Part I examined the Court’s transformative approach to fair dealing, its recognition of Section 52 as embodying user rights, and its emphasis on public interest. Part II turns to the remaining issues: whether ChatGPT memorized or substantially reproduced ANI’s works, and whether the Delhi High Court rightly assumed territorial jurisdiction.
No memorization, regurgitation or substantial reproduction – Issue II
The Court split Issue II into two sub-issues [Para 75]:
- whether Open AI memorizes and regurgitates ANI’s copyrighted literary works in the form of responses and,
- whether the ChatGPT responses are substantial reproductions of ANI’s copyrighted literary works.
The Court observed that ChatGPT could not have memorized (i.e., storage + reproduction of specific pieces of information from its training data) any of ANI’s works as all of such works in the dispute were published after ChatGPT’s training cut-off of that period – so for context, LLMs are trained periodically, and the model accessed by ANI in the documents produced had been last updated before ANI had published the works which it alleged OpenAI to infringe. Thus, the Court stated that it is practically not possible for ChatGPT to have memorized those works. [Paras 83, 84].
However, though not through training, ChatGPT had accessed those works through its web-search functionality using Retrieval Augmented Generation (RAG) technology [Para 84] (see my earlier explanation of RAG technology here.
The Court now examines whether ChatGPT’s outputs were substantial reproductions of ANI’s works; it observes that there is no substantial similarity between them [Para 114-115]. In the context of this analysis, the Court, relying on B.D Bhandari (DHC 2011), recognised the application of merger doctrine given that news which are facts can be expressed in limited number of ways and in such scenarios where the “facts and the expression would become so intrinsically connected that it would be difficult to distinguish one from another” would make “the expression itself noncopyrightable.” [Paras 95-97]
The Court’s reading of the merger doctrine is puzzling. After discussing the doctrine, it concludes only that there is no copyright in facts and that “in the context of news, copyright would subsist only in the form and manner of expression of news and not on the underlying facts” (Para 97). But this merely restates the fact-expression dichotomy. The merger doctrine applies where facts and expression are so inseparable that the expression itself becomes uncopyrightable. As Prof. Scaria argued, this would imply that news receives either no copyright protection or only a thin layer of protection (Para 66.4). Instead, the Court proceeds with a conventional substantial similarity analysis, making its discussion of the merger doctrine appear unnecessary.
The Court then proceeds to address ANI’s reliance on a few foreign cases, a notable one includes GEMA v. OpenAI (2025), a German case where OpenAI was held to be liable for memorizing and reproducing lyrics of a song. The Court distinguished GEMA (2025) on the grounds already discussed above regarding memorization and substantial reproduction. It further noted that, unlike the non-adversarial prompts used in GEMA, ANI had employed “fairly detailed, repeated prompts” including the instruction “exactly” which the Court characterised as adversarial prompts (Para 119). [ANI’s follow-up prompt – “Prompt : Please try to tell me exactly what she said” – owing to its usage of “exactly” qualifies as an adversarial prompt per DHC.]
It may be a stretch to argue that the usage of the word ‘exactly’ would make a prompt an adversarial prompt which requires explicit instructions to bypass guardrails or wear down the guardrails via repeated prompting (a few examples of adversarial prompts can be seen here). Well, so what if adversarial prompts were used? Do these observations imply that a relevant factor for Courts to consider in such cases is also whether adversarial prompts were used to elicit infringing responses?
What happens if only adversarial prompts were used to elicit infringing responses, based on which a suit is filed? Who is liable then? In a different piece, I have argued that in cases where there exist sufficient guardrails that are being circumvented by users using adversarial prompts, the liability must not lie with the AI developers but with the users (as, apart from AI chatbots having substantial non-infringing uses, akin to intermediaries, AI developers have done their due diligence and it is the users deliberately circumventing it). In cases where chatbots reproduce copyrighted material verbatim (or those bearing substantial similarity to copyrighted works) for simple prompts (without extractive techniques employed by the users) the liability should vest with the AI developer.
DHC has jurisdiction (well, when does it not?) – Issue IV
The Court holds that the DHC has territorial jurisdiction based both on the CPC (Section 20), and the Copyright Act [Section 62(2)]. Jurisdiction flowing from Section 62(2) of the Copyright Act is fairly obvious given ANI’s principal place of business is in Delhi. However, the Court offers little explanation for how territorial jurisdiction is attracted under the CPC. It merely notes that OpenAI targets Indian users, including those in DHC’s jurisdiction, and ChatGPT’s responses based on ANI’s prompts took place within DHC’s jurisdiction (Paras 46-47). To the extent these facts are treated as constituting part of the cause of action under Section 20(c), the judgment appears to adopt an expansive understanding of territorial jurisdiction, where territorial jurisdiction is conferred based on where a defendant markets and a plaintiff prompts their chatbot! Such an interpretation would have wide ramifications on forum shopping in future AI-related disputes.
On OpenAI’s objection of Indian copyright law being inapplicable owing to its servers being located in the US, the Court observed that OpenAI servers being located in the US is a “terminal step in the chain of events which begin from access of copyrighted works from India and transmission of the same abroad.” The implications of this observation is unclear, is the Court saying that the location of the servers is only the final link in a larger chain of acts that begins in India, so the foreign location of the servers alone does not defeat the jurisdiction of Indian courts? However, the copyrighted works were accessed online and not “from India” per se, while the alleged acts of copying and training occurred on servers located in the US, thus this reasoning is open to question.
The Court applied the reasoning in Neetu Singh v Telegram (DHC 2022) – the mere fact that a defendant locates their servers outside India “cannot divest the Indian Courts from dealing with copyright disputes or divest copyright owners from availing their remedies in Indian Courts.” (Para 46 of Neetu Singh) [Paras 48-55 of ANI v OpenAI]
That said, the reliance on Neetu Singh (DHC 2022) ought not to be overstated. That decision arose in a different context, where the DHC directed Telegram to disclose the Basic Subscriber Information (BSI) of alleged infringers despite the data being stored on foreign servers. It did not determine whether Indian courts possessed territorial jurisdiction over acts of copyright infringement allegedly committed abroad. The issue there was the enforceability of procedural directions against an intermediary, rather than the substantive question of territorial jurisdiction over acts of copyright infringement.
Conclusion
The ANI v. OpenAI judgment marks a significant shift in Indian copyright jurisprudence. By recognising Section 52 as embodying user rights, adopting a technologically informed understanding of LLMs, and foregrounding public interest, the DHC has reaffirmed that copyright exists not merely to protect authors, but also to facilitate access to knowledge and innovation. The judgment lays down a principled framework that resists copyright maximalism while preserving the balance at the heart of the Copyright Act.
I would like to thank Swaraj Barooah and Praharsh Gour for their valuable inputs and review of this post.
