Summary
In the case of ANI Media Pvt. Ltd. v. OpenAI OPCO LLC, the Delhi High Court refused to restrain OpenAI from using ANI’s literary works for training the large language models underlying ChatGPT. At the interim stage, the Court held that it had jurisdiction despite OpenAI’s servers being located outside India, but found that ANI had not established copyright infringement through ChatGPT’s outputs. The Court also held that temporary storage of publicly available literary works for training an artificial intelligence model fell within fair dealing for private use and research under Section 52(1)(a) of the Copyright Act, 1957.
Background
Copyright Dispute Over AI Training and ChatGPT Outputs
ANI filed a copyright infringement suit against OpenAI and sought an interim injunction concerning the use of its news articles and interviews by ChatGPT.
ANI raised two principal claims. First, it alleged that OpenAI copied and stored its literary works while collecting data and training the large language models underlying ChatGPT. The Court referred to this as the training claim.
Second, ANI alleged that ChatGPT reproduced its copyright protected content while responding to user prompts. The Court referred to this as the reproduction or output claim.
OpenAI disputed both claims. It also questioned the jurisdiction of Indian courts because its models were trained and its data was stored on servers located outside India.
The Court therefore considered four interconnected questions concerning jurisdiction, model training, ChatGPT outputs and fair dealing.
Questions Before the Court
- Whether the Delhi High Court had jurisdiction when OpenAI’s training and storage servers were located outside India.
- Whether storing ANI’s literary works for training ChatGPT amounted to copyright infringement.
- Whether ChatGPT’s responses substantially reproduced ANI’s copyright protected expression.
- Whether OpenAI’s use of ANI’s works for model training qualified as fair dealing under Section 52(1)(a) of the Copyright Act.
- Whether ANI had established the requirements for an interim injunction.
Arguments Presented By the Parties
ANI’s Arguments
- OpenAI accessed and transmitted ANI’s works from India, offered its services to Indian users and generated the allegedly infringing outputs in India. Indian courts therefore had jurisdiction.
- Training an artificial intelligence model required copying, storing, tokenising and processing copyright protected works. These acts fell within the exclusive reproduction right under Section 14 of the Copyright Act.
- Public availability did not remove copyright protection or give OpenAI an implied licence to use ANI’s works.
- ChatGPT memorised ANI’s content and reproduced exact or substantially similar portions in response to prompts.
- OpenAI operated commercially and could not characterise its model training as private use or research under Section 52(1)(a).
- ChatGPT diverted readers from ANI’s content and damaged ANI’s licensing, subscription and advertising interests.
OpenAI’s Arguments
- The relevant training and data storage occurred outside India, and Indian copyright law could not apply beyond India.
- Large language models did not ordinarily retain complete copies of training works. They learned statistical relationships, language patterns and information from a large body of material.
- ANI’s examples concerned articles published after the training cut off dates of the relevant models. Those examples therefore could not prove memorisation through model training.
- ChatGPT’s responses conveyed news facts in different language and did not substantially reproduce ANI’s expression.
- Facts, events, grammar, syntax and ideas could not receive copyright protection.
- Temporary storage for model training amounted to private use and research and qualified as fair dealing under Section 52(1)(a).
- Commercial use was not expressly excluded by Section 52(1)(a).
Court’s Analysis of Jurisdiction
The Court first held that ANI had its principal and registered office in Delhi. Section 62(2) of the Copyright Act therefore gave the Delhi High Court jurisdiction over the infringement suit.
OpenAI also offered its services to users and paid subscribers in India, including users located within Delhi. The responses relied upon by ANI were generated and accessed within the Court’s jurisdiction. The Court therefore found jurisdiction under both Section 62 of the Copyright Act and Section 20 of the Code of Civil Procedure.
The location of OpenAI’s servers did not change this conclusion. According to the Court, storage on servers outside India was only the final step in a chain of events that began with accessing copyright protected works from India and transmitting them abroad.
The Court said that it could not sever the final act of overseas storage from the earlier acts forming part of the same transaction. Accepting OpenAI’s argument could permit a person to avoid Indian copyright law merely by locating the final server outside India.
The Court also treated the training and output claims as connected. As the output claim arose from the functioning of the trained model and the disputed outputs were produced in India, the Court declined to separate the claims for jurisdictional purposes. It consequently held that OpenAI had not established, at the prima facie stage, that the Court lacked territorial jurisdiction.
Court’s Analysis of ChatGPT Outputs
The Court accepted that original news articles and interviews could receive copyright protection even when they were freely available on ANI’s website. Public accessibility did not destroy copyright.
The protection, however, extended only to ANI’s original expression and not to the facts or events reported in its articles. News facts remained available for use by others, though another person could not substantially reproduce ANI’s particular form, arrangement and expression.
The Court applied the substantial similarity principles laid down in R.G. Anand v. Deluxe Films. It stated that the rival works had to be considered as a whole. A claimant could not select a few similar words or sentences from a longer work and compare them separately with the alleged copy.
Training Dates Defeated the Memorisation Claim
ANI relied on ChatGPT responses relating to articles published in August and September 2024. The training cut off dates of the models concerned were April 2022 for GPT 4 and April 2024 for GPT 4o.
The Court therefore found that the articles relied upon by ANI could not have formed part of the training data of those models. The responses could not prove that the models had memorised the articles during training.
The Court considered that the responses were more likely to have been generated through retrieval augmented generation, known as RAG. Through RAG, a system retrieves current information from external sources when responding to a prompt rather than relying only on its original training data.
As ANI had framed its case around training, memorisation and regurgitation, examples obtained through later retrieval could not establish that OpenAI had permanently stored and memorised ANI’s works during training.
No Substantial Reproduction
The Court compared ANI’s news reports with ChatGPT’s responses and found that the responses used different language, structure and commentary.
One example concerned a statement made by Neeraj Chopra’s mother after the Paris Olympics. ChatGPT first summarised her comments in its own words. ANI then asked ChatGPT to state “exactly” what she had said.
The Court treated the second instruction as an adversarial prompt intended to extract a particular response. Even then, ChatGPT reproduced only part of the speaker’s statement and added its own explanation.
The Court also observed that the speaker, rather than ANI, was prima facie the first owner of copyright in the statement under Section 17(cc) of the Copyright Act. ANI had not shown an assignment of that copyright in its favour.
After comparing the articles and responses as complete works, the Court held that ANI had not established substantial reproduction. It therefore rejected the output infringement claim at the interim stage.
Court’s Analysis of AI Training and Fair Dealing
The Court accepted that OpenAI electronically stored literary works, at least temporarily, while training its models. It therefore considered whether the storage came within Section 52(1)(a), which permits fair dealing with a work for private or personal use, including research.
Section 52 Creates User Rights
The Court rejected a narrow interpretation of Section 52. It stated that the provision balanced the exclusive rights of copyright owners with the public interest in learning, creativity and dissemination of knowledge.
Section 52 did not merely operate as a narrow exception or loophole. It independently defined activities that did not amount to infringement and had to receive a broad and liberal interpretation.
The Court adopted a two step test. First, it asked whether the use served one of the purposes mentioned in Section 52(1)(a). Second, it asked whether the dealing was fair.
Commercial Use Was Not Automatically Excluded
The Court noted that Section 52(1)(a) did not expressly limit fair dealing to non-commercial activity. Other clauses of Section 52 expressly referred to non-commercial institutions or activities. The absence of similar language in Section 52(1)(a) indicated that commerciality alone did not defeat the defence.
Research conducted by lawyers, institutions, companies or commercial organisations could still amount to research. The fact that OpenAI ultimately offered commercial services did not automatically place model training outside Section 52(1)(a).
Publicly Available Copies
The Court also rejected the contention that OpenAI had stored infringing copies.
ANI did not allege that OpenAI had bypassed a paywall, defeated a technological restriction or obtained ANI’s articles from an unlawful source. The works were freely available on ANI’s website.
The Court stated that the requirement concerning a non-infringing copy in the Explanation to Section 52(1)(a) applied specifically to the incidental storage of computer programs and did not prevent electronic storage of other works for the permitted purposes.
Model Training Was Private Use and Research
The Court stated that OpenAI stored training material within a closed system. Members of the public could not access or download the training dataset. Only the models used the material during the training process. The Court therefore treated the use as private.
It also treated model training as research. Training required screening, organising and analysing data, converting it into machine readable inputs and repeatedly adjusting the model to improve its statistical predictions.
According to the Court, research generally occurred before the final output. A researcher conducted research before writing a book, a court conducted research before preparing a judgment, and a scientist conducted research before producing an invention. The fact that the resulting product was later communicated to the public did not make the research process itself public.
The Court consequently held that training large language models amounted to research directed towards generating new knowledge and advancing artificial intelligence systems.
The Fairness Test
The Court then considered the nature and purpose of the use, the extent of copying, commercial competition, market injury and public interest.
ANI used its works to report and distribute news. OpenAI used a very large body of information to train a multipurpose language model capable of generating text, translating content, assisting research, writing software and providing information. The purposes were materially different.
The Court also found no evidence that ChatGPT acted as a market substitute for ANI’s services. ANI had not produced material showing that it had lost subscribers, advertising revenue or market share because of OpenAI.
The Court placed weight on the public benefits of trained language models. It referred to their use in education, scientific research, software development, translation, communication, access to information and tools for persons with disabilities.
Based on these considerations, the Court held that the purpose test and the fairness test were satisfied. OpenAI’s temporary storage of ANI’s literary works for training its models therefore fell within Section 52(1)(a) and did not amount to infringement at the interim stage.
Findings
The findings of the Court are as follows:
- The Delhi High Court had jurisdiction despite OpenAI’s servers being located outside India.
- Publicly accessible literary works continued to receive copyright protection, but copyright did not extend to the underlying news facts.
- ANI’s examples could not establish memorisation because the articles were published after the training cut off dates of the relevant models.
- The ChatGPT responses relied upon by ANI were not substantially similar to ANI’s articles when the works were compared as a whole.
- ANI failed to establish memorisation or regurgitation of its literary works through ChatGPT’s responses.
- Temporary storage of ANI’s publicly available works for model training amounted to private use and research under Section 52(1)(a).
- Commercial use did not, by itself, exclude the fair dealing defence.
- OpenAI’s use satisfied both the purpose test and the fairness test.
- ANI did not establish loss of subscribers, market share or revenue caused by OpenAI.
- ANI’s proposed licensing fee showed that any proved loss could be calculated and compensated through damages.
- An injunction could affect OpenAI’s services, the development of Indian artificial intelligence systems and members of the public using those systems.
- ANI failed to establish a prima facie case, balance of convenience or irreparable injury.
- The interim injunction application was dismissed.
- All findings were limited to the interim application and would not affect the final decision after trial.
Relevant Paras
Paragraph 55
“On a prima facie view, it cannot be said at this stage that this Court does not have the territorial jurisdiction to entertain the present suit or that the alleged infringement on account of training would involve extra-territorial application of the Indian Copyright Act.”
Paragraph 125
“Based on discussion above it cannot be concluded that the responses produced by ChatGPT are a substantial reproduction of the news articles of ANI. Hence, at a prima facie stage, ANI has failed to make out a case of copyright infringement based on the responses generated by ChatGPT i.e. the output claim.”
Paragraph 256
“In light of the discussion above, both the purpose test as well as the fairness test under Section 52(1)(a) stand fulfilled. Hence, in my prima facie view, Open AI’s acts of storage of the literary works of ANI for the training of its LLMs would fall under Section 52(1)(a) of the Copyright Act and hence, would not amount to infringement.”
Paragraphs 271 to 274
“Based on the discussion above, I am of the prima facie view that Open AI’s act of storing ANI’s original literary works for training LLMs underlying ChatGPT falls under Section 52(1)(a) of the Copyright Act and therefore, does not amount to infringement under Section 51 of the Copyright Act.”
“In view of the discussion above, ANI has failed to make out a prima facie case for grant of interim injunction. Balance of convenience is also against grant of interim injunction. Irreparable injury would be caused not only to Open AI but also to the public at large, if an interim injunction is granted in favour of ANI at this stage.”
“Accordingly, I.A. 45300/2024 is dismissed.”
“Needless to say, any observations made herein are only for the purpose of adjudication of the aforesaid application and would have no bearing on the final outcome of the suit.”
Case Citation
ANI Media Pvt. Ltd. v. OpenAI OPCO LLC, CS(COMM) 1028/2024, I.A. 45300/2024, High Court of Delhi, decided 24 July 2026, 2026:DHC:5900.
The order is available here.
Disclaimer
This case blog is based on the author’s understanding of the judgment. Understandings and opinions of others may differ. An AI application was used to generate parts of this case blog based on user inputs and prompts.