Expert Explains | Why ‘lawful access’ may not be required for AI training in India

Though subject to appeal, the ruling gives AI developers such as OpenAI and Sarvam a measure of legal cover to train on contents without licensing agreements.

Arul George Scaria, law professor at National Law School of India University, Bengaluru, and an expert in intellectual property and competition law, assisted the Delhi High Court as an amicus curiae in the case. In an interview, he discussed the broader global context of how courts across the globe are approaching copyright protections surrounding LLM training.

Arul George Scaria, law professor at National Law School of India University, Bengaluru

Edited excerpts:

How does the Delhi High Court’s interim verdict bode for AI development and LLM training in the Indian jurisdiction?

In its interim order, the Delhi High Court has held that it is permissible for a machine to learn by training on copyrighted works. The Court reached this conclusion through a very careful reading of the copyright law in India, particularly the broader objectives of the fair dealing exception.

Overall, this is a very balanced and forward-looking judgement. It is good for innovation because people can now engage in training with more confidence. But at the same time it is important to note that the judgement isn’t a blank cheque because there could still be liability on AI companies for copyright infringement on the output side of LLMs, depending on the specific facts of the case. So overall, it balances the interests of both the rights holders and the users of copyrighted works.

Also, imagine the reverse scenario: if the judgment ruled that using copyrighted materials to train LLMs was infringement, it may have resulted in shutting down all indigenous generative AI development.

How does the Delhi High Court’s fair dealing exception compare with the fair use exception cited in two key US district court verdicts on LLM training and copyright infringement — involving Anthropic and Meta — last year?

One thing we must always remember is that different jurisdictions have very different approaches to protecting exclusive rights and crafting exceptions to those exclusive rights.

In the US context, courts have the “fair use” exception, a broad, open-ended exception based on a set of four factors to determine “fair use”: the purpose and character of use, nature of the copyrighted work, amount and substantiality of the portion taken, and the effect of the use upon the potential market. In the two US cases, both found that using copyrighted works for LLM training constitutes “fair use”, though there are nuances.

The framework in India is slightly different. We have the “fair dealing” exception, in which the analysis happens in two stages. First, the court looks at whether the use was for one of the specific purposes mentioned in the fair dealing exceptions. Second, the court engages in a “fairness analysis.”

The Delhi High Court has rejected the blind adoption of the US’ four-factor test in the fairness analysis because our fair dealing provision is much narrower. It would be terribly wrong to import those factors directly into the Indian analysis.

When ANI vs OpenAI was underway, many people doubted whether Indian copyright law could allow machine learning, given our relatively narrower “fair dealing” exception. However, what they overlooked is the ultimate purpose, i.e., learning, which is covered under the fair dealing exception The Delhi High Court has taken a liberal and dynamic approach to defining the terms ‘private use’ and ‘research’ in the fair dealing exception.I’m glad that it took this approach, because unless the first stage is cleared, we can’t move on to the next stage of fairness analysis.

Equally important, the Delhi High Court has laid down a unique fairness test. It asks us to look at three factors, including whether there will be market harm for the plaintiff to balance the interests of both the rights holders as well as users of copyrighted works. In this case, since ANI is a syndicating agency, this involved the question of whether ANI’s customers will substitute their services with LLM responses and the Court reached the conclusion that the answer is no.

Most importantly, the Delhi High Court also noted the public interest dimensions of LLMs. Overall, the court concluded that training activity can fall within the ambit of the fair dealing exception.

How has the Delhi High Court interpreted copyright infringement on the output side — the responses generated after being trained? What rights do copyright-owners exercise here?

ANI argued that ChatGPT users could reproduce their content verbatim, which would be copyright infringement. But the court observed that all the training happened before the publication of the articles which the plaintiff had cited as examples of infringement. The examples cited had dates after August 2024, whereas the relevant cut off months for training of the LLMs underlying ChatGPT were April 2022 for GPT 4 and April 2024 for GPT 4o.

The Court also looked at the question of whether the outputs were “substantially similar” to the copyrighted content.Most outputs in ANI’s case were not substantially similar because the machine generated different versions while responding to the user prompts. This means it didn’t meet the threshold for copyright infringement for the examples cited.

I must also add that news materials receive relatively lesser copyright protection because you cannot own facts, and there are only so many ways a fact can be presented. The net result is that the amount of copyrightable subject matter is much less as compared to other subject matters like novels or music.

This also means that while a finding of non-infringement was the outcome in this case, it doesn’t mean that tomorrow anyone can reproduce copyright-infringing outputs. In a different case with different facts, there might be a finding of infringement on the output side.

For example, if paragraphs are being reproduced verbatim, the publisher/copyright owner can still sue for output liability. As we move into specialised domains, and with increasing use of technologies such Retrieval Augmented Generation (RAG), LLMs may reproduce content in verbatim, making licensing the potential pathway to avoid infringement liability. This is why AI companies are entering licensing agreements with big publishers—to prevent future trouble.

OpenAI says it does not train on paywalled content or material blocked by web crawlers. Since ANI’s works were not behind a paywall, does bypassing such barriers to train LLMs amount to copyright infringement?

Definitely yes. If an LLM circumvents technological protection measures, then there might be liability coming their way.

The Delhi High Court has a pending suit regarding “shadow libraries” (like Sci-Hub) despite a preliminary observation that such libraries infringe on copyright. Since LLMs are often trained on these libraries, does this present a legal grey area in India?

In Elsevier vs Alexandra Elbakyan, I feel the court did not have the opportunity to hear the perspectives of the academic and research community. ANI vs OpenAI rightly clarifies that “lawful access” is not a general requirement under the fair dealing provision, or under most of the other exceptions to infringement listed under Indian copyright law. So if you can prove as a user that your use comes under the fair dealing exception or any of the other exceptions wherein lawful access is not mandated, the source shouldn’t really matter. Wherever “lawful access” is intended under our copyright statute, it is specifically mentioned.

Most jurisdictions don’t demand “lawful access” across the board because it would kill fair use and its broader purpose. The best example is Google LLC vs. Oracle America, Inc: Oracle claimed that Google had copied copyrighted application programming interface code from the Java programming language to train its Android operating system.

The US Supreme Court, in 2021, ruled in a 6–2 majority that Google’s use of the Java APIs was within fair use, even thoughGoogle copied the code without taking any permission. The same was also the case with respect to the famous Google Books decision, wherein you will notice that Google had scanned the books without taking permission, for the purpose of providing snippets to its consumers.