Authored By: Bhavya Anand
Manipal University Jaipur
Introduction
In November 2020, Indian Copyright Office broke new ground when it registered the AI software program “RAGHAV” as a co-author of an artwork “Suryast” along with its human user. However, in November 2021, it published a notice withdrawing the said registration. At around the same time, yet another matter of accountability was brought before the Delhi High Court, whether OpenAI could be sued for using copyrighted news reports from Asian News International to train ChatGPT without permission, and reproducing excerpts of the material in ChatGPT’s responses.[1]
The first and the second episode are two extremes of the same problem. The first question is who the author can be when a machine has helped create; the second question is who is responsible when a machine’s creations violate someone else’s work. Neither question can be answered with a definite answer, since the Copyright Act, 1957 were written for the era of human authors and identifiable infringers, not for generative systems trained on a large document corpus and able to create new expressions on demand.
This piece of writing suggests that Indian copyright law today does not have a satisfactory system for distributing the liability of infringement and authorship in the AI value chain, and instead of patching together regulations as ad hoc fixes, it is essential that Indian copyright law be targetedly improved.
The article will be divided into five sections. Part II outlines the current legal landscape for authorship, infringement and intermediary liability. Part III covers the Suryast affair and the looming lawsuit between the two against ANI and OpenAI. The resulting gaps are critically reviewed in Part IV. Part V compares the EU, UK and the United States. The proposed agenda for reform of Indian law is presented in Part VI.
The Existing Legal Framework
Authorship and Originality
Copyright exists in original literary, dramatic, musical and artistic works as per section 13 of Copyright Act, 1957.[2] The definition of the term “author” can be found in section 2(d). Sub-clause (vi) introduced in 1994 to deal with developments in computer technology states that for “computer-generated” work, the author is “the person who causes the work to be created. [3] This requirement was made concerning the systems, whose output consists of the algorithms’ output, such as macros for Excel or compiled databases, because their output is predictable since they were configured by a person. This requirement was made concerning the system, which cannot do something unpredictable; its output is a result of statistical prediction of the entire corpus. Originality, on the other hand, must be evaluated through the test of “modicum of creativity,” a standard set by the Supreme Court in the case of Eastern Book Company vs D.B. Modak where it overturned the earlier standard of “sweat of the brow.”[4] Whether the sheer effort of engineering a detailed prompt satisfies the legal requirement for human creativity remains an open question. As of now, Indian courts have yet to test or define the legal boundaries of AI-generated ownership.
Infringement and Fair Dealing
Under Section 14, the copyright proprietor has the exclusive right to reproduce, adapt, and communicate a work to the public. Infringement is defined in section 51 and section 52 provides exceptions for fair dealing, such as for private use, research, or criticism/review. India has no text and data mining exception as explicit in the text of its copyright law – does the ingestion of copyrighted text into a training corpus fall under the definition of “reproduction” and can any use that comes out of it be said to be protected by Section 52? – are these the questions the Delhi High Court is facing.
Intermediary Safe Harbour
Under Section 79 of the Information Technology Act, 2000, intermediary is not liable for the action taken by any third party as long as it is acting in good faith and has not actually known of the illegal act. The Information Technology Rules, 2021, and the advisory issued by Ministry of Electronics and Information Technology on 1 March 2024, on labelling and due-diligence obligations for generative AI services, but the safe harbour rule was based on an assumption that it was a conduit of third party material, rather than a machine that generates new expressive output from existing material.
III. Case Law and Regulatory Analysis
The Suryast Episode
“The surreal” is a painting that combines a portrait of Ankit Sahni (who has also painted various other portraits in this series) with a transfer of style inspired by Van Gogh’s Starry Night, which was created by Ankit Sahni with RAGHAV, an AI painting application. He submitted his first, where the sole author was RAGHAV, which was rejected. In November 2020, a second application was registered that names Sahni and RAGHAV as co-authors; so far, this is the first case in the world where a copyright office has recognised an AI system as a co-author. [5] In November 2021 the Copyright Office sent Sahni a notice of withdrawal stating that, in its opinion, the Copyright Office was directed to clarify RAGHAV’s legal status and that section 2(d)(iii) and (vi) of the Act required a natural person.[6] But Sahni has fought the withdrawal, saying the Act does not provide any mechanism to make a granted registration “re-open.[7] The registration is still on the list, the withdrawal is not finalized, and there is still no ruling on the question whether an AI system can be called co-author. The episode is more of a signal that the Copyright Office, when faced with an AI co-author, reverts to a human being, than a precedent.
ANI Media Pvt Ltd v OpenAI OpCo LLC
In November 2024, Asian News International filed a lawsuit against OpenAI, claiming that it used its copyrighted news articles without a licence in training ChatGPT and that the answers it generated from the platform copied parts of its content. The main justification for OpenAI has been a jurisdictional argument – that its servers and training processes are not located within India, so this Copyright Act does not apply, and the Delhi High Court has no territorial jurisdiction. On the merits, OpenAI has contended that it does not reproduce the content in the AI’s responses, that news content is not expressive or therefore subject to copyright, and that copyright protection for factual reporting is too narrow to cover ANI. With the help of two amici curiae, the Court asks four questions: First, does the use of the copyrighted material to train a LLM constitute infringing on the copyright? Second, does the use of the copyrighted material to generate a response infringe? Third, does either use fall under the “fair dealing exception” in section 52 of the Copyright Act? Fourth, does the Court have jurisdiction? The Federation of Indian Publishers and the Digital News Publishers Association have stepped in, as have other industry groups, highlighting the importance of this to the broader publishing community. At this time, judgment has also been kept, with the outcome likely to have an impact on the prosecution and defence sides of generative AI liability in court proceedings.
Critical Evaluation
This survey reveals two types of accountability deficits. The first is creation: section 2(d)(vi) was never designed to address who is the “causing” party of the creation of a generative output by a developer’s training of the model, a deployer’s fine-tuning and hosting of it, and a user’s prompt to trigger a specific output. The Suryast withdrawal indicates the Copyright Office’s bias toward a human author, but there’s no test for distinguishing between a user who shows creativity vs. a user who just presses a button. While this is not the case, the applicant and the examiner are left guesswork.
The second gap is with regards to infringement, which is more significant as it involves both ends of the AI pipeline. On the input side, it is unclear whether the ingestion of copyrighted text for training is a “reproduction” within the meaning of section 14, and whether any of the exceptions to that section (none of which was drafted with machine learning applications in mind) extends to cover it. There is some truth in OpenAI’s non-expressive use argument, copyright is about expression, not facts or ideas that the model might derive from a text. If accepted on its own terms however, then it would effectively render ingestion immune to the rights holder, regardless of the size of the intake, if there was no exact match of a passage in any single output – meaning that even if a particular rights holder’s text is used for a particular model, that model would not protect them, because it would not protect any other rights holder.
The liability is not as clear on the output side. If a generative system generates a substantial amount of text that is like a copyrighted news report, the immediate infringer may be the developer who created the model, the deployer who made it commercially available or the user who caused the specific output by prompting the generative system. Section 79’s safe harbour, which applies to platforms that passively host or transmit third-party content, appears to be a poor fit for a system that is actively creating new expressive content using the third-party material, and can only reasonably be said to be merely a conduit.
Comparative Perspectives
The European Union
The European Union has directly tackled the issue of inputs. The Digital Single Market Directive introduces a ‘text and data mining exception’ that allows the reproduction of copyrighted works already lawfully accessible to the public for the purposes of training, with the right for the rights holder to object to commercial text and data mining. The AI Act imposes transparency requirements on general-purpose AI service providers, such as a summary of the content of the training process, so as to make enforcement easier without prohibiting training itself.
The United Kingdom
United Kingdom has been more conservative. The Copyright, Designs and Patents Act, 1988 already considers computer-generated works and attributes the authorship to the person who made the arrangements necessary for the creation of the work, which is like the Indian concept of computer-generated work in section 2(d)(vi). After being met with some publisher resistance, the government’s plans for a blanket exception for “text and data mining” were stalled, and UK law is left to rely on current licencing practice rather than an exception set out in legislation.
The United States
The United States has taken the approach of litigation, not legislation, to creation. In Thaler v Perlmutter, the District Court for the District of Columbia affirmed the Copyright Office’s refusal to register the work created by an autonomous AI system and found that copyright law requires human authorship. Meanwhile, an AI training data lawsuit by news publishers is pending considering its potential as a bellwether for the input-side issue and is also similar to ANI’s theory.
Proposals for Reform
First, Parliament should clarify the requirement for human creative control in the creation of an AI-assisted work in section 2(d)(vi) of the Bill, since there is a clear desire in the Copyright Office to have a human as the author, and a need for a test that is not left to the arbitrary decisions of withdrawal notices.
Secondly, the Act should include a clearly formulated exception for reproduction for training AI models, based on the EU model, along with an opt-out option for rights holders, and for large-scale commercial training, a collective license system that directs compensation to the authors and publishers whose works are incorporated. This would clear the doubt on the input side at the Delhi High Court and would not deprive the rights holders of getting any compensation.
Thirdly, liability for infringing outputs should be based mainly on the deployer most likely to be able to make a generative system commercially available, on a standard of ‘reasonable technical measures’ to prevent ‘substantial reproduction’ of identifiable copyrighted material, and only subject to the section 79 safe harbour where such measures are shown to be in place. This shifts the burden of filtering outputs from the claimant to the party most suited to doing so, not the developer, deployer or end user.
Fourth, the developers of general-purpose AI systems sold in India should be obligated to provide a summary of the sources from which the training data was sourced, allowing the rights holders to find out whether their works have been used, and to assert their rights and/or withdraw their consent to the use of their works.
VII. Conclusion
The Suryast pullout and the pending ANI case are manifestations of the same underlying structural issue: India’s copyright regime assumed the presence of a human author as well as an identifiable infringer, and AI-based generative technology undermines both premises simultaneously. As this paper contends, there is simply no way to bridge the accountability void thus created through judicial rulings or regulation.
The reforms recommended herein – a clearer threshold for authorship, a defined and paid training data exception, output liability focused on the deployer, and mandatory training data transparency – offer the courts, the Copyright Office, and AI developers a framework in lieu of the hodgepodge that includes a vague 1994 statute and an open case. Although the Delhi High Court will eventually rule on ANI v OpenAI and thus resolve the immediate dispute, only legislative reform can address the issue revealed by the case: not just whether a machine infringes, but whom to blame when it does.
Reference(S):
Cases
ANI Media Pvt Ltd v OpenAI OpCo LLC CS(COMM) 1028/2024 (Delhi HC, filed 19 November 2024).
Eastern Book Company v D B Modak (2008) 1 SCC 1.
Thaler v Perlmutter 687 F Supp 3d 140 (DDC 2023).
Legislation
Copyright Act 1957 (India).
Information Technology Act 2000 (India).
Copyright, Designs and Patents Act 1988 (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 (AI Act) [2024] OJ L1689.
Secondary Sources
Altacit Global, ‘AI-Generated Works in India: Examining the “Raghav” Signal in Copyright Law’ (Medium, January 2026).
Sukanya Sarkar, ‘Exclusive: Indian Copyright Office Issues Withdrawal Notice to AI Co-Author’ Managing IP (14 June 2022).
‘ChatGPT Data Not Stored in India, Copyright Act Does Not Apply: OpenAI to Delhi High Court in ANI Suit’ Bar & Bench (2 April 2025).
‘News Reports Have Limited Copyright Protection: OpenAI to Delhi High Court in ANI vs ChatGPT Case’ Bar & Bench (22 April 2025).
‘Hearing in Copyright Case Against OpenAI Brought by Indian Publishers Set for March’ Tech Policy Press (14 February 2025).
‘Delhi High Court Reserves Judgment in ANI v OpenAI’ The New Publishing Standard (1 April 2026).
Ministry of Electronics and Information Technology, Advisory on Due Diligence by Intermediaries/Platforms under the IT Rules 2021 in Dealing with Artificial Intelligence Tools (1 March 2024).
[1] ANI Media Pvt Ltd v OpenAI OpCo LLC CS(COMM) 1028/2024 (Delhi HC, filed 19 November 2024).
[2] Copyright Act 1957, s 13(1)(a) (India)
[3] Copyright Act 1957, s 2(d)(vi) (India).
[4] Eastern Book Company v D B Modak (2008) 1 SCC 1.
[5] Altacit Global, ‘AI-Generated Works in India: Examining the “Raghav” Signal in Copyright Law’ (Medium, January 2026).
[6] Sarkar (n 1).
[7] ibid.
