A) ABSTRACT / HEADNOTE

ANI Media Pvt. Ltd. v. Open AI Opco LLC, CS(COMM) 1028/2024, I.A. 45300/2024, decided by the Delhi High Court on 24 July 2026, is a landmark Indian decision on copyright, artificial intelligence, large language models, training data, fair dealing, and territorial jurisdiction. The Court considered whether Open AI’s alleged storage of ANI’s copyrighted news content for training ChatGPT amounted to infringement under the Copyright Act, 1957. It also examined whether ChatGPT outputs reproduced ANI’s works in a legally substantial manner. Justice Amit Bansal refused interim injunction. The Court held, at a prima facie stage, that storage of ANI’s literary works for LLM training fell within Section 52(1)(a) as private or personal use, including research. It further held that ANI failed to prove memorisation, regurgitation, or substantial reproduction in ChatGPT outputs. The decision is important because it interprets old copyright provisions for modern AI systems. It balances authorial rights with innovation, public interest, education, research, and access to information.

Keywords: Copyright Infringement, Artificial Intelligence, Large Language Models, Fair Dealing, Section 52, Training Data, ChatGPT, RAG, Territorial Jurisdiction, Interim Injunction

B) CASE DETAILS

i) Judgement Cause Title: ANI Media Pvt. Ltd. v. Open AI Opco LLC.

ii) Case Number: CS(COMM) 1028/2024, I.A. 45300/2024, I.A. 45301/2024 and I.A. 26192/2025.

iii) Judgement Date: Judgment reserved on 27 March 2026. Judgment pronounced on 24 July 2026.

iv) Court: High Court of Delhi at New Delhi.

v) Quorum: Hon’ble Mr. Justice Amit Bansal.

vi) Author: Justice Amit Bansal.

vii) Citation: ANI Media Pvt. Ltd. v. Open AI Opco LLC, CS(COMM) 1028/2024, I.A. 45300/2024, Delhi High Court, decided on 24 July 2026. Indian Kanoon identifies the author and bench as Amit Bansal, J.

viii) Legal Provisions Involved: Section 13, Section 14, Section 17, Section 51, Section 52(1)(a), Section 62 of the Copyright Act, 1957, and Section 20 of the Code of Civil Procedure, 1908. The Court mainly interpreted Section 14(a)(i), Section 14(a)(iii), Section 51, and Section 52(1)(a).

ix) Judgments Overruled by the Case: No judgment was overruled. The Court distinguished several decisions. It applied existing copyright principles to AI training.

x) Related Law Subjects: Copyright Law, Intellectual Property Law, Artificial Intelligence Law, Technology Law, Civil Procedure, Private International Law, Commercial Litigation, Internet Jurisdiction, and Constitutional Public Interest Principles.

xi) Counsels Appearing for Parties: ANI was represented by Mr. Sidhant Kumar, Mr. Akshit Mago, Ms. Manyaa Chandok, Ms. Anshika Saxena, and Ms. Lahar Jain. Open AI was represented by Mr. Amit Sibal, Senior Advocate, with Mr. Sanjeev Kapoor, Mr. Nirupam Lodha, Mr. Madhav Khosla, Mr. Gautam Wadhwa, Ms. Moha Paranjpe, Mr. Abhi Udai Singh Gautam, Ms. Rebecca Cardoso, Mr. Hardik Malik, Ms. Vanshika Thapliyal, and Mr. Rajat Bector. Several intervenors appeared. Mr. Adarsh Ramanujan and Professor Arul George Scaria assisted as Amici Curiae.

C) INTRODUCTION AND BACKGROUND OF JUDGEMENT

The dispute arose because ANI alleged unauthorised use of its copyrighted news content by Open AI. ANI claimed that Open AI copied and stored its works for training large language models underlying ChatGPT. ANI also argued that ChatGPT generated outputs reproducing ANI’s works. The Court treated the case as a difficult meeting point between traditional copyright law and modern artificial intelligence. The judgment records that copyright law has faced similar disruptions before. Printing, broadcasting, photocopying, video recording, streaming, and now AI have each disturbed settled legal arrangements. The Court therefore approached the matter with caution. It did not treat AI as legally exempt. It also did not treat copyright as mechanically prohibiting every technical use of public data. The central task was to read the Copyright Act, 1957 in a technologically updated manner. That approach is visible throughout the judgment. The Court accepted that LLMs use enormous text corpora. It also noted that data may come from publicly available web sources and licensed sources. It then asked whether such use is infringement, fair dealing, or a matter requiring trial.

The Court first explained how LLMs work. It stated that an LLM predicts the next word from preceding words. The model is trained on vast text corpora. The data is processed through tokenisation and embedding. Tokens become numerical representations. During training, the system repeatedly predicts masked words and adjusts its parameters. After training, the model is fine-tuned to follow instructions. A user then interacts with the model through prompts. The Court also discussed Retrieval-Augmented Generation, or RAG. RAG does not rely only on training data. It retrieves current or external information before generating responses. This technical explanation mattered legally. ANI’s claim depended on whether training involved copying. ANI also alleged that outputs showed memorisation. Open AI responded that outputs were not copies but generated responses. The Court’s technical discussion allowed it to separate training claim, output claim, and RAG-based responses.

D) FACTS OF THE CASE

ANI is a news agency. It publishes and licenses news reports, articles, interviews, images, videos, and related content. ANI claimed ownership over original literary works published on its platforms. It relied on agreements with professionals to show that copyright vested in ANI. Open AI operates ChatGPT. ChatGPT is a generative AI application based on large language models. ANI alleged that Open AI copied and stored ANI’s works for training ChatGPT. ANI also claimed that ChatGPT outputs reproduced ANI’s content. ANI’s interim application sought restraint against Open AI. The relief effectively required Open AI to stop storing or using ANI’s works and to prevent infringing outputs. Open AI resisted the injunction. It argued that its training happened outside India. It also stated that the specific fourteen ANI works relied upon were published after the relevant training cut-off dates for GPT-4 and GPT-4o. Therefore, those works could not have been memorised during training.

The proceedings became larger than an ordinary copyright dispute. Several industry bodies intervened. The Federation of Indian Publishers, Digital News Publishers Association, and Indian Music Industry broadly supported ANI. Flux AI Labs, IGAP Project LLP, and Broadband India Forum broadly supported Open AI. The Court appointed two amici. This shows the systemic importance of the case. The matter did not only concern ANI and Open AI. It affected publishers, news businesses, music producers, AI developers, start-ups, users, researchers, and the public. The Court had to decide an interim injunction. It repeatedly clarified that its findings were prima facie. Therefore, the judgment does not finally decide every question of AI copyright liability. It decides whether ANI had shown enough to justify immediate interim restraint.

E) LEGAL ISSUES RAISED

The first issue was whether storage of ANI’s data for training ChatGPT amounted to copyright infringement. This involved Section 14(a)(i), which includes reproduction in material form and electronic storage. It also involved Section 51, which explains when copyright is infringed. The second issue was whether ChatGPT outputs infringed ANI’s copyright by reproducing ANI’s works. This required the Court to test memorisation, regurgitation, substantial similarity, and communication to the public. The third issue was whether Open AI could rely on fair dealing under Section 52(1)(a). Though the framed issue used “fair use”, the Court clarified that Indian law uses fair dealing. The fourth issue was whether Indian courts had jurisdiction when Open AI’s servers were located in the United States.

The jurisdiction issue was foundational. If Delhi courts lacked jurisdiction, the remaining questions could not be considered. The output issue was examined before the training and fair dealing issues because it influenced the Court’s understanding of actual harm and reproduction. The Court later considered the training and fair dealing issues together. That method was logical. If storage was reproduction, the next question was whether such storage was protected by Section 52(1)(a). The Court also had to consider whether interim injunction was justified. This required the familiar triad of prima facie case, balance of convenience, and irreparable injury. Public interest was treated as an important factor.

F) PETITIONER / PLAINTIFF’S ARGUMENTS

ANI argued that Open AI used its copyrighted works without permission. It said that training an LLM necessarily involves copying and storing copyrighted text. According to ANI, tokenisation and vectorisation did not remove copyright concerns. They were merely technical forms of converting protected expression into machine-readable form. ANI argued that such conversion still involved reproduction or adaptation. ANI also submitted that Open AI’s model could detokenise or reconstruct language into readable outputs. Therefore, ANI claimed that the entire process showed use of its protected expression. ANI further argued that public availability is not equal to public ownership. A work available on the internet can still be copyrighted. This argument is doctrinally strong. Copyright protects expression even when the work is freely accessible. The Court accepted this broad proposition. It treated ANI’s publicly accessible literary works as capable of copyright protection at the interim stage.

ANI also argued that ChatGPT outputs reproduced its works. It relied on examples from the plaint. It said that some outputs were close to ANI articles and interviews. ANI placed emphasis on memorisation and regurgitation. It argued that LLMs can retain training data and reproduce it when prompted. ANI relied on foreign decisions such as The Associated Press v. Meltwater U.S. Holdings, Inc., 931 F. Supp. 2d 537 (S.D.N.Y. 2013), GEMA v. Open AI, Case No. 42 O 14139/24, Munich Regional Court, 11 November 2025, and Advance Local Media LLC v. Cohere Inc., 1:25-cv-01305 (S.D.N.Y.). ANI submitted that commercial AI companies cannot take protected news content without licence. It also stressed that Open AI had entered licensing arrangements with other news organisations. Therefore, according to ANI, Open AI understood the value of news content but avoided paying ANI.

G) RESPONDENT / DEFENDANT’S ARGUMENTS

Open AI argued that ANI failed to show infringement. It stated that the fourteen works relied upon by ANI were published after the relevant training cut-off dates. Therefore, Open AI said those works could not have been included in training data for the models then used by ANI. Open AI also argued that copyright does not protect facts, ideas, events, or news as such. It protects only expression. This principle comes from classic copyright doctrine. It was reinforced through R.G. Anand v. Deluxe Films, (1978) 4 SCC 118 and Eastern Book Company v. D.B. Modak, (2008) 1 SCC 1. In R.G. Anand, the Supreme Court explained that copyright does not protect ideas. It protects expression, form, arrangement, and treatment. In EBC v. Modak, the Supreme Court rejected mere “sweat of the brow” and required skill and judgment. These cases were important because ANI’s works were news-based. News reporting has protectable expression. Yet the underlying facts remain free.

Open AI also argued that ChatGPT was not designed to reproduce training material. It said outputs differ between users and prompts. It submitted that any regurgitation was rare and actively mitigated. Open AI emphasised that ChatGPT generates responses from learned statistical patterns. It also argued that RAG-based search outputs are more like summaries or snippets with source references. On fair dealing, Open AI argued that training was private use and research under Section 52(1)(a). It submitted that commercial character does not automatically defeat fair dealing. It also contended that requiring licences for every public web source would seriously harm AI development. On interim relief, Open AI argued that ANI’s claim was compensable in money. ANI had itself offered a licence for USD 7.5 million. Open AI also said it had blocked ANI’s website for training and ChatGPT search/RAG purposes.

H) RELATED LEGAL PROVISIONS

Section 13 of the Copyright Act, 1957 recognises copyright in original literary, dramatic, musical, and artistic works. ANI’s news articles and interviews were treated as original literary works at the interim stage. Section 17 concerns first ownership of copyright. ANI relied on professional services agreements to show ownership in works created by its personnel. Section 14(a)(i) gives the owner the exclusive right to reproduce a literary work in material form, including storing it electronically. Section 14(a)(iii) protects communication to the public. Section 51 defines infringement when a person, without licence, does something exclusively reserved to the copyright owner. These provisions formed the base of ANI’s claim.

Section 52(1)(a) was the heart of Open AI’s defence. It protects fair dealing with works, other than computer programmes, for private or personal use, including research, criticism or review, and reporting current events. The Explanation protects electronic storage for those purposes. The Court held that this provision should not be read narrowly. It treated Section 52 not merely as an exception, but as part of the internal balance of copyright law. Copyright rewards authors. It also serves learning, research, education, innovation, and public access. The Court therefore used a purpose test and fairness test. The purpose test asked whether Open AI’s storage came within private or personal use, including research. The fairness test asked whether Open AI’s use prejudiced ANI’s market and whether ChatGPT served public interest.

I) PRECEDENTS ANALYSED BY COURT IN THIS CASE

The Court relied on R.G. Anand v. Deluxe Films, (1978) 4 SCC 118 for the core test of copyright infringement. The case holds that copyright protects expression, not ideas. Similarity must be substantial, material, and based on the work as a whole. This helped Open AI because the Court compared ANI articles and ChatGPT outputs as wholes. It found no substantial similarity at the interim stage.

The Court referred to Eastern Book Company v. D.B. Modak, (2008) 1 SCC 1 for originality. The Supreme Court rejected pure labour-based copyright. It required skill and judgment. This mattered because news content often contains facts. The Court accepted ANI’s copyright in original literary expression. Yet it recognised that facts and events themselves are not monopolised.

The Court used Banyan Tree Holding (P) Ltd. v. A. Murali Krishna Reddy, 2009 SCC OnLine Del 3780 on internet jurisdiction. It supported the view that purposeful availment and targeting users in a forum may create jurisdiction. Open AI offered services to users in India, including Delhi. This helped ANI on jurisdiction.

The Court relied on Neetu Singh v. Telegram FZ LLC, 2022 SCC OnLine Del 2637. That case held that foreign server location cannot defeat Indian copyright remedies where infringement affects India. The Court applied similar reasoning. It held that Open AI could not avoid Indian jurisdiction merely by locating servers abroad.

The Court also discussed Syndicate of the Press of the University of Cambridge v. B.D. Bhandari, 2011 SCC OnLine Del 3215. That case accepted transformative or derivative use in the context of guidebooks. It helped the Court develop the fairness inquiry. Rameshwari Photocopy Services was considered for educational copying and Section 52. Blackwood v. A.N. Parasuraman, AIR 1959 Mad 410 was used for understanding “research”. State through CBI v. S.J. Choudhary, (1996) 2 SCC 428 supported updating construction. Foreign cases such as Authors Guild v. Google Inc., 804 F.3d 202 (2d Cir. 2015), Bartz v. Anthropic PBC, and Kadrey v. Meta Platforms were used cautiously. The Court treated foreign law as persuasive only where compatible with Indian law.

H) JUDGEMENT

a) RATIO DECIDENDI

The ratio of the judgment is that, at the interim stage, Open AI’s storage of ANI’s original literary works for training LLMs underlying ChatGPT falls within Section 52(1)(a). Therefore, it does not amount to infringement under Section 51. The Court held that research under Section 52(1)(a) must receive an updated interpretation. Research is not limited to human study with books and notes. Machine learning can also be research when it is performed for developing knowledge systems for human benefit. The Court also held that private use can include use by a private company in a closed technical environment. The training data was not made publicly available in natural language or tokenised form. This made the use private in the Court’s prima facie view.

The Court further held that ANI failed to show that ChatGPT outputs were substantially similar to ANI works. The examples relied upon were published after the model training dates. Therefore, they could not establish memorisation from training. The Court found that the outputs were summaries or differently worded responses. It applied R.G. Anand and examined the whole works. The Court also held that RAG outputs did not amount to infringement because they were not substantially similar. ANI failed to show memorisation or regurgitation. Hence, the output claim failed at the interim stage. Finally, the Court held that Delhi High Court had jurisdiction. ANI had its principal office in Delhi. Open AI targeted Indian users. The alleged outputs were generated in India. Foreign server location was not decisive.

b) OBITER DICTA

The judgment contains important observations beyond the narrow refusal of interim injunction. The Court observed that copyright law must adapt to technological change. It recognised AI as a major public-interest technology. It stated that traditional copyright provisions must be interpreted with sensitivity to new technological realities. The Court also observed that Section 52 should receive broad and liberal interpretation because it represents user rights and public interest, not a grudging exception. This observation may influence future copyright cases.

The Court also made significant comments on public interest. It noted that AI tools support education, research, translation, software development, accessibility, and knowledge dissemination. It warned that requiring licences from multiple web sources at the interim stage could make LLM development economically unviable. This was not a final policy ruling. Yet it shows judicial concern for innovation. The Court also recognised that copyright owners remain important. Its reasoning therefore rests on balance, not absolute freedom for AI companies.

c) GUIDELINES / PRINCIPLES EMERGING FROM THE CASE

First, foreign server location does not automatically defeat Indian copyright jurisdiction. If the plaintiff is based in India, the defendant targets Indian users, and outputs are generated in India, Indian courts may exercise jurisdiction.

Second, publicly available content may still be copyrighted. The Court did not hold that internet availability destroys copyright. ANI’s works remained protectable at the interim stage.

Third, facts and news events are not protected as such. Only original expression is protected. This is especially relevant for news agencies.

Fourth, LLM training may involve electronic storage. The Court accepted that at least temporary storage occurs during training.

Fifth, Section 52(1)(a) may protect AI training where the use is private, research-oriented, fair, non-substitutive, and publicly beneficial.

Sixth, commercial use does not automatically defeat fair dealing. The absence of “non-commercial” language in Section 52(1)(a) mattered.

Seventh, memorisation and regurgitation require evidence. Mere possibility is insufficient for interim injunction.

Eighth, substantial similarity must be assessed as a whole. Isolated similarities are not enough.

Ninth, public interest matters in AI injunctions. Courts must consider effects on users, research, innovation, and access to information.

Tenth, trial remains open. The findings are prima facie and do not finally decide liability.

I) CONCLUSION & COMMENTS

The decision is a major starting point for Indian AI copyright law. It does not give AI companies blanket immunity. It does not abolish copyright in news. It does not say that every scraping activity is lawful. Its holding is narrower. On the facts placed at the interim stage, ANI could not prove substantial output infringement. It could not prove memorisation or regurgitation. Its examples were weak because they post-dated the training cut-off dates. The Court also found that Open AI’s internal storage for training was protected by Section 52(1)(a). This conclusion rests on a broad reading of private use and research.

The judgment is persuasive because it treats AI seriously. It avoids emotional conclusions. It studies how LLMs work. It separates facts from expression. It recognises author rights but also protects innovation. Its strongest contribution is the two-step Section 52 test. The Court first asks whether the purpose fits Section 52. It then asks whether the dealing is fair. For fairness, it considers limited training use, market harm, and public interest. This structure gives Indian courts a workable method for future AI disputes.

The judgment also has limits. It is an interim order. Trial evidence may change conclusions. A stronger case involving verbatim outputs, paywalled content, repeated copying, or clear market substitution may be decided differently. The Court’s reasoning may also face appellate scrutiny. Still, the ruling is likely to shape Indian debates on text and data mining. It also places India near global conversations on AI training, fair dealing, and copyright balance.

J) REFERENCES

a) Important Cases Referred

1. ANI Media Pvt. Ltd. v. Open AI Opco LLC, CS(COMM) 1028/2024, I.A. 45300/2024, Delhi High Court, decided 24 July 2026.
2. R.G. Anand v. Deluxe Films, (1978) 4 SCC 118.
3. Eastern Book Company v. D.B. Modak, (2008) 1 SCC 1.
4. Banyan Tree Holding (P) Ltd. v. A. Murali Krishna Reddy, 2009 SCC OnLine Del 3780.
5. Neetu Singh v. Telegram FZ LLC, 2022 SCC OnLine Del 2637.
6. Syndicate of the Press of the University of Cambridge v. B.D. Bhandari, 2011 SCC OnLine Del 3215.
7. University of Oxford v. Rameshwari Photocopy Services, 2016 SCC OnLine Del 6229.
8. Blackwood & Sons Ltd. v. A.N. Parasuraman, AIR 1959 Mad 410.
9. State through CBI v. S.J. Choudhary, (1996) 2 SCC 428.
10. Authors Guild v. Google Inc., 804 F.3d 202 (2d Cir. 2015).
11. Associated Press v. Meltwater U.S. Holdings, Inc., 931 F. Supp. 2d 537 (S.D.N.Y. 2013).
12. GEMA v. Open AI, Case No. 42 O 14139/24, Munich Regional Court, decided 11 November 2025.
13. Advance Local Media LLC v. Cohere Inc., 1:25-cv-01305 (S.D.N.Y.).
14. Bartz v. Anthropic PBC, 3:24-cv-05417 (N.D. Cal.).
15. Kadrey v. Meta Platforms, Inc., N.D. Cal..

b) Important Statutes Referred

1. The Copyright Act, 1957, Section 13.
2. The Copyright Act, 1957, Section 14(a)(i).
3. The Copyright Act, 1957, Section 14(a)(iii).
4. The Copyright Act, 1957, Section 17.
5. The Copyright Act, 1957, Section 51.
6. The Copyright Act, 1957, Section 52(1)(a).
7. The Copyright Act, 1957, Section 62.
8. The Code of Civil Procedure, 1908, Section 20.
9. Berne Convention for the Protection of Literary and Artistic Works, 1886, Article 9.
10. TRIPS Agreement, 1994, copyright enforcement principles and international IP balance.