New Delhi: The Delhi High Court has the jurisdiction to decide the outcome in the news agency ANI’s copyright lawsuit against OpenAI, said two amici curiae appointed by the court in the case in their written submissions.
They said that under the Indian Copyright Act, jurisdiction is determined by the place of business of the plaintiff, which is New Delhi in this case. However, they differed on whether OpenAI’s use of ANI’s data constituted copyright infringement and was permitted as “fair use”.
ANI’s copyright infringement lawsuit against American artificial intelligence giant OpenAI filed earlier this month, said that OpenAI trained its software on copyrighted literary works published by the member-publishers.
The issue of jurisdiction is the key argument that OpenAI has made, saying that since its servers that train and store data are not located within India, Delhi HC has no jurisdiction.
ANI’s lawsuit, that an industry body of book publishers, and digital arms of legacy news media including Hindustan Times have sought to join, is the first of its kind in India and follows a slate of similar lawsuits filed across the world against AI companies.
On November 18, the single-judge bench of Justice Amit Bansal had appointed academic Arul George Scaria and lawyer Adarsh Ramanajun as amici curiae to assist the court in the matter. Scaria and Ramanujan submitted their reports to the court on Monday.
Scaria said that even though OpenAI’s servers are located in USA, it is making its “interactive services” available to users in India, including those in Delhi, where ANI’s headquarters are located. “In similar circumstances, the Courts in India, have assumed jurisdiction,” his submission read.
Ramanjuan said that the principal place of the plaintiff’s (ANI) business is New Delhi, hence Delhi HC retains jurisdiction. He submitted that the location of OpenAI’s servers is “prima facie irrelevant” as far as establishing jurisdiction for copyright infringement under the Indian Copyright Act is concerned.
Both amici said that the court will have to examine whether OpenAI’s “opt out” option for people who do not want their content to be used for training was effective and easy to use.
Scaria suggested that during the trial, both ANI and OpenAI could also answer whether “unlearning” the information from content used during training was “technically and practically feasible”.
Ramanujan said that OpenAI’s opt-out policy was unclear and did not exempt infringement. The policy would only prevent future occurrences, and that since OpenAI relies on third-party web crawlers, limiting only OpenAI’s web crawlers may not even solve the issue, he said.
In the November order, the court had asked three other questions of law apart from that of appropriate jurisdiction — first, if storage of ANI’s news data to train OpenAI’s software was copyright infringement; second, if using this copyrighted data to generate responses was copyright infringement; and third, if OpenAI’s use of ANI’s data qualified as “fair use”.
Scaria said that storage of all kinds of copyrighted material — news related or otherwise – as part of training large language models (LLMs) is permitted under the Copyright Act. The court, however, will have to assess whether OpenAI store copyrighted materials for purposes other than learning, and if such purposes are allowed under the Act, he said. He submitted that preventing “temporary or permanent storage for learning” might undermine the broader coal of a copyright law, which is “to promote the creation and dissemination of new works”.
Ramanujan, on the other hand, submitted that “reproduction” by OpenAI during the training process was not protected under Section 52(1)(b) of the act. He submitted, “the iterative reproduction of the training data, whether in the original text or its corresponding numerical representations, amounts to infringement by reproduction”.
On whether using ANI’s content to generate responses was an infringement of its copyright, Scaria said that first, ANI must establish that it owns the copyright over the works in question, especially as “a substantial chunk of news related materials might be outside copyright protection in India”. He added that ANI’s position as a news agency, instead of a newspaper, magazine, or a periodical, will need to be examined where the authors/contributors are not necessarily ANI employees and thus a licensing or copyright assignment may exist between ANI and the contributors.
Second, for works over which ANI’s copyright is established, the court must assess to what extent copyright protection can be claimed. And third, “non-expressive uses of copyrighted works”, that is, paraphrasing the facts reported by ANI, “to provide more optimal responses to use prompts” may not be a copyright infringement, Scaria said.
Even in cases where exact phrases have been used in ChatGPT’s responses, he said, that the court will have examine the whether the purposes for which such responses were generated in verbatim were allowed under the Copyright Act and Indian jurisprudence. For instance, he said, “right to quote” is allowed in India in certain contexts.
Ramanujan, on the other hand, argued that using ANI’s content in outputs, that is, “the communication to the public at the output stage”, would be a copyright infringement.
He said that the court might need a trial to assess whether LLMs actually permanently stored all “raw data” but most technical literature indicated that the LLMs did not. Since ANI’s content is syndicated across multiple publications, the LLM may have “memorised” it from other publications to produce duplicates of the “raw data”, that is, ANI articles themselves, he submitted.
Ramanujan said that OpenAI’s use of ANI’s content was not permitted as “fair dealing” under the Indian act as the company is not a news agency or a newspaper reporting on current events or affairs, nor is it using ANI’s content for criticism/review. He said that by using the content for a commercial motive, it may not be allowed as “private or personal use” either.
Scaria, however, submitted that since the learning process for LLMs barely gave access to “external human beings”, and the “learning is primarily done with the objective of promoting research”, OpenAI’s storage of ANI content may be allowed as fair dealing/fair use. He added that since users are using services such as ChatGPT to gather “information about events and facts”, use of source materials to fulfil this “emerging function of LLMs” might be considered fair dealing.
Ramanujan submitted said that the court could consider whether OpenAI’s use of ANI’s works posed any “major economic threat” to the news agency whereby newspapers and news channels could cease subscribing to ANI and opt for responses from ChatGPT instead.
Scaria warned that the court must keep three things in mind: first, the court’s decision will affect future of AI development and deployment in India; second, given that responses from LLMs at times contain disinformation and LLMs may be misused to that end, access to legitimate information, including copyrighted materials, is important; third, since LLMs falsely attribute sources, the court could direct OpenAI to address such grievances promptly. For ANI, false attribution by ChatGPT was a major reputation concern.