Aniruddh Bhatia & Dr. Victor Vaibhav Tandon
In a recent judgement, the Delhi High Court dismissed an application under Order XXXIX, Rule 1 and 2 of the Civil Procedure Code, 1908 (“CPC”) seeking an interim injunction in a suit for permanent injunction filed by ANI Media (“ANI”) against OpenAI OPCO LLC (“Open AI”). ANI sought to restrain OpenAI from the unauthorized storage and reproduction of content from its copyrighted news articles.
In a significant ruling, the Court clarified the scope of the doctrine of Fair Dealing under Section 52 of the Copyright Act, 1957 (“the Act”), and extends its application to the storage of copyrighted News articles for training Large Language Models (“LLMs”) and the subsequent use of such data in generating responses.
Key-Facts
ANI had filed the suit in 2024, seeking inter- alia an injunction restraining OpenAI from “storing, publishing, reproducing or in any manner using” ANI’s works, published on its platforms or by its licensees.
ANI’s case of infringement was two-pronged:
- Training Claim: ANI alleged that Open AI accessed, stored and processed its original literary works published online to create datasets for training its LLMs under lying its user interface (“ChatGPT”).
- Output Claim: ANI further alleged, such copyrighted content was reproduced through outputs generated by ChatGPT, amounting to infringement under the Copyright Act, 1957 (“the Act”).
Key-Findings
- Indian Courts have the Jurisdiction to the adjudicate claims regarding processing of Data stored in Foreign Servers
- Indian courts maintain territorial jurisdiction over processing of data by foreign AI developers in terms of Section 62(2) of the Act and Section 20 of the CPC since ANI has its principal place of business and registered office within the Court’s jurisdiction and Open AI offers its services to users in India, including within the Court’s jurisdiction.
- The Court applied the principle of the “chain of events”, and held that the location of Open AI’s server’s being outside the territory of India does not preclude the Court from applying the provisions of the Act extra territorially, since the process of infringement begins with the data scraping and transmission from India. The same is consistent with the reasoning of the Delhi High Court in Neetu Singh v Telegram 2022 SCC Online Del 2637, wherein the court held that the availability and circulation of the data within the territory of India attracted the application of the Court’s Jurisdiction
- Output Claim- Outputs Generated by Open AI do not amount to Infringement
- Copyright protects original form and expression, therefore, the scope of copyright protection available to news publisher is limited to the work’s linguistic expression, not extending to underlying news facts or direct quotes. Accordingly, news outlets need to establish substantial and material reproduction / regurgitation of its linguistic expressions in AI user interface generated outputs.
- The outputs filed with the Plaint to demonstrate unlawful storage of ANI’s works relied on information published by ANI after April 2022, i.e., the cut-off date of Open AI’s dataset for the model integrated into ChatGPT during the institution of the suit (ChatGPT 4.0). As ChatGPT generated these responses using the Retrieval-Augmented Generation (“RAG”) technique, which retrieves relevant information from online sources in response to a user’s query, the Plaintiff failed to establish that Open AI unlawfully stored or memorized its works.
- Since ANI was unable to illustrate that the responses amount to substantial reproduction of the contents contained in its news articles alleged to be misappropriated for creating the data set, ANI did not succeed in its Output Claim.
- Training Claim- Machine Learning of News Articles is permissible under the Fair Dealing Doctrine
- While the News Articles are publicly accessible, ANI retains the exclusive right to reproduction and communication to the public in terms with Section 14 of the Act.
- The Court analysed the scheme of Section 52(1)(a), and held that the commercial nature of Open AI’s use of the works does not estoppel Open AI from claiming the defence of Fair Dealing under the research and private use exception. It was held that the deliberate omission of the term “non-commercial” from Section 52(1)(a) in contrast to other provisions under Section 52 showcases clear legislative intent to permit commercial fair dealing.
- “Private” use includes closed corporate processing where raw data remains inaccessible and unreadable to third parties post training.
- Multi- functional AI platforms are not in direct competition with news publishers and therefore ANI was unable to show case that the Open AI’s use of ANI’s articles caused prejudice.
- The Court applied the doctrine of updating construction and expanded the scope of “Research” from human cognitive reading to include machine learning, automated text extraction and pattern analysis.
- Machine learning involves processing of raw copyrighted data to extract non-expressive elements of the works. The copyrighted material is not published or accessible by the public, therefore copyrighted material is not reproduced to consumers for profit, thereby the usage of the copyrighted material was held to be fair.
- OpenAI’s access to articles owned by ANI, thereafter licensed and published on the platforms of ANI’s licensees is lawful.
- Noting that a ruling of infringement under Section 51 of the Act is subject to the provisions of the Act, the Court held that training of the LLMs on ANI’s works fell within the scope of Fair Dealing under Section 52 of the Act and did not amount to infringement as per the scheme of the act.
- Balance of Convenience and Irreparable Harm not established
- The Court noted that ANI’s offer to license its content to Open AI for USD 7.5 million indicated that the injury claimed by ANI was quantifiable and could be compensated in monetary terms, whereas an injunction could significantly disrupt Open AI’s operations.
- Artificial Intelligence development relies on free access to information in public domain. It would unfairly prejudice AI developers to impose a prerequisite to seek a license from multiple sources for the development of a tool beneficial across several sectors.
- ANI retained the technical ability to block web crawlers and paywalls to protects its content but chose not to engage such measures, significantly weakening its claim for interim relief
Accordingly, the Court dismissed the application seeking interim injunction against Open AI, holding that ANI was unable to make out a prima facie case of infringement against Open AI based on Open AI’s utilization of ANI’s works.
Key-practical implications
- Indian Courts retain the jurisdiction to adjudicate on claims of copyright infringement against foreign entities unauthorized utilizing works published by Indian individuals / entities for LLM training.
- Copyright in News Articles is limited to the form and manner of expression published and reproduction of the subsisting facts in any manner adapted thereafter does not amount to infringement thereof.
- Multi-functional AI tools are beneficial to the public at large and do not replace news syndication services, therefore cannot be deemed to be causing financial harm unless direct competition in services / commercial operations is established.
- Publishers seeking to protect their literary works from being processed must proactively engage measures such as paywalls and crawler blockers.
- Machine Learning involved in LLM training utilizing works lawfully accessed may be exempt under the private use and research principles under Section 52 of the Act.
