use” (similar to the concept of ‘fair dealing’ in India) which leaves it to courts to determine use after considering four statutory factors[3]: (i) the purpose and character of the use; (ii) the nature of the copyrighted work; (iii) the amount and substantiality of the portion used; and (iv) the effect upon the potential market for the copyrighted work.

In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc.,[4] the United States District Court of Delaware held that Ross Intelligence’s use of copyrighted Westlaw headnotes for training its legal AI constituted copyright infringement even absent direct output reproduction.

The Court held that potential market harm caused to the copyright owner was sufficient, where the “competitor intended to compete with platform by developing a substitute in platform’s original market for legal research, and competitor’s conduct affected a potential derivative market for data for training legal AI”.

In Bartz v. Anthropic PBC[5], authors of literary works sued Anthropic alleging copyright infringement, claiming their books were pirated and used to train the large language model, Claude.

The US District Court of Northern District of California held that the purpose of using copyrighted works to train Claude was quintessentially transformative, aimed not at replicating the books but enabling the AI to create something different, constituting “fair use”. However, downloading and retention of the copyrighted content from pirate sources was not “fair use” — Anthropic should have paid for the over seven million books it kept in its internal library.

The Court thus drew a critical distinction: copying for training was protected as “fair use”, but maintaining a permanent library of pirated books was not.

A contrary position, albeit non-binding, was taken in Kadrey v. Meta Platforms, Inc.,[6] wherein thirteen authors instituted copyright infringement proceedings against Meta for unauthorised use of their books to train Llama. The Court criticized Bartz v. Anthropic, noting it brushed aside concerns that trained AI could generate countless competing works. However, since this argument was not raised by the plaintiffs, the Court granted summary judgment in Meta’s favor.

A more focused perspective into ‘fair dealing’: a UK outlook

Unlike the US, the UK does not recognise a broad “fair use” doctrine. Instead, it follows ‘fair dealing’ under the [English] Copyright, Designs and Patents Act 1988 (“CDPA”), permitting use only within specific statutory exceptions.

In Getty Images (US) Inc. v. Stability AI Ltd.,[7] the High Court of England and Wales dealt with alleged use of copyrighted images for training Stable Diffusion. The Court held that an AI model which does not store or reproduce any Copyright Works could not itself constitute an “infringing copy” for secondary infringement under the CDPA.

The Court rejected Getty’s argument that the trained model embodied infringing reproductions, but recognised that copyrighted images were downloaded, stored, and processed during training. It distinguished between temporary reproductions made during training and the trained model ultimately produced.

The Court recognised that copyright legislation may apply to unforeseen technological developments, while cautioning that courts cannot “fill gaps” where Parliament has not expressly legislated for GenAI.

Europe’s Most Protective Position

In Germany, copyright protection is governed by the [German] Copyright Act, 1965,[8] which provides a closed catalogue of statutory limitations to use of copyrighted works. In GEMA v OpenAI, the Munich Regional Court determined whether using copyrighted song lyrics in training and operating large language models constituted infringement under German and European copyright law.[9]

The Court substantially upheld GEMA’s claims, holding that both the incorporation of copyrighted lyrics into the model and their subsequent reproduction through chatbot outputs infringed copyright exploitation rights. The Court found that the ability to reproduce protected lyrics upon prompting demonstrated the works had been “memorised” within the language model.

The Japanese Outlier

Article 30-4 of the [Japanese] Copyright Act, 1970 permits the use of copyrighted works for purposes of “information analysis” rather than “enjoyment” of expressive content. It permits “exploitation” for: (a) testing technology connected with recording sounds or visuals; (b) data analysis; and (c) computer data processing not involving perception by human senses.

Article 30-4 may permit ingestion of copyrighted works for model training without prior authorisation, provided the use does not unreasonably prejudice the interests of the copyright owner. The Japanese Agency for Cultural Affairs clarified in its “General Understanding on AI and Copyright” (“Guidance”) that AI training ordinarily qualifies as “information analysis”.[10]

However, the Guidance identifies limitations: the exception does not apply where use is directed toward reproducing or allowing users to experience the expressive content itself, particularly where AI is intentionally trained to generate outputs closely resembling copyrighted works.

Article 30-4 represents one of the most permissive statutory frameworks globally for AI training. No analogue exists in Indian law, which prompted the Hon’ble Delhi High Court in ANI Media v. Open AI to expand the scope of ‘fair dealing’ under Section 52(1)(a)(i) of the [Indian] Copyright Act.

Balancing the Growth of Gen-AI against Copyright Protection

The global judicial trends as also the position taken in India reveal a visible direction towards acceptance of Gen-AI’s use of copyrighted works for training purposes. The overwhelming weight of global jurisprudence tilts in favor of recognizing and categorizing the output by Gen-AI within the existing copyright framework. This supports technological advancement and deters stifling of Gen-AI’s learning by strict application of copyright law.

The Hon’ble Delhi High Court’s findings in ANI Media v. Open AI are prima facie observations and cannot be accepted as the final legal position in India. That said, given the formulation of three-factor test by the Hon’ble Delhi High Court, it is possible that courts in India may develop the law along similar lines. These factors serve as a guiding mechanism for courts to determine, most importantly, that whether the output generated is used for private use, or for commercial use by the user. It is also to be borne in mind that the prima facie observations are consistent with global jurisprudence, buttressing the view that it is the first step towards a confirmed position in India on this aspect.

If the position of fair dealing is not adopted and the law instead leans toward restricting AI’s use of copyrighted content for training, the consequences would be far-reaching. The factors enumerated by the Hon’ble Delhi High Court must be looked at to serve broader public interest, i.e., dissemination of information and knowledge, and to curb exploitation of copyrighted content under the pretense of ‘commercial use’ under the [Indian] Copyright Act. Courts in India would have to arrive at a balanced approach, on the one hand protecting copyright laws, and on the other, accepting technological advancement consistent with global approach. It will be interesting to see as to how courts seek to balance the interests of copyright owners while accepting the training by Gen-AI.

[1] Section 52, [Indian] Copyright Act, 1957.

[2] Ani Media Pvt. Ltd. v. Open AI Opco LLC, I.A. 45300/2024 in CS(Comm) 1028/2024, judgment dated July 24, 2026.

[3] 17 U.S.C. § 107.

[4] Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., 765 F. Supp. 3d 382 (D. Del. 2025).

[5] Bartz v. Anthropic PBC, No. 3:24-cv-05417 (N.D. Cal. 23 June 2025).

[6] Kadrey v. Meta Platforms, Inc., No. 23-cv-03417 (N.D. Cal. 25 June 2025).

[7] Getty Images (US) Inc. v. Stability AI Ltd., [2025] EWHC 2863 (Ch).

[8] Copyright Act of 9 September 1965 (Federal Law Gazette I, p. 1273), as last amended by Article 28 of the Act of 23 October 2024 (Federal Law Gazette 2024 I No. 323).

[9] GEMA v OpenAI (Munich I Regional Court, Case No. 42 O 14139/24, decided on 11 November 2025).

[10] General Understanding on AI and Copyright in Japan, Legal Subcommittee under the Copyright Subdivision of the Cultural Council, May 2024, available here.