ANI v. OpenAI is India’s first copyright challenge against generative AI. Filed in 2024, it has since drawn in other industry groups, reflecting its wide impact.
As on date, the Amici Curiae, plaintiff and defendant have nearly completed their submissions (presumably on the interim application). Unfortunately, these submissions are not available to the public. The case is listed next on May 16, 2025.
Generally, the case is about copyright infringement for training AI models and for AI-generated output. One side argues that AI companies infringe copyright by reproducing news content without permission; the other, contends it’s either non-expressive use or protected by fair dealing. There are no clear answers which has caused some worry, much confusion and to lawyers like me – a lot of excitement (my previous piece here). Bloggers on SpicyIP have also covered this case across several posts here.
Given the law is unsettled, it may be helpful to reframe the key issues to better navigate tensions between AI generated content and publishers/creators. Revisiting Indian copyright law’s treatment of ‘derivative works’ and ‘transformation’ could offer a path for addressing these challenges.
Indian ideas of ‘derivative works’ and ‘transformation’
The Delhi High Court has framed the following issues on the interim injunction application. The problem has been dissected into three distinct parts:
(a) whether unauthorised reproduction/storage of copyrighted content for training AI is permitted,
(b) whether unauthorised use of copyrighted content for generating AI content is permitted, and
(c) though unauthorised, whether (a) and (b) are saved as ‘fair use/dealing’.
“25. The present suit raises novel legal issues arising on account of recent technological advancements. At this stage, some of the key issues that warrant consideration include the following:
I. Whether the storage by the defendants of plaintiff’s data (which is in the nature of news and is claimed to be protected under the Copyright Act, 1957) for training its software i.e., ChatGPT, would amount to infringement of plaintiff’s copyright.
II. Whether the use by the defendants of plaintiff’s copyrighted data in order to generate responses for its users, would amount to infringement of the plaintiff’s copyright.
III. Whether the defendants’ use of plaintiff’s copyrighted data qualifies as ‘fair use’ in terms of Section 52 of the Copyright Act, 1957.
IV. Whether the Courts in India have jurisdiction to entertain the present lawsuit considering that the servers of the defendants are located in the United States of America.”
This framing helps publishers as I discuss in my piece here. The problem for OpenAI is that if the first issue is decided against them, which is likely, it may not be easy for them to win for reasons discussed here and here.
Instead, if the issues were reframed, the outcome may be different. The court will be able to resolve the issue more ‘fairly’ by removing the first issue on reproduction for training, and focusing only on the second issue i.e. whether AI generated output is infringing. This way the court will be able to concentrate on the ‘misuse’ of copyrighted material and not merely on their use. (see Rebecca Tushnet who asks these questions here) In fact, the second issue addresses the first and third issues and is backed by the Supreme Court and by legislative intent.
The Indian Supreme Court, in EBC v Modak, has held that the mere use (which includes reproduction) of a prior work is not itself infringement. While this case in better known for establishing standards of ‘originality’, it is important for establishing the Indian concept of ‘derivative works’.
In EBC, the court classifies literary works into – primary works or ‘prior works’ and derivative works or ‘secondary works’. The court holds that ‘in many cases, a work is derived from an existing work. Whether in such a derivative work, a new copyright work is created, will depend on various factors.’ If these factors of originality are satisfied ‘a new work is created making the creator of the derivative work its author’.
Therefore, an existing work can be reproduced and transformed into something different without attracting copyright liability in India. In such cases, initial reproduction becomes irrelevant i.e. the first issue; and fair dealing doesn’t arise since a new copyrightable work is created – addressing the second issue. This idea is central to copyright law in India as it makes ‘independent creative expression’ possible.
This logic can apply to AI models. Today, AI models like ChatGPT are presumably trained on existing copyrighted works to create ‘derivative works’. AI generated content can be creative and human-like. AI-generated content may qualify as non-infringing ‘derivate works’ – if they are not substantially similar to existing works – and should qualify for an assessment of copyright protection. The question of authorship of AI generated content – whether of the human prompter, the AI model, or author-less, etc. is the next question, to be decided separately.
As a result of EBC, reproduction of existing copyrighted works for training AI models which results in a ‘transformed’ output, may not be considered infringing.
Legislative intent also supports this argument. Copying for technical/hidden uses is not considered infringement in several contexts. The legislature has designed fair dealing provisions to enable the smooth functioning of the internet and to facilitate ‘technical copying’ of works. For example, copies for the purpose of caching or linking are not infringing. For output related issues, Section 52(1)(c) establishes a notice-and-take-down mechanism and not a blanket ban. This reflects a legislative preference for negotiated, complaint based responses over immediate statutory liability.
With such reframing, there may be no need for a radical re-interpretation of Section 52 (fair dealing) and copyright holders will be compensated for AI-generated output that is substantially similar to their works.