Summary: The Delhi High Court has refused ANI Media’s interim injunction application against OpenAI holding on a prima facie basis that the scraping and storage of copyrighted content to train LLMs constitutes fair dealing under Section 52(1)(a) of the Copyright Act, 1957. The ruling, India’s first major judicial decision on AI training and copyright, signals a broad pro-AI path, though substantive questions remain open for trial.
OVERVIEW
In a landmark decision, the Delhi High Court (Justice Amit Bansal) has dismissed ANI Media’s application for an interim injunction against OpenAI in India’s first major judicial ruling on whether the scraping and storage of copyrighted content to train large language models (“LLMs”) constitutes copyright infringement under the Copyright Act, 1957. The Court’s prima facie findings chart a broad pro-AI path on the training question, while leaving substantive issues open for trial.
BACKGROUND
ANI Media Pvt. Ltd. (“ANI”), one of India’s largest news agencies, filed a copyright infringement suit against Open AI OpCo LLC (“OpenAI”), alleging that OpenAI scraped, stored, and used ANI’s copyrighted news articles without authorisation to train the LLMs underlying ChatGPT, and that ChatGPT’s outputs reproduce ANI’s works. ANI sought an interim injunction restraining OpenAI from continuing to store or use its works.
ISSUES FRAMED BY THE COURT
The Court framed four issues:
- Issue 1 (Training/ Storage): Whether storage of ANI’s data to train ChatGPT amounts to copyright infringement.
- Issue 2 (Output/ Reproduction): Whether ChatGPT’s responses reproducing ANI’s data amount to copyright infringement.
- Issue 3 (Fair Dealing): Whether OpenAI’s use qualifies as “fair dealing” under Section 52 of the Copyright Act, 1957.
- Issue 4 (Jurisdiction): Whether Indian courts have jurisdiction, given that OpenAI’s servers are in the United States.
KEY HOLDINGS
1. Jurisdiction – Indian Courts Have Jurisdiction (Issue 4)
The Court held, on a prima facie basis, that Indian courts have territorial jurisdiction under both Section 62(2) of the Copyright Act and Section 20 of the Code of Civil Procedure, 1908. ANI’s registered and principal office is in New Delhi, and OpenAI actively targets Indian subscribers and collects subscription fees from them.
Critically, on the training claim, OpenAI argued that because training takes place on US servers, the Indian Copyright Act has no extra-territorial reach. The Court rejected this, holding that the storage of ANI’s works on US servers is merely the “terminal step” in a chain of events that begins with access to and transmission of copyrighted works from India. Severing that chain to look only at the final link would allow infringers to evade Indian copyright law by routing operations offshore.
Implication: A foreign AI company that accesses and copies Indian-published copyrighted content for LLM training cannot escape Indian copyright jurisdiction simply by locating its servers abroad.
2. Output/ Reproduction Claim – No Prima Facie Infringement (Issue 2)
ANI presented various articles and several screenshots of ChatGPT prompts and responses as evidence of reproduction. The Court rejected the output claim at the interim stage on at least the following distinct grounds:
- Training cut-off: All ANI articles relied upon were published in August or September 2024, after the training cut-off dates of GPT-4 (April 2022) and GPT-4o (April 2024). Accordingly, those specific articles could not have been part of the training data. Therefore, the claim of memorisation and regurgitation of those works could not be sustained.
- No substantial similarity: The Court found, on its own analysis of the examples provided by ANI, that ChatGPT’s responses were not substantially similar to ANI’s articles when compared in its entirety.
- The Court also confirmed that public availability of ANI’s news content does not nullify its copyright. The fact/ expression dichotomy means copyright in news articles protects the specific expression, not the underlying facts. The threshold for establishing substantial similarity in the expression of factual news content is accordingly higher than for creative works such as song lyrics.
Implication: LLM developers face lower immediate litigation risk on output claims where (a) the challenged content post-dates the training cut-off, (b) outputs differ substantively in expression from source articles, and (c) reproduction requires adversarial prompting. However, the Court left open the possibility that evidence of memorization and verbatim reproduction could be established at trial.
3. Training/ Storage Claim and Fair Dealing – Storage Is Prima Facie Fair Dealing (Issues 1 and 3)
This is the most significant part of the judgement. The Court took Issues 1 and 3 together because Section 52 of the Copyright Act is not a proviso or exception to infringement but an integral part of the statute that must be read alongside Section 14 and Section 51. If an act falls within Section 52, it is not an infringing act at all.
Section 52(1)(a) of the Copyright Act provides that “fair dealing” with any work (other than a computer programme) for the purposes of (i) private or personal use, including research; (ii) criticism or review; or (iii) reporting of current events shall not constitute infringement. The Court applied a two-step test to determine whether OpenAI’s storage of ANI’s works satisfies Section 52(1)(a):
STEP 1 – Purpose Test
- Commercial use does not bar the defence. The Court held that the absence of an explicit non-commercial limitation in Section 52(1)(a)(i) is deliberate. The legislature expressly added “non-commercial” in other sub-sections (e.g. Sections 52(1)(ad), (k)(ii), (l), (n), (o)) but not in Section 52(1)(a)(i). Accordingly, a commercial entity may still invoke the defence.
- No requirement for the stored copy to be non-infringing. The Court rejected ANI’s argument that the Explanation to Section 52(1)(a), which deals with incidental storage of computer programmes, imposes a general condition that any work stored must not itself be an infringing copy. On the Court’s reading, the “not itself be an infringing copy” limitation applies only to the incidental storage of computer programmes, not to literary works stored in electronic form. In any event, since ANI’s works are freely accessible on its website (without a paywall) and OpenAI scrapes publicly available content, the copies used are not infringing copies.
- Training is “private or personal use, including research”. Applying the doctrine of updating construction (following the Supreme Court in State (Through CBI) v. S.J. Choudhary), the Court held that “research” and “private use” under Section 52(1)(a)(i) must be read in light of modern technological conditions and must not be confined to human actors. The training of LLMs, which involves machine learning of stored data, extraction of patterns, iterative error correction, and generation of new knowledge, is a form of research, and because the training dataset is never made available to the public (it is accessible only to the LLM system), the use is private. The Court held that research conducted commercially can still qualify as “research”.
STEP 2 – Fairness Test
The Court drew on principles from Indian and international decisions to formulate three fairness factors applicable to the facts:
- Factor 1 – Use is limited to training: OpenAI has not been shown to use ANI’s works for any purpose other than training its LLMs. The training data is never communicated to the public in natural language or tokenised form.
- Factor 2 – No market substitution: ChatGPT’s functions (content creation, research assistance, translation, summarisation) are fundamentally different from ANI’s business of news syndication. The Court found no evidence of ANI losing subscribers, advertising revenue, or other demonstrable market harm from OpenAI’s activities. ANI’s own offer to license its content to OpenAI for USD 7.5 million demonstrated that any loss is quantifiable and can be compensated financially.
- Factor 3 – Public interest in AI development: The Court took cognisance of the substantial societal benefits of trained LLMs in healthcare, education, financial services, agriculture, and accessibility, as well as the NITI Aayog’s position on AI-inclusive development. Requiring AI developers to obtain individual licences from every copyright holder would be economically unviable and would stifle the development of Indian LLMs, including those in early-stage development by domestic start-ups.
Prima facie, the Court held that both the purpose test and the fairness test under Section 52(1)(a) are satisfied, and that OpenAI’s storage of ANI’s works for training purposes does not amount to copyright infringement.
The Court also noted that even the balance of convenience and irreparable harm weighed strongly against granting an injunction. The Court identified several factors in this respect:
- ANI’s claim is quantifiable in monetary terms (ANI itself offered a licence for USD 7.5 million).
- OpenAI had already voluntarily blocked ANI’s website from its web crawlers for both training and RAG.
- An injunction requiring OpenAI to delete training data would potentially contradict US legal obligations to preserve it.
- A blanket injunction would harm millions of ChatGPT users in India and impede the development of domestic LLMs, harming broader public interest.
- ANI has the technical ability to block web crawlers from its website and has not done so.
The present order is an interim order and many questions have been left open for trial. Further, the judgement uses the doctrine of updating construction to interpret existing provisions in light of current AI realities, but notes Parliament must address legislative gaps. This may accelerate legislative debate on whether India should introduce a dedicated AI training exception, and under what terms.
