A fascinating ruling from the High Court of Delhi in India regarding the very high-profile case of ANI Media v OpenAI (decision here). Asian News International (ANI) is one of India’s largest news agencies, which supplies text and video syndication to broadcasters and publishers across South Asia. They sued OpenAI alleging copyright infringement during the training phase, as well as reproduction of its articles on the output phase. This decision is an interim order deciding ANI’s application for a temporary injunction against OpenAI, so while interesting on its own right, the trial continues. This is however very important because it deals both with a few questions regarding memorisation, RAG, and jurisdiction which we have covered repeatedly here at Llama Towers.
Spoiler alert, OpenAI won this first stage.
The case and jurisdiction
The complaint from ANI Media follows a familiar pattern that we have seen in various cases across different jurisdictions, they claimed that OpenAI copied works that they own and used them to train ChatGPT, and they also claimed that the LLM could reproduce various articles as a result of that training, providing several allegedly infringing outputs. So this is both an input and output case.
The court had first to decide whether it had jurisdiction. The defendants argued that the alleged infringement took place in the USA, where the training occurred; that the training happened there was not in dispute between the parties, or among the various intervening third parties and amici. What was disputed was what followed from it. The court held that Indian courts had jurisdiction because the claimant’s main place of operations is in India, with the result that copying carried out in the USA fell within its reach.
The defendants argued separately that the Indian Copyright Act has no extraterritorial effect. The judge accepted that the Act could not apply abroad, but held that, prima facie, training and the allegedly infringing outputs formed part of a single chain of events, so the Act could apply to the output phase. This is the only part of the ruling I find dubious. I have argued repeatedly that there are good reasons to treat training and outputs as separate acts, and collapsing them into one chain of events does exactly what the extraterritoriality limit is meant to prevent.
The judge then goes on to discuss the core infringing points, namely whether training an AI is copyright infringement, and whether there were infringement in the outputs, and here is where OpenAI scores its first victories. The judge went directly to the outputs as this would be the main action that took place under Indian jurisdiction.
Outputs
The plaintiffs alleged that ChatGPT was capable of producing directly infringing outputs, presenting several examples that they argued were substantial reproductions of their works, and therefore evidence that ChatGPT had memorised them. However, this did not go well. The outputs they provided came from either ChatGPT 4 or ChatGPT 4o, and OpenAI was able to show that these models had been trained in April 2022 and April 2024 respectively, while the articles allegedly infringed were published later, mostly in August and September 2024. I checked the dates, there are various dates floating around, but one thing is true, none of the models used in the complaint were trained after the articles were published, so none of the articles, could have been in the training data. So how was ChatGPT able to reproduce some of the works? Thanks to RAG (retrieval augmented generation). I’ve written about this before, but for present purposes RAG operates like a search engine, retrieving information from the Web in real time, which allows models to give up-to-date responses about facts postdating their training.
The training dates were a killing blow for the output infringement claims, as none of the outputs were reproductions of the works in question because models do not store the training data. Much in line with what we have continuously argued here, models do memorise sometimes, but this has to be proven in the outputs. ChatGPT was able to discuss the articles presented by the claimants, but these were not substantial reproductions, but summaries of publicly-facing websites retrieved on demand. Moreover, facts in news are not copyrightable, the threshold for originality in India is the “skill and judgement” test, and in this case just reproducing facts and some syntax from the articles was not infringing. The judge commented that ANI had not been able to produce any verbatim reproduction of their works, and it was even worse than that, as ANI had used adversarial prompts to try to get the model to reproduce the works, even telling the model to try to reproduce something from an article “exactly”.
The discussion by the judge here is really worth a read, the fact that the articles were not in the training data was damning, and the judge argued accurately that any summary or similarity was a result of RAG. Because this had not been argued to the court, this could not be ruled prima facie, and it would have to wait for the trial. The judge commented that this could fall under a communication to the public, but this was never argued by the claimants. I believe that RAG is not a communication to the public, so I am not surprised that this was not argued by the plaintiffs, but I digress… So in the end the judge analysed the outputs and could not find anything that amounted to a substantial reproduction of the articles.
On the matter of memorisation leading to outputs, the judge said:
“As highlighted above, the illustrations given in the plaint are post the training of Open AI’s LLMs and a case for memorization of ANI’s works on the basis of the said illustrations cannot be made out. Therefore, at this prima facie stage, the contention of ANI that Open AI permanently stores the training data in order to memorize and regurgitate ANI’s works cannot be accepted. At best, these are disputed questions which can only be determined during trial upon parties leading evidence.”
So the whole output part of the judgement is an undisputed win for OpenAI, with a few aspects to be determined during trial.
Inputs
The input question did not go well for the plaintiffs either. The plaintiffs argued that their works were copied and used to train OpenAI’s models, which is copyright infringement. The defendants argued that this was true, but that works from ANI make an incredibly small part of the totality of training data, and that only non-expressive elements of the works were extracted, namely grammar, syntax, and linguistic patterns. The defendants also argued that any reproduction amounted to private use, and it was also temporary or incidental use, which are exceptions under Indian copyright law.
That reproductions of the work had been made was uncontested, so the judge was happy to take this as given. The issue was whether such a copy falls under one of the aforementioned exceptions. Here the judge has a lengthy and interesting discussion of the nature of private copying under Indian law that I won’t cover in detail, but the highlight is that the commercial nature of the training does not in itself defeat OpenAI’s reliance on the exception. The judge was also clear that as long as the copies were made legitimately, not from shadow libraries or by circumventing paywalls, the use would be legitimate, and that it is private because at no point are the copies distributed to the public (a point I have also made repeatedly here). Interestingly, India has no text and data mining exception, so the judge applies the private-copying-for-research provision, interpreting “research” broadly enough to cover AI training. The judge then concludes that:
“Therefore, on a prima facie view, from the above analysis, this Court is of the view that the process of training LLMs underlying ChatGPT undertaken by Open AI 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.”
This is huge, as to my knowledge this is the first time that a court has ruled that AI training falls under private use. This was helped by the fact that no output was found to be infringing, and therefore no copies of the works were made available to the public. Because the use was private, the court did not need to delve into the transient or incidental exception.
The decision then goes on to discuss whether AI training can be considered fair, following the Indian fairness test. Here the court asks three questions:
“a. Whether Open AI’s use of ANI’s original literary works is limited to training its LLMs underlying ChatGPT?
b. Whether the usage 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?
c. Whether the functions performed by Open AI through ChatGPT serve the overall public interest?”
The answer to the first question is yes because the plaintiffs weren’t able to provide any evidence of memorisation or reproduction of their works in any output.
The answer to the second question is also negative, both OpenAI and ANI are in very different businesses, and people use ChatGPT for all sorts of uses that do not enter into competition with ANI’s products. The judge pointed out that even when a user asks for headlines from ANI, the output does not provide anything more than topics and at most some titles of current articles, which is in no way direct competition. Here the judge cites Bartz v Anthropic, and Kadrey v Meta, noting that there the outputs generated by an LLM are considered to be transformative. The judge also cites Google Books.
The answer to the third question was positive, and the judge makes a strong case of why chatbots serve the public interest:
“The public benefits flowing from such trained LLMs underlying ChatGPT are considerable. They are designed to assist users in analysing and generating text, improving access to information, supporting education, assisting scientific research, facilitating software development, enabling translation and communication, and creating tools for persons with disabilities. The technology is also capable of supporting research, learning, and knowledge dissemination at scale, thereby advancing science and the useful arts in a broader societal sense. […] Training LLMs underlying ChatGPT therefore contributes to advancing scientific knowledge, developing innovative computational tools, disseminating information, promoting education, enhancing accessibility, and advancing artificial intelligence research”
Very interesting and strong defence of chatbots from a judge.
Concluding
The judge concluded by denying a preliminary injunction, so the trial continues. It will be interesting to follow, but we can already draw some takeaways.
First, the jurisdiction question appears to be leaning in favour of allowing lawsuits to proceed where the effects of training are experienced. Second, this case proves once again that if you cannot produce a credible infringing output, the case will rest on the input phase, where judges seem less willing to damn an entire model for copies made during training. Third, I found the reliance on the private use exception to be an astute move in a jurisdiction that has no text and data mining exception.
And finally, I think we often underestimate how willing judges are to look at the actual usefulness of LLMs when ruling in these copyright cases. Judges are AI users too, and here we have the Honourable Amit Bansal, who appears to understand the technology well, and this evidently shows in the decision. I was really impressed with the clarity of the opinion, particularly in the memorisation discussion, something which I hope other judges will replicate when dealing with this question in the future.
This is one to watch. OpenAI scored an important first victory here; we will see if it translates at trial.
0 Comments