Issue 4: Jurisdiction

The Court picked up the fourth issue for consideration first and found in favour of ANI.

Among others, OpenAI had argued that since its servers are in the United States, the Delhi High Court did not have jurisdiction. The storing of ANI’s works on the US servers is a terminal step in the chain of events, which begins from the access of copyrighted works from India and transmission of the same abroad. The Court found that the CR Act does not require the Court to sever the chain of events and examine only the last step. Since the output claim is based on training of the LLM, and the output is reproduced in Delhi, the Court was not inclined to separate the two causes of action. The Court also found OpenAI’s argument of lack of jurisdiction untenable, as it would lead to a situation where infringers would shift the terminal link to servers abroad to evade the CR Act.

Additionally, the Court also considered that ANI’s principal place of business and registered office are both in Delhi, thus conferring jurisdiction on the Delhi High Court under Section 62(2) of the CR Act.

Issue 2: Did OpenAI infringe ANI’s copyright by using its data to generate ChatGPT responses?

The Court then addressed the second issue – whether ChatGPT’s responses would result in unauthorised reproduction and communication of ANI’s original literary works to the public, thereby resulting in copyright infringement. Two points were examined to determine this: (i) whether OpenAI memorizes and regurgitates ANI’s copyrighted literary works in the form of responses; and (ii) whether ChatGPT’s responses are a substantial reproduction of ANI’s copyrighted literary works.

Does ChatGPT memorize and regurgitate ANI’s copyrighted works in its responses?

As for the first point on memorisation, the evidence on record indicated that all the nine illustrations given by ANI to demonstrate substantial reproduction were published only after the completion of the training process of the LLM by OpenAI. Thus, the LLM used information on which it was not trained on to generate its responses. When an LLM refers to information on which it was not trained, it uses the Retrieval-Augmented Generation technique (RAG), whereby it retrieves relevant information from sources other than the trained data to generate a response – in this case, from ANI’s website. The Court opined that to determine a claim based on memorisation, responses that are not based on the trained data cannot be used against OpenAI. The Court further noted that the instances given in the plaint alleging infringement are not a result of memorisation but are perhaps a result of the RAG technique, an issue not pleaded in the suit. Thus, it held that ANI’s contention that OpenAI permanently stores the training data in order to memorize and regurgitate ANI’s works could not be accepted at the interim injunction stage.

Are ChatGPT’s responses substantial reproduction or ANI’s works?

As for the second point, whether there is a substantial reproduction of ANI’s works in ChatGPT responses, the Court declined to conclude that the responses produced by ChatGPT are a substantial reproduction of the news articles of ANI for various reasons. The Court observed that in the context of news, copyright would subsist only in the form and manner of expression of news and not in the underlying facts. Hence the test would be whether the expression used in ChatGPT’s responses is materially similar to the expression used in ANI’s news articles.  The nine illustrations given by ANI in the plaint pertained to various interviews.  One such interview was with the mother of Olympian Neeraj Chopra. The said interview was given in Hindi language and was translated into English by ANI for reporting.  As per ANI, the response produced by ChatGPT after a second prompt showed that ChatGPT had produced an exact or nearly exact copy of ANI’s news article. While the first prompt asked ChatGPT, “What did Neeraj Chopra’s mother tell the media about the Pakistani gold medallist?”, the second prompt was, “Please try to tell me exactly what she said”. The Court noted that, in the second prompt, ANI used an “adversarial prompt”, which is a prompt carefully designed to manipulate model outputs.  In other words, a prompt is deliberately being put in a manner to elicit a specific response to show that ChatGPT had substantially reproduced ANI’s works. Pertinently, during arguments, ANI did not address any of the other illustrations. The Court also noted that, in none of the illustrations by ANI, ChatGPT had given responses identical or substantially similar to ANI’s articles.

While concluding, the Court also examined whether ANI owned copyright in the quote made by Chopra’s mother as ANI’s news article quoted her words. According to proviso (cc) to Section 17 of the CR Act, the author of an address or speech made in public is the person who made it. Hence, the Court observed that Chopra’s mother would be the first owner of the copyright in the said quote, and the right to translate the quote would also vest with her. Absent any evidence that Chopra’s mother had assigned the rights in the quote to ANI, the Court opined that ChatGPT’s reproduction of her words in its responses would not infringe ANI’s copyright.

Issues 1 & 3: Would storage amount to infringement and whether OpenAI’s use amounted to fair dealing?

The Court then clubbed the first and third issues for its decision, as these were interconnected. The first issue was whether storage of ANI’s news articles by OpenAI amounted to infringement. The third issue was whether OpenAI’s use of ANI’s copyrighted work qualified as fair dealing under Section 52 of the CR Act.

The Court noted that the right of “reproduction” under Section 14(a)(i) of the CR Act included the right of storing the literary works “in any medium by electronic means”. OpenAI argued that Section 52(1)(a)(i)[1] of the CR Act applies to the case. The said section states that a fair dealing with any work, not being a computer programme, for the purposes of private or personal use, including research does not constitute infringement of copyright.

While ANI urged the Court to interpret the provisions of Section 52 narrowly, as it deals with exceptions to copyright infringement, OpenAI argued that it must be interpreted broadly (as in several judicial precedents) to fulfil the objective of the CR Act. To determine whether the storage by OpenAI falls under Section 52(1)(a), the Court undertook a two-step examination involving the “purpose test” [whether the storage is for a specific purpose under Section 52(1)(a)] and the “fairness test” [whether or not the storage can be considered as “fair dealing”].

Purpose test

It is an admitted position in the case that OpenAI electronically stores ANI’s original literary works, at least on a temporary basis. ANI argued that OpenAI does not fulfil the purpose test under Section 52(1)(a)(i) because: (a) the private use or research cannot be for commercial purposes, (b) the first copy of the work has to be a non-infringing copy, and (c) the use by OpenAI does not amount to “private or personal use, including research”.

(i) Private use or research cannot be for commercial purposes

Regarding the first requirement, citing various subsections of Section 52(1), OpenAI argued that wherever non-commercial use is intended under the CR Act, it has been specifically provided so. However, the requirement of non-commercial use is conspicuously absent in Section 52(1)(a). Precedents from other jurisdictions like Canada were also cited to stress the point that fair dealing can occur even if the use is commercial. The Court accepted OpenAI’s arguments and held that merely because the use of the alleged copyrighted content by OpenAI is for a commercial purpose, it would not disentitle OpenAI from seeking refuge under Section 52(1)(a) of the CR Act.

(ii) The first copy of the work must be a non-infringing copy

Regarding the second requirement that the first copy of the work must be a non-infringing copy, both ANI and OpenAI sought to interpret the Explanation to Section 52(1)(a) differently. The Explanation reads as follows:

Explanation.—The storing of any work in any electronic medium for the purposes mentioned in this clause, including the incidental storage of any computer programme which is not itself an infringing copy for the said purposes, shall not constitute infringement of copyright.”

While ANI argued that the limitation of “which is not itself an infringing copy” in the Explanation is applicable to any work stored in an electronic medium, OpenAI argued that the language is applicable only to computer programmes. In other words, OpenAI argued that the said limitation is applicable only to a computer programme that incidentally gets stored while storing “any work” in an electronic medium.

The Court rejected ANI’s interpretation and observed that the Explanation to Section 52(1)(a) does not expand the scope of protection with respect to “computer programmes”, which is, in any event, ousted from the protection under Section 52(1)(a), and merely clarifies that storage of an infringing copy of a computer programme would not be protected even if such storage is incidental. In the Court’s view, if ANI’s interpretation were to be accepted, it would create an ironic situation where a researcher photocopying a page from a book in a public library for the purposes under Section 52(1)(a) of the CR Act would be entitled to the defence of fair dealing, but if the same researcher scans and stores the said page electronically, no such defence would be available.

(iii) The use by OpenAI does not amount to “private or personal use, including research”

Regarding the third requirement under the purpose test, that the use by OpenAI does not amount to “private or personal use, including research”, relying on judicial precedents and dictionary definitions, the Court held that the term “private” cannot be confined to an individual and can apply to a closed group or a company. While the expression “personal” may be confined to individual persons, the term “private” would include other private entities, including private companies. The Court further noted that OpenAI stores the literary works in a closed space without access to the public. The data obtained by the LLMs for training purposes is used for private purposes and is accessible only to the LLM models themselves. Considering that no human entity has access to the said data, the Court opined that the use is purely private.

The Court also noted that the process of training of the LLMs underlying ChatGPT can be considered as a form of “research” for generating new knowledge and advancing AI systems. It was the Court’s view that research is normally a closed activity and is not disclosed to the public – it is always the output of the research that is communicated to the public. The Court applied the ‘doctrine of updating construction’ to the meaning of “research”, considering that in 2012 when Section 52(1)(a) was last amended, the legislature could not have imagined the advent of AI or LLMs. The doctrine of updating construction gives effect to the true intention of the legislature by interpreting words in a statute in accordance with changing social conditions and developments in science and technology. Accordingly, the Court held that the process of training LLMs undertaken by OpenAI using stored literary work of ANI falls under “private or personal use, including research” as provided in Section 52(1)(a) of the Copyright Act and fulfils the purpose test.

Fairness test

The Court observed that different Indian Courts have adopted different tests according to the facts and circumstances of the cases before them and there is no single test which has been adopted uniformly or consistently by the Indian Courts. Further, it was noted by the Court that there is a broad consensus among counsel in this case that the four-factor test from United States is not applicable in India and that there is no single uniform test for examining whether or not an act under Section 52(1)(a) constitutes fair dealing.

In these circumstances, the Court identified three factors for determining fair dealing in the use by OpenAI of ANI’s work.

The first factor is whether OpenAI’s use of ANI’s original literary works is limited to training ChatGPT. In the absence of any evidence to the contrary, the Court concluded that the storage by OpenAI is only for the purpose of training OpenAI’s ChatGPT.

The second factor is whether the use of ANI’s literary works by Open AI would result in economic competition and would prejudice the legitimate interests of ANI, thereby causing actual or potential damage to ANI. Noting that, except for bare averments, nothing was placed on record by ANI to show that they have lost any market share or there has been a reduction in subscription revenues due to operations of OpenAI, the Court opined that there is no resultant market substitution of ANI’s works. 

The third factor is whether the functions performed by OpenAI through ChatGPT serve the overall public interest. The Court found in favour of OpenAI here as well, noting that the public benefits flowing from such trained LLMs underlying ChatGPT are considerable and go beyond analysing and generating text, and extend to improving access to information, supporting education, assisting scientific research, facilitating software development, enabling translation and communication, and creating tools for persons with disabilities.

Thus, the Court found that OpenAI satisfied both the purpose test and the fairness test under Section 52(1)(a) and held that OpenAI’s acts of storage of the literary works of ANI for the training of its LLMs would not amount to infringement.

Balance of convenience

Finally, the Court found that the balance of convenience also favoured OpenAI in rejecting ANI’s interim injunction application for the following reasons: