Generative AI and Copyright Infringement: Can Training AI Models on Copyrighted Works Constitute Infringement in India?

Introduction

Generative Artificial Intelligence (Generative AI) has changed the way content is created. Tools such as ChatGPT, image-generation systems and AI music platforms can produce text, photographs, music, videos and other creative content within seconds. However, these systems do not learn in the same way as humans. They are trained using very large amounts of existing data, which may include books, photographs, articles, music, websites and other material protected by copyright.1

This has created an important legal question: If an AI developer uses copyrighted works to train an AI model without taking permission from the copyright owner, does that amount to copyright infringement?
The question is particularly difficult because copyright law was developed before modern AI systems existed. The Copyright Act, 1957 does not specifically mention "artificial intelligence training" or "machine learning datasets". Therefore, the issue has to be examined through existing concepts such as reproduction, infringement and fair dealing.

The World Intellectual Property Organization (WIPO) has also recognised this as one of the major unresolved copyright issues created by Generative AI. WIPO notes that AI training datasets can contain large amounts of copyright-protected material and that both AI developers and copyright owners have competing arguments regarding the legality of such use.2

In India, this issue has now received direct judicial consideration in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, where the Delhi High Court, at the interim stage, considered whether OpenAI's use of copyrighted news articles for training its AI models could be protected under the fair-dealing exception.3

Understanding How Generative AI Uses Copyrighted Material

Before examining infringement, it is important to understand what happens during AI training.
A Generative AI model is trained using large datasets. These datasets may contain millions or even billions of pieces of information. For example, a language model may be trained using books, news articles, websites and other textual material. Similarly, an image-generation model may be trained using large collections of photographs, drawings and other images.

WIPO explains that Generative AI systems are trained using enormous quantities of data, which may consist of freely available information as well as copyright-protected works.4
In simple terms, the AI system studies patterns in the training material. It learns relationships between words, images, sounds or other elements and uses those patterns to generate new content when a user gives it a prompt. This creates the first copyright problem. The AI developer may have to copy, download, store or process copyrighted material in order to train the model. The second problem arises when the trained model produces an output that is very similar to, or reproduces, part of an existing copyrighted work.

Therefore, there are actually two related but different questions:

Is the use of copyrighted material during AI training itself an infringement?
Can the output generated by the AI infringe the copyright in an existing work?

These questions should not automatically be treated as the same.

Copyright Protection Under Indian Law

The Copyright Act, 1957 provides copyright protection to original literary, dramatic, musical and artistic works, cinematograph films and sound recordings.5

Section 14 of the Act gives the copyright owner certain exclusive rights over the work. Depending upon the type of work, these rights include reproduction of the work, issuing copies, communicating the work to the public and making certain adaptations or other forms of exploitation.

Therefore, if an AI developer makes copies of copyrighted books, photographs or articles while collecting and preparing its training dataset, the question is whether this activity falls within one of the exclusive rights of the copyright owner.

Section 51 of the Copyright Act deals with infringement. In simple terms, copyright infringement can occur where a person, without the permission of the copyright owner, carries out an act which the Copyright Act gives exclusively to the copyright owner. 6

At first sight, therefore, copying copyrighted works for AI training may raise a genuine infringement concern. However, Section 51 cannot be read separately from Section 52.

The Importance of Section 52 and Fair Dealing

Section 52 of the Copyright Act provides certain situations where an act will not amount to copyright infringement.One of the most important provisions for the AI-training debate is Section 52(1)(a). It permits fair dealing with a work, other than a computer programme, for certain purposes, including:

private or personal use, including research
criticism or review
reporting of current events and current affairs.7

This provision is important because Indian copyright law does not provide an unlimited "fair use" defence for every socially useful or transformative use. Instead, the Act specifies particular purposes for which fair dealing may be available.

The difficult question is therefore:Can training an AI model be treated as "research" under Section 52(1)(a)? Until recently, there was no direct Indian judicial answer to this question.

The ANI Media v. OpenAI Decision

A major development came on 24 July 2026, when the Delhi High Court delivered an important interim ruling in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC.8

ANI alleged that OpenAI had used and stored its copyrighted news articles for training the large language models underlying ChatGPT without authorisation. ANI also raised concerns regarding the reproduction of its content through ChatGPT's responses.

OpenAI, on the other hand, argued that its use of publicly accessible material for training was protected by the fair-dealing exception under Section 52(1)(a)of the Copyright Act.9 The Delhi High Court declined to grant ANI an interim injunction against OpenAI. Importantly, at this stage, the Court took the prima facie view that OpenAI's storage and use of ANI's works for training its AI models could fall within the fair-dealing exception relating to research under Section 52(1)(a).10 The Court also considered the difference between using material for training and reproducing the copyrighted expression in the final output.

This is significant because it shows that, at least at the interim stage, the Court did not treat every act of copying copyrighted material for AI training as automatically infringing copyright.
However, the decision should be understood carefully. It was an interim decision and not a final determination of every copyright issue arising from AI training. The underlying litigation continues.11 Therefore, it would be incorrect to say that Indian courts have finally declared that "AI training on copyrighted works is always legal."

Why the Word "Research" Becomes Important

The interpretation of "research" under Section 52 is one of the most interesting parts of the Indian AI copyright debate. Traditionally, research might involve a student, researcher or institution studying books, articles or other material to develop knowledge. AI training is different in scale and method. An AI developer may process enormous quantities of information using automated systems rather than reading the material in the ordinary human sense.

The question is therefore whether such computational processing can still be considered research.

In ANI Media, the Delhi High Court considered the nature and purpose of AI training while examining the fair-dealing defence.12 This reasoning is important because if AI training can, in appropriate circumstances, come within "research", the use of copyrighted works for training may not automatically result in infringement. At the same time, this does not mean that every AI training activity will necessarily qualify as fair dealing. The purpose, nature of the material, manner of use and other circumstances can become relevant.

Copyright Owners' Rights and Concerns

First, their works may have considerable commercial value. A publisher may spend money and resources creating articles, photographs or other content.

Secondly, AI companies may use thousands or millions of copyrighted works to develop commercial products. Copyright owners may therefore argue that AI companies are commercially benefiting from their creative works without obtaining licences.

Thirdly, there is a possibility of market substitution. If an AI system can provide users with information that substantially reproduces a copyrighted article, users may have less reason to access the original article.

Fourthly, copyright owners may have concerns about attribution and compensation. WIPO has specifically identified questions of rights management, attribution and compensation as major issues arising from Generative AI.13

WIPO has also noted that copyright owners argue that unauthorised use of their works in AI training datasets can constitute infringement, while AI developers argue that training does not necessarily result in the preservation or distribution of permanent copies of the original works.14 Thus, there are legitimate arguments on both sides.

The Argument of AI Developers

AI developers generally argue that training is different from simply republishing a copyrighted work.

For example, when a model is trained on millions of books, the developer may argue that the model is not intended to provide users with copies of those books. Instead, the model learns statistical and linguistic patterns from the material.

WIPO has described this argument as one of the major positions in the international AI copyright debate. Developers have argued that AI training can be compared, in a broad sense, to humans learning by reading books or viewing artworks and that the model does not necessarily retain a conventional permanent copy of every work used during training.15 However, copyright owners challenge this argument by pointing out that the AI system still needs access to and may make copies of the protected works during the training process.
Therefore, the legal issue is not simply whether the final AI model contains a readable copy of the original work. The process through which the copyrighted work is accessed, copied, stored and processed may itself need to be examined.

What Does Indian Copyright Law Say About Copying?

The concept of reproduction is central to this debate. Copyright protects the expression contained in a work rather than merely the underlying idea. This principle is well established in Indian copyright law. In Eastern Book Company v. D.B. Modak, the Supreme Court explained the importance of originality in copyright law and rejected the idea that mere labour and expenditure automatically creates copyright. The Court emphasised the importance of skill, judgment and a degree of creativity.16

This principle becomes relevant to AI because not everything present in a training dataset will necessarily be copyright-protected.For example, facts themselves are generally not protected by copyright, although the particular manner in which facts are expressed may be protected.WIPO similarly explains that copyright protects expression and does not generally protect ideas, procedures, methods of operation or mathematical concepts as such.17 Therefore, an AI developer using raw facts is in a different position from an AI developer copying an author's original article, photograph or song.

Does Public Availability Mean That Copyright Disappears?

One of the important points in the AI copyright debate is that material being available on the Internet does not automatically mean that it is free from copyright. A photograph uploaded publicly on a website may still be protected by copyright. Similarly, an online newspaper article may remain protected even though anyone can read it without paying. This distinction is important because "publicly available" and "public domain" are not the same thing.
A work is not automatically free to copy merely because it can be accessed online. The ANI Media proceedings also involved this question. ANI argued that the fact that its content was publicly accessible did not give OpenAI a general licence to copy and exploit the content.18 Therefore, an AI developer cannot simply assume that everything found on the Internet is copyright-free.

When AI Training May Create a Stronger Infringement Concern

Although ANI Media provides important guidance, several situations may create greater copyright concerns. Although ANI Media provides important guidance, several situations may create greater copyright concerns.

For example, the risk may become stronger where:

copyrighted works are copied and stored without a clear legal basis

the material is used for purposes outside an applicable statutory exception
the AI system reproduces substantial portions of the original work
the model is capable of reproducing books, articles, songs or images substantially as they originally appeared
the AI output acts as a substitute for the original copyrighted work
the use is clearly outside the purpose for which a fair-dealing exception is available.

WIPO has also highlighted the problem of AI systems memorising portions of their training data and potentially producing copyright-protected material in their outputs.19 This creates an important distinction: Using a copyrighted work to train an AI model and reproducing that copyrighted work through the AI output are separate copyright questions.Even if training is ultimately considered permissible in a particular situation, that does not automatically make every AI-generated output lawful.

Can Section 52 Be Applied to Every Type of AI Training?

Section 52 does not contain a general provision saying that all technological or commercial uses of copyrighted works are fair. The provision specifically refers to purposes such as private or personal use including research, criticism or review, and reporting of current events.20 Therefore, the exact purpose and circumstances of the AI training remain important.

For example, there may be a difference between:

training a model for scientific research
training a commercial AI product
fine-tuning a model using a particular author's books
building an AI system that reproduces copyrighted works
using licensed material under an express agreement.

These situations should not automatically be treated as legally identical.

The Need to Balance Copyright and Technological Development

The central challenge is finding a balance between two legitimate interests. On one side are authors, publishers, photographers, musicians and other copyright owners who depend on copyright protection for their creative work.On the other side are researchers, technology companies and developers who need access to large amounts of data to develop new technologies. If copyright law is interpreted too broadly, technological development and research could become unnecessarily difficult.

At the same time, if copyrighted material can be freely used for every AI-related purpose without any protection for creators, the economic value of creative work could be affected. WIPO has repeatedly recognised this need to balance the interests of copyright owners with technological innovation. Its work on AI and IP specifically discusses transparency, attribution, compensation and rights management as important parts of the emerging copyright infrastructure.21

Therefore, the objective should not simply be to choose between "AI" and "copyright owners". The more practical question is how copyright law can protect creators while allowing responsible technological development.

Possible Future Solutions for India

India currently does not have a separate comprehensive statutory provision specifically regulating the use of copyrighted works for AI training. The existing Copyright Act therefore has to deal with a technology that was not contemplated when the legislation was enacted.One possible solution is greater transparency regarding training datasets. AI developers could maintain records of the broad categories and sources of material used for training.

Another possibility is licensing. AI companies could enter into agreements with publishers, music companies, image providers and other copyright owners for access to protected material.
A further possibility is an opt-out mechanism through which copyright owners can communicate that their works should not be used for AI training, subject to whatever legal framework lawmakers may ultimately adopt. WIPO's recent discussions have specifically examined emerging approaches such as opt-out mechanisms, transparency requirements, rights management and compensation systems.
India could consider these approaches while ensuring that any new framework does not unnecessarily prevent research and innovation.

Conclusion

The question of whether training Generative AI on copyrighted works constitutes copyright infringement in India does not have a simple "yes" or "no" answer.

Under the Copyright Act, 1957, unauthorised reproduction of a copyrighted work can amount to infringement under Section 51. However, Section 52 provides specific exceptions, including fair dealing for research. The important question is therefore whether a particular AI training activity falls within an applicable exception.

The 2026 decision in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC is an important development because the Delhi High Court, at the interim stage, found that OpenAI's use and storage of ANI's works for training could fall within the fair-dealing exception for research. However, the decision does not establish a blanket rule that every use of copyrighted material for AI training is lawful.23

The issue is likely to become more complicated as AI systems become capable of reproducing portions of their training data and as copyright owners become more concerned about the commercial value of their works.

India therefore faces the challenge of maintaining a balance. Copyright owners should have meaningful protection against unauthorised exploitation of their creative works, while AI developers and researchers should have sufficient legal space to develop useful technologies. Ultimately, Indian copyright law may need to evolve with Generative AI.

Whether this happens through judicial interpretation, licensing mechanisms, greater transparency, statutory amendments, or a combination of these approaches remains an important question for the future of Indian intellectual property law.

REFERENCES

[1] World Intellectual Property Organization (WIPO), Generative AI: Navigating Intellectual Property (2024), explaining that Generative AI systems are trained using enormous amounts of data, including potentially copyright-protected works.

[2] WIPO, WIPO Conversation – Generative AI, noting the competing positions of AI developers and copyright owners concerning the use of copyrighted works as training data.

[3] ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, CS(COMM) 1028/2024, I.A. 45300/2024, Delhi High Court, decided 24 July 2026.

[4] WIPO, Generative AI: Navigating Intellectual Property (2024).

[5] Copyright Act, 1957, Sections 13 and 14.

[6] Copyright Act, 1957, Section 51.

[7] Copyright Act, 1957, Section 52(1)(a).

[8] ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, CS(COMM) 1028/2024, Delhi High Court, 24 July 2026

[9] WIPO, Learning Machines: An Introduction to AI and IP for Small and Medium-sized Enterprises, discussion on copyright-protected training material and AI systems.

[10] ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, Delhi High Court, 24 July 2026.

[11] The ruling concerned interim relief and does not finally dispose of all issues raised in the underlying copyright suit.

[12] Ibid

[13] WIPO, Artificial Intelligence and Intellectual Property, discussing the growing importance of rights management, attribution and compensation in the context of Generative AI.

[14] WIPO, WIPO Conversation – Generative AI.

[15] Ibid.

[16] Eastern Book Company & Ors. v. D.B. Modak & Anr., (2008) 1 SCC 1.

[17] WIPO, Frequently Asked Questions: Copyright, explaining that copyright protects expression and does not generally extend to ideas, procedures, methods of operation or mathematical concepts.

[18] ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, Delhi High Court, 24 July 2026

[19] WIPO, Intellectual Property Rights in Healthcare-related AI, discussing the possibility of AI models memorising portions of training data and reproducing copyright-protected material

[20]Copyright Act, 1957, Section 52(1)(a).

[21] WIPO, Artificial Intelligence and Intellectual Property, particularly its discussion of copyright infrastructure, rights management, attribution and compensation.

[22] The WIPO Conversation on Intellectual Property and Frontier Technologies, discussions concerning transparency and opt-out mechanisms in relation to AI training data.

[23] ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, Delhi High Court, 24 July 2026.