ANI v. OpenAI Judgment Explained: Delhi High Court on ChatGPT, AI Training & Copyright in India

Contributor: Adv. Anish Sinha

Introduction & Background of ANI v. OpenAI

On 24 July 2026, a Single Judge of the Delhi High Court, Justice Amit Bansal, delivered judgment in ANI Media Pvt. Ltd. v. Open AI Opco LLC, dismissing an interim injunction application filed by Asian News International (“ANI”) against Open AI Opco LLC, the entity behind ChatGPT.[1] The suit, registered as CS(COMM) 1028/2024, was the first infringement action instituted in India against a large language model (“LLM”) developer, with summons issued in November 2024. The judgment, running to 135 pages, addressed four principal issues: territorial jurisdiction, infringement by output generation, infringement by training-data storage, and the applicability of the fair-dealing exception under Section 52(1)(a) of the Copyright Act, 1957.

The proceedings drew an unusually wide field of intervenors, including the Indian Music Industry, the Federation of Indian Publishers, the Digital News Publishers Association, the Indian News Media Group, Broadband India Forum, and Flux AI Labs, together with two Amici Curiae that were Professor Arul George Scaria of the National Law School of India University and Advocate Adarsh Ramanujan.[2] ANI’s case, in essence, was that ChatGPT reproduced its copyrighted news content verbatim in response to user queries, that this reproduction occurred because OpenAI’s models memorised and regurgitated the training data, and that OpenAI’s storage of ANI’s articles for training purposes was itself an act of infringing reproduction not saved by any statutory exception.

The Jurisdictional Threshold (ANI v. OpenAI)

Since the objection went to the root of the matter, the Court addressed jurisdiction first. Two sub-questions arose, which were whether the Delhi High Court possessed territorial jurisdiction at all and whether the Copyright Act could apply to training activity that Open AI contended occurred exclusively on servers located in the United States.[3] On the first sub-question, the Court found it undisputed that ANI’s registered office and principal place of business lay within the Court’s jurisdiction and that Open AI targeted and offered its services to users in India, including within Delhi, satisfying Section 62(2) of the Copyright Act read with Section 20 of the Code of Civil Procedure.[4]

On the second sub-question of whether the extraterritorial location of training servers ousted the Court’s jurisdiction over the training claim, the Court relied on the Division Bench ruling in Blueberry Books, where a similar objection raised by Amazon regarding infringing conduct occurring in the United States had been rejected.[5] The Court also drew on the reasoning in Neetu Singh, concerning Telegram’s choice to locate servers in Singapore, where a Coordinate Bench had held that the mere location of servers abroad could not leave Indian copyright owners without a remedy in an era of cloud computing and diminishing national boundaries in data storage.[6]

Applying this reasoning, Justice Bansal held that the storage of ANI’s works on US servers was merely “a terminal step in the chain of events” that began with the scraping of copyrighted material accessible from India, and that severing the analysis to examine only the final step would permit evasion of Indian copyright law by relocating servers abroad.[7] On this prima facie footing, the Court answered the jurisdictional issue in favour of ANI.

Memorisation, Regurgitation, and Substantial Similarity of Outputs

The second issue concerned whether ChatGPT’s outputs, generated in response to user prompts, amounted to infringement. ANI’s central submission was that LLMs inherently memorise training data through a process of tokenisation and detokenisation and that this data is regurgitated, sometimes verbatim, when the model generates a response, thereby retaining and reproducing the expressive elements of ANI’s articles.[8] OpenAI countered that LLMs are not designed or trained to reproduce extracts of training data, that identical prompts submitted by different users would elicit different responses, and that the model instead generates novel text based on patterns learned from voluminous data rather than reciting stored content.[9]

The Court examined comparative jurisprudence on this question, particularly the reasoning of the United States District Court for the Northern District of California in Bartz v. Anthropic, where it was noted that even adversarial prompting by the plaintiffs’ own expert could extract no more than a small number of words from the underlying copyrighted books, and that the model concerned could not be used to “read or otherwise meaningfully access” the works in question.[10] Professor Scaria’s written submissions, extracted by the Court, similarly observed that empirical literature places the incidence of memorisation in a small fraction of cases, occurring predominantly during deliberate “extraction attacks” using adversarial prompts, and that the majority of LLM outputs constitute non-expressive use of copyrighted material.[11]

Turning to the specific illustrations pleaded by ANI, the Court applied the classical test for substantial similarity laid down in R.G. Anand, comparing ANI’s articles as a whole against ChatGPT’s responses and found no substantial similarity, and noting, for instance, that even the titles generated by ChatGPT differed materially from ANI’s headlines.[12] The Court also distinguished the German decision in GEMA v. Open AI, where memorisation of song lyrics had been established on the basis of non-adversarial prompts and in the present case, by contrast, ANI’s prompts were found to be detailed, repeated, and expressly designed to extract content using terms such as “exactly,” rendering them adversarial in character and even such prompting failed to produce a substantial reproduction.[13]

The Court further noted that the illustrations pleaded in the plaint post-dated the completion of Open AI’s training, meaning that no case of memorisation could be made out on that material at all.[14] On this basis, the Court recorded its prima facie finding that ANI had failed to establish either memorisation and regurgitation, or substantial reproduction of its works through ChatGPT’s outputs.[15]

Storage of Training Data as an Act of Reproduction

The third issue examined whether OpenAI’s storage of ANI’s articles for training purposes independently infringed the right of reproduction under Section 14(a)(i) of the Copyright Act. The provision, since its 1994 amendment, expressly extends the right of reproduction to “the storing of it in any medium by electronic means,” a phrase the Court read as deliberately broad, covering both temporary and permanent storage regardless of the storer’s underlying purpose.[16] ANI and the intervenors supporting it argued that tokenisation and vectorisation of raw data necessarily involve copying, that this copying persists in the resulting vector database, and that the entire training pipeline therefore falls within Section 14(a)(i), irrespective of whether OpenAI ultimately reproduces the expression to end users.[17]

The Court accepted, at least for the purposes of Section 14(a)(i), that storage of ANI’s literary works during the training process would prima facie constitute an act of reproduction. This finding, however, did not conclude the enquiry, since Open AI had pleaded the exception under Section 52(1)(a) of the Copyright Act as a complete defence to what would otherwise amount to infringement, as an issue the Court proceeded to examine jointly with the third issue framed in the suit.

The Section 52(1)(a) Fair-Dealing Defence

The fourth and most substantial issue concerned whether Open AI’s storage of ANI’s works qualified as fair dealing under Section 52(1)(a) of the Copyright Act, which exempts “fair dealing with a literary… work… for the purposes of private or personal use, including research.” The Court traced the legislative history of the provision, noting that it originally referred only to “research or private study” and was amended in 1994 to its present form of “private use, including research,” a change the Court treated as widening rather than narrowing the exception’s scope.[18] The Court held that the word “research” must be read broadly and liberally, and rejected the contention advanced by ANI that machine training could never fall within “private use” merely because the resulting model is subsequently deployed for commercial purposes through ChatGPT.[19]

The Court structured its analysis of Section 52(1)(a) around two components: a “purpose test” and a “fairness test.”[20] Under the purpose test, the Court examined whether Open AI’s use of ANI’s copyrighted material was confined to training its models, without any accompanying reproduction or communication of the original works to the public. As a factual conclusion following directly from the Court’s findings on the second issue, where memorisation and regurgitation had not been established.[21] The Court further invoked the “transformative use” doctrine developed in American jurisprudence, including Bartz v. Anthropic, to hold that training an LLM on copyrighted text serves a fundamentally different purpose from the original works themselves, and does not displace demand for the underlying news articles.[22]

Under the fairness test, the Court weighed whether Open AI’s use resulted in economic competition prejudicial to ANI’s legitimate interests. It found that ANI’s commercial model rests on subscription licensing and advertising revenue generated from user traffic, whereas ChatGPT performs a materially different function, like serving as a general-purpose tool for research, translation, summarisation, and content creation, only incidentally touching upon news queries, and typically returning summaries or snippets accompanied by source attribution rather than a substitute for ANI’s syndicated feed.[23] The Court also placed weight on the broader public interest served by generative AI across sectors such as education, healthcare and agriculture, referencing a NITI Aayog policy paper on AI for inclusive societal development, before concluding that both the purpose test and the fairness test stood satisfied, such that Open AI’s storage of ANI’s works fell within Section 52(1)(a) and did not amount to infringement.[24]

Balance of Convenience and the Refusal of Interim Relief

Having found no prima facie case of infringement on any of the three substantive issues, the Court nonetheless proceeded, in the alternative, to examine the balance of convenience and irreparable injury. Several factors informed this analysis. First, the Court noted that ANI possessed and had not exercised the ability to block web crawlers from accessing its website, a step Open AI stated it had itself taken with respect to ANI’s site for both training and retrieval-augmented generation (“RAG”) purposes.[25] Second, the Court observed that ANI had, by a communication dated 3 October 2024, offered Open AI a licence for its content at a fee of USD 7.5 million, a fact the Court treated as demonstrating that ANI’s claim was capable of quantification and compensation in damages, whereas an injunction against Open AI would cause harm not readily reducible to monetary terms.[26]

Third, the Court invoked public interest as the fourth and independent factor relevant to interim relief in intellectual property matters, citing Zydus Lifesciences v. E.R. Squibb, F. Hoffmann-La Roche AG v. NATCO Pharma, and Astrazeneca AB v. Intas Pharmaceuticals, and held that an injunction would adversely affect the millions of Indian ChatGPT users, many of whom are non-paying subscribers, and would be detrimental to the development of LLMs in India more generally.[27]

On this cumulative reasoning, the Court held that the balance of convenience lay against the grant of an interim injunction, and that irreparable injury would be caused both to Open AI and to the public at large were such an injunction to be granted.[28]

Conclusion

The Court’s final conclusion recorded three prima facie findings that firstly Open AI’s storage of ANI’s works for training fell within Section 52(1)(a) of the Copyright Act and did not amount to infringement under Section 51, secondly that ChatGPT’s outputs, including those generated through RAG, were not substantially similar to ANI’s works and therefore did not infringe and thirdly that ANI had failed to establish memorisation or regurgitation of its works in ChatGPT’s responses.[29] Accordingly, I.A. 45300/2024 was dismissed, with the Court expressly clarifying that its observations were confined to the disposal of the interim application and would have no bearing on the final adjudication of the suit after trial.[30]

The judgment is, by its own terms, a prima facie ruling rendered at the interlocutory stage, and the underlying suit, including the disputed questions of fact concerning training methodology and the scope of “research” under Section 52(1)(a), remains to be tried on evidence.

[1]ANI Media Pvt. Ltd. v. Open AI Opco LLC, CS(COMM) 1028/2024, I.A. 45300/2024, Judgment dated 24 July 2026 (Delhi High Court) (“Judgment”), Cause Title; Judgment 271–273.

[2]Judgment, Cause Title, listing intervenors including Indian Music Industry (I.A. 4027/2025), Federation of Indian Publishers (I.A. 470/2025), Digital News Publishers Association (I.A. 2199/2025), Broadband India Forum (I.A. 10766/2025), and Flux AI Labs (I.A. 4615/2025); Amici Curiae Professor Arul George Scaria and Advocate Adarsh Ramanujan.

[3]Judgment 43–44.

[4]Judgment 45–48.

[5]Judgment 49–50, citing Blueberry Books (Division Bench, Delhi High Court).

[6]Judgment 52–53, citing Neetu Singh (Delhi High Court) ¶ 46.

[7]Judgment 51, 54.

[8]Judgment 57.1–57.3, 76.

[9]Judgment 77.

[10]Judgment 80, citing Bartz v. Anthropic, U.S. District Court, Northern District of California.

[11]Judgment 78, extracting the written submissions of Professor Arul George Scaria.

[12]Judgment 113–115, applying R.G. Anand (Supreme Court of India).

[13]Judgment 117–119, distinguishing GEMA v. Open AI (Munich Regional Court).

[14]Judgment 124.

[15]Judgment 124–125.

[16]Judgment 145–149.

[17]Judgment 127.1–127.4, 128.1–128.3.

[18]Judgment, discussion at “Legislative History of Section 52(1)(a).”

[19]Judgment, discussion under the interpretation of Section 52(1)(a) of the Copyright Act.

[20]Judgment, Index at “I. Purpose Test” and “II. Fairness Test,” pp. 96, 115 of the Judgment.

[21]Judgment 241–243.

[22]Judgment 247, citing Bartz v. Anthropic.

[23]Judgment 244–246.

[24]Judgment 256, 267, citing NITI Aayog, AI for Inclusive Societal Development (October 2025).

[25]Judgment 262, 258.4.

[26]Judgment 265, referring to ANI’s communication dated 3 October 2024.

[27]Judgment 266, 269, citing Zydus Lifesciences v. E.R. Squibb, FAO(OS)(COMM) 120/2025; F. Hoffmann-La Roche AG v. NATCO Pharma, CS(COMM) 567/2024; Astrazeneca AB v. Intas Pharmaceuticals Limited, 2020:DHC:3125.

[28]Judgment 270.

[29]Judgment 271.

[30]Judgment 273–274.