- with Inhouse Counsel
- with readers working within the Accounting & Consultancy and Law Firm industries
Since the launch of large language models(‘LLMs’), there have been debates regarding the legality of use of publicly available copyrighted works for training of LLM models. In a significant recent development, the Delhi High Court(‘Court’) has settled that debate by holding that use and storage of copyrighted works for training of LLMs do not, in the circumstances of the case, constitute copyright infringement and qualifies as ‘fair dealing’1.
Background of the Case:
A suit was filed by ANI Media Pvt. Ltd. (‘ANI’), a news agency, against OpenAI Opco LLC (‘OpenAI’) under the Copyright Act, 1957 (‘Act’) seeking injunction against OpenAI’s use and storage of ANI’s copyrighted works (i.e., news articles, interviews etc.) for training its LLMs and generating responses through its AI chatbot ChatGPT.
After filing of the suit, multiple intervention applications were filed by the relevant stakeholders – some supporting the case of ANI and others supporting OpenAI.
This article shall cover the Court’s findings on the major issues, as under:
Jurisdiction of Indian Court
At the outset, the Court dealt with the issue whether the Indian courts even have jurisdiction to entertain the suit, considering that OpenAI’s servers are located in the United States of America (US). OpenAI claimed that two separate causes of action had been joined in the suit, namely: (i) the use of ANI’s works on OpenAI’s servers located outside India to train LLMs underlying ChatGPT and (ii) the responses received from ChatGPT that are allegedly similar to ANI’s copyrighted works. As OpenAI’s LLMs are trained in the US, it questioned the applicability of the Act outside the territory of India.
On the aspect of jurisdiction, the Court delved into the issues of territorial jurisdiction as well as applicability of the Act. The Court held that it had territorial jurisdiction because: (i) ANI’s principal place of business and registered office are located within its jurisdiction, (ii) OpenAI specifically targets/ offers its services to users across India, including those located within the Court’s jurisdiction; and (iii) the responses generated by OpenAI also took place within the Court’s jurisdiction.
As regards the applicability of the Act in relation to the training of LLMs outside India, the Court observed that the storing of ANI’s works on US servers is a terminal step in the chain of events which begins with access of copyrighted works from India and transmission of the same abroad. While the Court acknowledged that the Act cannot have jurisdiction outside the territory of India, it clarified that the Act does not require it to sever the chain of events and examine only the last step. Accordingly, the Court held that act of training LLMs outside India will also be subject to the provisions of the Act, because the responses are generated in India and are based on the scraping of ANI’s copyrighted works and the training of OpenAI’s LLMs.
Use of Copyrighted Data to Generate Responses
It was not disputed that news articles and interviews published on ANI’s digital platform are “original literary works” and hence, ANI enjoys the exclusive right to reproduction of these works and the right to “communicate it to the public”.
However, ANI claimed that OpenAI’s LLM model memorizes the content of ANI’s copyrighted works and specific responses are generated by ChatGPT (OpenAI’s famous AI chatbot) through memorization. Consequently, original literary works of ANI are communicated to the public by OpenAI, which amounts to infringement under the Act. To determine ANI’s claim, the Court framed two issues:
(i) Memorization of Copyrighted Works
The first issue was whether OpenAI memorizes and regurgitates ANI’s copyrighted literary works in the form of responses. To allege that OpenAI’s LLMs memorize the training data verbatim/ substantially, ANI submitted a few illustrations demonstrating substantial reproduction of its copyrighted works.
However, OpenAI informed the Court that the cut-off date for training of its models (relevant for the suit) was April 2024 and all of publications claimed to have been infringed by ANI were authored and published post April 2024. As such, the Court did not accept ANI’s memorization contention as ChatGPT responses could not have been produced on account of training of OpenAI’s LLMs on ANI’s literary works because the publications were made post the training phase.
As the claim of memorization was not accepted, the Court inferred that OpenAI is using Retrieval Augmented Generation (RAG) technique whereby LLMs retrieve relevant information from external sources (ANI’s website) before generating a response. Hence, the Court concluded that the instances given in the plaint are not a result of memorization, but they are in the nature of live links. However, as ANI did not plead infringement for outputs produced using RAG technique, the Court did not analyse this aspect.
(ii) Substantial Reproduction of Copyrighted Works
While the Court did not accept ANI’s claim of memorization, the Court agreed that OpenAI may likely be using Retrieval Augmented Generation (RAG) technique to generate the responses. Using RAG technique, LLMs retrieve relevant information from external sources (ANI’s website) and hence, the responses may have been generated using live links, rather than memorization. As such, the Court then decided to analyse whether ChatGPT’s responses are substantial reproduction of ANI’s copyrighted literary works.
It is a settled legal position that there cannot be any copyright protection in respect of facts and particularly in the context of news, copyright would subsist only in the form and manner of expression of news. Therefore, the Court analysed whether the expression used by OpenAI in its ChatGPT responses is materially similar to the expression used by ANI in its news articles.
The Court compared the similarities between the works of ANI and the responses produced by ChatGPT and observed that the response given by ChatGPT was not a substantial reproduction or nearly exact copy – though the essence of the response is similar as it is based on facts. As ChatGPT adds its own commentary to the news article published by ANI, it bears its own expression, which is dissimilar from the expression used in ANI’s articles.
As the Court did not find any substantial reproduction of the news articles of ANI by OpenAI, it held that there is no copyright infringement based on the responses generated by ChatGPT.
Fair Dealing
In its plaint, ANI contended storing of its copyrighted works by OpenAI (though temporarily) amounts to infringement. To counter, OpenAI argued that their storage falls under the scope of ‘private or personal use, including research’2 and therefore qualifies as ‘fair dealing’. To settle the issue, the Court undertook a two-step examination:
(i) Purpose Test
The Court first anaylsed whether the storage by OpenAI is for one of the specific purposes mentioned in Section 52(1)(a) under the Act, which lists certain acts which are not considered infringement of copyright.
OpenAI contended that their storage of ANI’s works fell within the scope of ‘private or personal use, including research’.ANI, on the other hand, argued that such private use/ research could not qualify for the statutory exception if undertaken for commercial purposes. The Court rejected this submission, observing that where Section 52(1) intends to restrict the protection to non-commercial use, it expressly says so3. As Section 52(1)(a) contains no such limitation, the availability of the fair-dealing defence is not excluded merely because the underlying use is commercial.
ANI then contended that OpenAI stores a copy of ANI’s works that has been obtained in an unauthorized manner. ANI argued that defence under Section 52(1)(a) can only be availed if the first copy of the work is a non-infringing copy4. However, the Court rejected this argument as well and held that the limitation of not being an infringing copy is only applicable to the incidental storage of computer programmes. The Court also observed that OpenAI does not obtain copies of ANI’s works from unauthorized sources or by breaking through the paywall of ANI, rather OpenAI obtains copies from information/ data that is freely available on ANI’s website.
Lastly, ANI argued that use by OpenAI does not amount to private or personal use, including research. On this aspect, the Court held that the word ‘private’ is not limited to individuals and it also includes a closed group or a company. Further, the Court gave a liberal and purposive interpretation to the term ‘research’ and held that the process of training LLMs should be considered ‘research’.
(ii) Fairness Test
As the term ‘fair dealing’ is not defined under the Act, the Court analysed various judgements of the Indian courts and observed that different courts have adopted different tests according to the facts and circumstances of the case before them. In line with Article 9 of Berne Convention for the Protection of Literary and Artistic Works (1886), the Court devised the following factors relevant for determining ‘fair dealing’ for the dispute at hand:
- Whether OpenAI’s use of ANI’s literary works is limited to training its LLMs?
The Court observed that OpenAI’s LLMs are not trained to reproduce or communicate the training material to the public. Instead, the training material enables it to generate new responses and perform novel tasks when presented with novel prompts. As such, it held that OpenAI has not used the literary works of ANI for any purposes, other than for training. - Whether the usage of ANI’s literary works by OpenAI would result in economic competition and cause damages to ANI?
While ANI is a news agency, OpenAI is an application which allows users to interact with LLMs. As such, the functions performed by ChatGPT are quite distinct from the functions performed by ANI. In the context of news, ChatGPT only provides a summary of news articles or reports carried by ANI along with a reference to its website. Hence, the Court held that use of ANI’s works by OpenAI would not result in market substitution of ANI’s works and consequently not cause any damage to ANI. - Whether the functions performed by OpenAI serve the overall public interest?
The Court was of the view that public benefits flowing from such trained LLMs are considerable. The Court observed that training LLMs contributes to advancing scientific knowledge, developing innovative tools, disseminating information, promoting education, enhancing accessibility, and advancing artificial intelligence research. Hence, the Court held that the factor of public interest is also established in favour of OpenAI.
Accordingly, both the purpose test and fairness test were deemed to be in favour of OpenAI and thus, the storage of ANI’s copyrighted works by OpenAI was considered to qualify as ‘fair dealing’ under the Act.
In relation to the balance of convenience and irreparable injury, the Court noted that ANI had not exercised the available option of blocking third-party web crawlers. Treating this as an additional consideration, and in light of its prima facie finding under Section 52(1)(a), the Court concluded that the balance of convenience favoured OpenAI and declined to grant an injunction.
Remarks:
Being the first of its kind, the ruling sets a precedent in the long overdue battle between AI and copyright law. With a view to prioritise the larger public interest, the Court has allowed use and storage of copyrighted works by AI companies for training their LLMs. While the order has only been passed on an interim application and the matter is yet to proceed to trial, the Court’s decision to interpret the provisions relating to ‘fair dealing’ in the context of AI should have the effect of extending the applicability of the Act beyond human beings.
The stand taken by the Indian Court also broadly aligns with the global trend of the courts in different jurisdictions treating the use of copyrighted material by the AI companies as permissible. However, as the technology continues to evolve, this may just be the first of many decisions to come in this domain.
Footnotes
1. I.A. 45300/2024 in CS (COMM) 1028/2024
2. Section 52(1)(a)(i) of the Act
3. Sections 52(1)(ad), 52(1)(k)(ii), 52(1)(1), 52(1)(n), 52(1)(o) of the Act
4. Explanation to Section 52(1)(a) of the Act
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
