Shauryaveer Chaudhry & Shreya Banchhor

In this episode of the LSPR Podcast, Shreya Banchhor and Shauryaveer Chaudhry sit down with Mr. Adarsh Ramanujan, the amicus curiae in the ANI Media v. OpenAI proceedings.The conversation examines the copyright implications of AI training, including web scraping, storage, tokenisation, vectorisation, and memorisation. Mr. Ramanujan discusses Sections 14 and 52 of the Copyright Act and the Court’s approach to the research exception. The discussion also explores the implications of the case for news organisations and creative works. It concludes by considering the broader policy questions surrounding AI and the need for legislative intervention.

LISTEN TO THE PODCAST

Shreya Banchhor
Today, we are hosting an LSPR podcast on the ANI vs OpenAI Symposium with Mr. Adarsh Ramanujan, who is one of India’s leading patent and IP litigation counsels. He operates as an independent counsel with offices in Chennai and Delhi, and while IP remains the core of his practice, Mr. Adarsh regularly acts across commercial litigation, arbitration, insolvency, competition law, and regulatory disputes, and international trade and investment matters. Trained in both science and law, and a registered patent agent, Mr. Adarsh began his career in intellectual property and has since built a full lifestyle practice covering patent, trademark, design, copyright, plant variety, and trade secret disputes. On this note, today I, Shreya Banchhor, and Shauryaveer Chaudhry, are sitting down with Mr. Adarsh Ramanujan. Shauryaveer will give an introduction of the podcast series.

Shauryaveer Chaudhry
This is part of the LSPR symposium on the ANI Media vs OpenAI case, in which we have featured some distinguished professors as well as the other amicus in the case, Professor Arul George Scaria, along with other practitioners. We hope to understand the judgment and its policy implications. I will start us off with the first question for today. Thank you so much for agreeing to be here today, sir.
How did you approach the question of whether OpenAI’s use of AI and ANI’s content amounted to copyright infringement in the first place?

Adarsh Ramanujan
Yes, thank you Shreya and Shauryaveer. Happy to be here and share my thoughts now that the judgment is out. I approached the question like any other case that I would, which is, understand the law and apply the law to facts. What took me some time to digest was to get a thorough understanding of how LLMs operate. More than the legal issues, it took me a long time to get a hold and grip of the internal functioning of LLMs. And it is a tendency that I have regarding any subject that I pick up: I want to make sure that I fully understand the facts that are involved, and then only come to the law.

So, I approached the question to first understand the facts: how LLMs train? how is an LLM created? what are the steps involved in the training? And then I went to the question of law. I don’t know if that is the kind of answer you are expecting, but that was the approach that I had.

Shauryaveer Chaudhry
That really helps our viewers understand how to go about their own research when they come across questions like this. In your opinion, there is a distinction between strong copyright content and some of the technical processes involved in using that content. Could you, for our audience, please explain that distinction? What ultimately led you to conclude the strong copyright content for training amounted to infringement?

Adarsh Ramanujan
Yes, as part of that exercise, I have consumed about 3 technical books on LLMs and quite a few papers on how LLMs were created dating back to 2017-18, and some even going back earlier on machine learning per se. What I learned from all this was that all that what we call as training is really a multi-step process; in fact, I remember my amicus submission even carried a pictorial representation dealing with all the steps. I have been told that it is leaked in public; but from what I remember filing with the court, is that all training starts with an initial stage of scrubbing the data across the internet.

This web crawling, as people call it, could be direct or indirect, in which you either directly crawl for the data yourself, or somebody else has crawled it for you, which is, for example, a common crawler, and then you lift it from there. Then, you create a copy for yourself on the server that is accessible to you, which means your private servers or whatever servers that you are employing as an LLM creator. That data is thereafter, which I called as raw data, everything that is a scrape from the internet is put in there. This raw data is then filtered and deduplicated etc. In this process, most of the LLMs also remove offensive content etc., the degree of which depends upon the LLM. It is not because of any legal regulation on what you can or cannot do, because at the time most of these LLMs were trained, regulations were not in place. Even today, these regulations are non-existent in most jurisdictions, so there is more filtering of raw data there. Once that raw data is filtered, there is a further step called tokenization. Ultimately, everything is translated into numbers, so that every word is given a number. Even words are sometimes broken down into components, and those components are given token IDs so that everything is in a number format, not in English words. This thereafter goes through a process called as embedding or vectorization. This last stage is really the heart of an LLM – it took me an extremely long time to understand, because it has got nothing to do with law in that sense, but it is a very abstract mathematical concept. It is about data, largely metadata, and it is not even a metadata in the generic sense. For example, if I have a song, then the metadata about the song would be: how many times has it been played? which format is it in? where was it originally downloaded? This is the conventional understanding of metadata. But this is a very different form of metadata because you are trying to encode information about a particular token in a mathematical graph. So, if you say the word ‘black’, they will assume certain coordinates in time-space because of its relationship with, for instance, the color white. It will occupy a certain coordinate in another graph because of its relationship, for example, with the animal penguin. It might probably have a relationship with the black hole, and based on that, it will have a certain location on a graph, if you have to plot the data in an X-Y coordinate. But the funny part about this vectorization is that we usually understand graphs as consisting of 2 coordinates, X and Y, or sometimes 3 coordinates, X, Y, and Z; however, in LLMs, there are thousands of coordinates. There are not just 2 or 3 axes, there are hundreds and thousands of axes on which a particular data point is coordinated. That took a long time for me to observe.

When all of this is done and the training actually takes place, the LLM is fed with snippets of the data. The baby LLM, i.e., untrained LLM carries its own weights. For example, you feed it that time=0 and then you will forward a snippet of the data and ask it to predict what is next in the series. This prediction is compared with the original to see how far or how close it is to the original, which is also a mathematical relationship. Based on that, the weights in your LLM will be modified. You will run this process again and again, and you will compare the predicted output with the expected output, the raw root, and then you will tinker with the weights in the LLM. This is that entire gamut that actually takes place when you train an LLM.

When I understood the full gamut, the sense that I got was that none of these steps have got anything to do with copyright law and I think it would have been unfair to the court because as the amicus, my job and duty is to assist the court in reaching a conclusion. I am not for the plaintiff nor for the defendant, therefore, I am supposed to be objective and neutral even if I am a counsel for a party, because my duty as an officer of the court always stands ahead of the duty as a counsel for the party. But nonetheless, it was absolutely important when I understood how LLMs work. In my field of vision, I felt it was absolutely important to inform the court that training is this entire gamut of processes, some of which are completely alien to the application of copyright law. Further, to decide whether something is infringing, it is unfair to paint the whole thing as training and then color everything with that one color which is why I thought it was fair to divide it up.

Once I understood this, then it was the application of law to those facts. This is something that the court also recognized: if you see the final judgment around para 145, there is an acknowledgement in the order on the evolution of the law: once we became members of certain international conventions, Section 14 of the Copyright Act was specifically amended to include storage of works in electronic media or digital media as an exclusive right. There is a specific amendment including it, which means the first half or the first part of the training, which is comprised of accessing and storing everything in one place, is a clear case of impingement on section 14. The subsequent steps that come, which comprise the vectorization and the embeddings, were the ones I felt were a little tricky.

On vectorization, I think there is an arguable case from both the sides on whether it amounts to an ‘adaptation’ within the meaning of the Copyright Act. As you know, adaptation of a work is an exclusive right to the copyright owner. The funny part about this in our statute is that it is so broadly worded that you can include arrangement or rearrangement as equivalent to adaptation. I felt that that was an arguable point, but my opinion was that it is not an argument that should be decided at an interim stage; I thought that it required a much more detailed analysis.

Regarding the next step, which is comprised of the embeddings/vectorization: my view to the court was that that stage involves nothing of copyright, because the LLM or the software behind it is analyzing the interrelationship between words, the rules of grammar, the frequency with which some words are coming or not coming. That entire metadata about certain words and certain expressions has got nothing to do with the copyright in the original work. Further, as I said, embeddings or vectorization are essentially mathematical concepts, which gives you points in a mathematically imaginary graph with multiple dimensions. It is really a math concept. It has got nothing to do with the copyright aspect in the underlying work. I have given a clear opinion to the court that this cannot amount to any copyright infringement in any sense.

Then came the final step, which is the actual LLM. When you train the LLM, as I said, at t=0, the baby LLM has to be fed with something. It will provide an output, you will compare it with the original, you will again feed it, and you will keep correcting. This, I opined, involves a recursive use of extracts of the copyright work, because unless you do it multiple times, you cannot change the weights and get it to be an accurate LLM. My opinion stated that this is storage in electronic form, because to do this process i.e., to put the snippet in, pull it out and compare, you have to keep it stored, whichever work that you are using for training. So, that will involve storage.

But I would say there was another angle to this aspect, which I think you will notice in the judgment as well, which is memorization, which I think formed the heart of the judgment. And truly, memorization is a phenomenon that you will experience only in the output stage. As I said, the LLM’s job is to predict what comes next. Its job is not to memorize the training data. It should learn from the training data to predict what comes next. But data shows, including the statistical analysis, that there are multiple instances where the LLM, for reasons I think people are still figuring out, tends to memorize the raw data in its original form. Because in the recursive process of learning and retraining, it so happens that the LLM has decided that that is the best way to keep the answer accurate i.e., to memorize. I am sure you would have seen in the earlier versions of the LLM, and there are actually published papers from Stanford professors and Yale professors which show that actual literary works were being regurgitated verbatim in output.

So that is memorization where, although the training data is supposed to be removed, the LLMs weights, rightly or wrongly, have memorized the core work in its original form. My opinion expressed concern in the LLMs regarding this aspect. But whether such memorization has occurred in the training process is something you can only learn and detect at the output stage, and the plaintiff, of course, bears the burden to show that memorization has occurred. I think you will notice in the judgment that there is a prima facie finding by the court that it has not been proved based on the documentary evidence.

So, that is how I split the steps, and for each step, I applied section 14. My opinion was that when there is storing before the training and a temporary storage during the training, the recursive storage, those are the only two steps where I could clearly see infringement, at least under section 14, and not in the vectorization and tokenization.

Shauryaveer Chaudhry
Thank you for taking us through the entire process. I think that really helps break up the entire process of how LLMs work and how the court actually approached this.Do you think that the larger question at stake here is whether it is a policy question that the court should be deciding, or do they ultimately require legislative intervention?

Adarsh Ramanujan
If you are asking the question on whether the Court should be deciding this or not, it is not like the court is suo moto taking cognizance on a writ petition or a PIL. It is a commercial suit, where there is an aggrieved plaintiff and there is a defendant who feels that he is in the right. And it is the court’s duty to decide the facts before it according to the law. Of course, there are always instances where a court can come to the conclusion that a particular thing needs legislative intervention and, on a number of occasions, courts have taken this approach, even in commercial matters. For example, in patent cases, you will see exhortations in a lot of judgments saying that the Court have decided the case today, but the legislature must look into this and consider a clarification or an amendment, etc. Thus, there are sometimes such exhortations in judgments on commercial matters on such tricky points of law.

Nonetheless, if you are asking the question of whether the Court must decide or not, I think, there is not much choice involved. If a case comes, a court must decide based on the law as it stands. But if we take a step back and look at it from an academic perspective, I think the Parliament surely needs to intervene. There is a very tricky balancing exercise that is involved and perhaps it is going a little askew. Given the push towards developing AI and for not being dependent as a sovereign, there is a heavy push from all jurisdictions, including our country, to be independent in this respect. Sometimes I feel it is going too far. Perhaps, authors and creators don’t have the required protection that they need. And I also feel as an academic that sometimes the government wants to take the easy approach.

I am a big fan of the Hitchhiker’s Guide to the Galaxy. There is one character who always says this is a this is a problem covered by the SEP field. Somebody asked the character: what is this? It actually means Somebody Else’s Problem field. I have a feeling that sometimes we tend to take that approach. Why should I take it upon myself to deal with this? Let the courts solve the problem for us. In my opinion, on issues of this nature, it is not a great idea to leave everything up in the air for the court to call upon. Also, I don’t see the AI issue as that siloed question on copyright law, training and output. It is ultimately what the court’s narrow jurisdiction was in this case. I think there are much, much larger issues when it comes to AI. And it is really high time that the Parliament intervened with some kind of a sui generis legislation, which covers everything across that technological space.

Shreya Banchhor
Now that we have seen how the problem of AI is much more than the copyright jurisdiction: on research, the court applied the doctrine of updating construction, holding that research is no longer confined to humans and that limiting the exception to human actors would be regressive. On this basis, what, according to you, is the limiting principle, because if machine processing counts as research, what other commercial data processing activities now fall outside the purview of Section 52(1)(a)?

Adarsh Ramanujan
It is a tricky situation because I think that it is quite an expansive read. As you may know, or at least it may be evident from the recollection of my arguments from the judgment, I had suggested in my opinion to not take such an expansive approach.

At a fundamental level, Copyright law is human-centric, it evolved to be human-centric, and it continues to be human-centric. There are several potential dangers of taking this approach: if you apply updating construction to research, and certainly updating construction is a very established proposition on statutory construction. If it applies to one side of the supply chain, it should equally upgrade to this side of the supply chain, which is the generation of works. We already have an order from the copyright board, and a patent order has also been passed in this respect that no protection can be granted to AI generated works. Many people argue for taking an expansive interpretation of updating construction. But the answer is clearly in the negative: we can’t take this approach. The proposition supporting that conclusion is that copyright law is human-centric and is intended to enable humans to generate and be creative. Equally, when section 52 talks about research not being infringement, it is also human-centric. It is to enable human beings to conduct research. In my opinion, there is a certain element of risk in applying updating construction in an extreme manner, because then it removes the human out of the equation. And in fact, the answer to the question by the court was that because there is no human in the loop order, section 52 applies, which I think takes it far away from its purpose. Updating construction, in my opinion, cannot be de hors the object in purpose.

Shauryaveer Chaudhry
Thank you for that, sir. Section 52(1)(b) already deals with transient incidental storage in the technical process of transmission. Why then did the defense turn on 52(1)(a) instead? Does section 79 of the IT Act have any kind of role there?

Adarsh Ramanujan
That is a question best left to the defense counsel as to why they didn’t. But in my view, as I argued in my first opening submission to the court, Section 52(1)(b) does not apply. Because if you see the language of the section, it states, ‘transient or incidental storage in the process of transmission,’ which, as I explained, does not involve any transmission of the work to the public. That exception is really meant if there is a technical process of transmission. For example, when I open a browser and there is a cached version of the website, that is a copy or an incidental storage in the technical process of transmitting that website to me. There is thus, a transmission involved and, in that transmission, there is an incidental copy called a cache. That is the purpose of 52(1)(b). Not all incidental and transient storage are included, but only which are done in the process of transmission. I think the defence counsel didn’t argue it because there was no transmission in this stage.

Shauryaveer Chaudhry
What do you think this case could mean for news organizations and journalists going forward? What would they keep in mind provided the appeal yields the same result?

Adarsh Ramanujan
The message is that if you are an American news organization or journalist, you can sue and get a settlement in the US. But if you are going to come to India, you have got no hope for a settlement, because there is no injunction, as there is a prima facie finding of fair dealing.
The court specifically confines the questions to a factual pattern where the LLM developer is scraping publicly available information. You will see that expression in the judgment, and one main defense that the court has also taken into account is the fact that any news organization has a technical mechanism of preventing such scraping of information i.e., there is a technological measure possible. The Court did not get into whether that technological measure is effective or not, but there is a finding that there is a measure possible.
What I think is likely to happen is that you might be able to get a headline or a snippet, but you will never get the full content because it will be under a paywall or a subscription wall. You are likely to see an effect in this form; you will end up creating more hurdles to access to information because people now want to stop or prevent LLMs from using that data. That will be the likely consequence, which I don’t think helps anybody, but that is what people will start doing.

Shauryaveer Chaudhry
That is very interesting, sir. Coming to the last question, the court left open the question of whether the same reasoning extends to more expressive work, such as music, film, or literature, where the fact, expression, distinction carries less force. The Indian music industry was also among the interveners. What are your thoughts on this?

Adarsh Ramanujan
Much to the chagrin of a lot of the interveners, my opinion to the court was to not allow interveners, because of the fact that this is not a PIL.

The Court, for example, in this case, as you pointed out, says that when it comes to protection itself, there is a sliding scale. The less expressive, that is, more factual content like news, is on the lower end of the spectrum when it comes to protection, but less factual, more creative works like music or art, are at the higher end of the spectrum of protection. When a court assesses infringement, that balance may shift, depending on what is copied and what is a substantial copy.

The paragraph that you are talking about has a lot of connection to the memorization concept that I told you. Now, if you take the case of GEMA vs OpenAI, the German judgment where there is a finding of infringement, they were able to demonstrate memorization because the output was a verbatim reproduction of the work. Here, you will see that on the facts of this case, the court comes to the conclusion that the only reproduced part are facts. That is the finding and I think that it is what changes the scale. It means, in my view, that they are based on different set of facts and the court could have reached an opposite conclusion. The non-existence of memorization and the lesser expression involved in a news article tilted the case a little bit.

Shauryaveer Chaudhry
That is a very interesting point, sir. Thank you for being with us today. We really enjoyed having you here and I can speak for Shreya when I say this, we learned a lot from this interaction. It was lovely having you today.