The Doctrine of Updating Construction

A possible response to our argument is that Chintamani Rao concerned an attempt to extend fair dealing to a new class of works, whereas ANI merely adopts a broader interpretation of the existing statutory purpose of “research” through the doctrine of updating construction. The Court’s application of the doctrine in paragraphs 216 to 219 proceeds as follows. It acknowledges that Parliament could not have contemplated the use of AI when section 52(1)(a) was last amended in 2012. It then invokes S.J. Choudhary for the proposition that statutes are “always speaking,” and reasons that research and learning are no longer confined to humans, since machine learning now undertakes these activities at human behest and for human benefit. On this basis, the Court concludes that LLM training constitutes research aimed at generating new knowledge and advancing artificial intelligence systems.

However, the Court engaged in an incorrect application of the doctrine here. In S.J. Choudhary itself, the Supreme Court used the doctrine of updating construction to resolve an ‘uncertainty’ about whether a new form of activity fell within an existing statutory term. An earlier decision in Hanumant had held that expert opinion on typewriting identification did not fall within “science” under section 45 of the Evidence Act, and the authorities were divided on the issue. The Supreme Court thus used the doctrine to hold that typewriting identification involved the application of scientific methods and therefore fell within the ordinary meaning of “science”.

In Taurus Earthmovers Pvt. Ltd., an assessee similarly invoked updating construction to argue that a JCB excavator should be classified as “agricultural machinery” rather than an “earth mover” for tax purposes, given its modern use in farming. The Court rejected the argument. Creative or updating interpretation, it held, “requires to be done, if there is some ambiguity or if the language of the statute is not clear, and not under all circumstances”. Where the Legislature itself categorised activities separately, with earth-moving machinery on one hand and agricultural machinery on the other, it was “not permissible” for a court to recategorise them through interpretation, however sympathetic the case for a more “progressive” reading. Thus, it is clear from precedents that the doctrine of updating construction operates only where genuine ambiguity exists. It cannot be used to cross a categorical line that the legislature has deliberately drawn.

Reading section 52(1)(a)(i) through this logic exposes the incorrect application of this doctrine by the Court in ANI. No ambiguity comparable to that in S.J. Choudhary existed here. Unlike “science,” whose application to typewriting analysis had been genuinely unsettled, there was no prior authority suggesting that “research” could be read independently of private or personal use. If anything, as discussed above, Chintamani Rao held that the provision was ‘clear’ on its face, while Jiwan Publishing House had already held that it protects the act of research or private study itself, rather than commercial activity merely connected to an eventual research purpose pursued by others. And, as in Taurus, Parliament here “itself categorised” the relevant purpose in a specific way. Consequently, what the Court did was extend the scope of the provision beyond what Parliament had enacted and thus effectively legislated on the matter itself.

A Fairness Test sans Quantum

Having concluded that AI training fell within the enumerated exceptions, the Court proceeded to decide whether OpenAI’s use was fair. As noted in Part I, it answered that question by formulating a three-factor test of its own. The inquiry focused on whether the use was confined to training, whether it caused economic harm to ANI, and whether it served the public interest (ANI, ¶ 236). One notable omission from this inquiry is the amount and substantiality of what was taken. This is despite the fact that the very authorities the Court relied upon in formulating its test treat that consideration as part of the fairness inquiry.

But why should the amount taken matter? It should because the greater the appropriation, the more likely it is that the use substitutes for the original work or diminishes its value. A student conducting research may scan and store selected pages from books relevant to a project he is working on. However, that is different from a student scanning and storing entire books for later use. Both may invoke the same purpose, whether research, criticism or review. Yet those two uses cannot be treated as equally fair because the amount taken directly affects the impact on the copyright owner’s work.

The authorities relied upon by the Court reflect this understanding. In Blackwood v. A.N. Parasuraman, the Madras High Court held that fairness depends on both the quantity and the value of what is taken (ANI, ¶ 230), and that where the appropriation is substantial, the copier’s motive becomes of limited significance. Likewise, in ESPN Star Sports v. Global Broadcast News, the Delhi High Court treated the amount and importance of what was taken as part of the fairness inquiry (ANI, ¶ 231). Yet neither consideration finds a place in the three-factor test the Court ultimately adopts.

Perhaps what is more glaring is that Civic Chandran finds no mention in the ANI judgment, even though it remains the leading Indian authority on the role of quantum in the fair dealing inquiry. The case concerned a “counter-drama” written in response to Thoppil Bhasi’s celebrated play Ningal Enne Communistakki. The defendants, in doing so, reproduced substantial portions of the original, including its characters and dialogue, as part of a critique of its political message.

The Kerala High Court began from the premise that fair dealing does not ordinarily permit the reproduction of an entire work, or a substantial part of it “as such.” It accepted that the defendants had copied a substantial part of the original, but that did not end the inquiry. Instead, it weighed the quantum of copying against the purpose of the use and its competitive effect before concluding that the copying was necessary for the criticism, was not a substitute for the original, and did not amount to an appropriation of the author’s labour. The significance of Civic Chandran is that the amount taken was the starting point of the fairness inquiry, and substantial copying was excused only after the remaining considerations justified it.

It is this inquiry that AI training would have had difficulty in satisfying. Unlike criticism, AI training does not reproduce extracts limited to what a particular purpose requires. It reproduces each work in its entirety. A faithful application of Civic Chandran would therefore have required the Court to confront the wholesale nature of the copying and to explain what, if anything, displaced the presumption against it. The judgment does neither. Tellingly, even the SpicyIP commentary sympathetic to the outcome has observed that the Court’s factors amount to little more than a combination of the first and fourth United States fair use factors together with a public interest inquiry.

It might be objected that this is unrealistic given the technology. Large language models require vast amounts of data, and there may be no practical way to train them without copying works on a very large scale. But if the amount of copying is part of the fairness inquiry, this may simply mean that AI training does not fit comfortably within the existing exception. The answer cannot be to disregard quantum because the technology makes extensive copying necessary. That would be to expand the exception to accommodate a use that Parliament has not expressly provided for. Nor does this leave AI companies with no lawful means of obtaining training material. The Court itself recorded that ANI had offered OpenAI a licence for its content (ANI, ¶258.2), and that OpenAI already licenses material from the Financial Times, the Associated Press, and Condé Nast (ANI, ¶57.5). Licensing and paying for the material used remains a possible route even if the existing exception does not accommodate the scale of copying required for AI training.

A major part of the Court’s fairness analysis also relies on the doctrine of transformative use. This doctrine allows the use of an original work if it is for a substantially different purpose or character than the original. The Court derives it from Narendra Publishing House and B.D. Bhandari. Both decisions, however, arose under other limbs of section 52, and not under the “private or personal use, including research” limb in section 52(1)(a)(i) with which the ANI Court was concerned. In Narendra Publishing House, the Single Bench applied the transformative test and allowed the defence under section 52(1)(a)(ii), the “criticism or review” limb. On appeal in B.D. Bhandari, the Division Bench adopted the “derivative work” test primarily under section 52(1)(h), the course-of-instruction limb, and extended it to section 52(1)(a)(ii) later in its reasoning. On no view did either case concern section 52(1)(a)(i). The ANI Court itself acknowledged as much, recording the amicus’s submission (ANI, ¶ 226) that the transformative test had been applied under “review” in section 52(1)(a)(ii) and “not in the context of Section 52(1)(a)(i).” It nonetheless disposed of the difficulty in a single sentence, holding that “there is no reason why” the same principles could not apply under section 52(1)(a)(i) (ANI, ¶ 227).

Some commentators have in fact argued that Indian courts should read the fair dealing exceptions progressively, such that scraping and training fall within the “research” limb of section 52(1)(a)(i). Nonetheless, there appears to be a reason to resist that move, at least where it proceeds by transplanting the transformative use doctrine from another ground in the statute. Criticism and review are inherently referential. To criticise a work, one must reproduce enough of it to show the reader what is being criticised. For example, a critic may reproduce a passage from a novel to explain why its treatment of a particular theme is flawed. The copied passage is necessary because it is the subject of the criticism. In that setting, asking whether the material was ultimately for a ‘different purpose or character’ helps assess whether that copying is justified. Research operates on a different logic.  A researcher may reproduce a study because its findings are useful to the research being conducted, without the study itself being the subject of analysis or criticism. In such a case, the fact that the research is for a different purpose and ultimately produces something different from the study says little about whether the copying was justified. Indeed, if a difference between the original and the eventual output were enough to make the use transformative, almost any research use could be characterised as transformative. The rationale for applying transformative use in criticism or review therefore does not readily carry over to the research limb. The doctrine thus cannot simply be transplanted without the court explaining why exactly the distinction drawn by Parliament between the two statutory grounds should make no difference.

There is also an inconsistency in how the Court treats this line of authority. When ANI relied on B.D. Bhandari to argue that commercial exploitation defeats the defence of fair dealing, the Court declined to apply it and reasoned that the case had been decided under a different limb of section 52 whose scope was “entirely different” from the provision before it (ANI, ¶ 183). Yet, when discussing transformative use, the Court took the opposite approach and held that the doctrine could be carried across into section 52(1)(a)(i) without difficulty.

Moreover, the Court records the intervenor’s reliance on Andy Warhol Foundation v. Goldsmith for the proposition that a secondary work does not become transformative merely because it adds new expression, and that where the secondary work shares the purpose of the original, the claim to transformativeness is correspondingly weaker (ANI, ¶ 139.5). That qualification bears directly on AI training, whose outputs may serve the same informational purpose as the news reports on which they are trained. Yet when the Court later embraces transformativeness as part of its reasoning, it never engages with that qualification.

Conclusion

There is a recurring pattern in the Delhi High Court’s reasoning. On the purpose inquiry, it had Jiwan Publishing before it but read past what that decision actually held. On whether section 52 could be expanded at all, it had Chintamani Rao before it, yet relied on that decision only for propositions that did not bear on its main issue. On fairness, it formulated a test that omitted a consideration its own authorities consistently treated as part of the inquiry, and imported the doctrine of transformative use from a different statutory limb without answering the objection it had itself recorded. It is evident that the authorities relied upon by the Court were read closely enough to be cited, but not for the propositions that would have constrained the result.

None of this was necessary to decide the interim application. ANI’s claim based on the outputs had already failed on narrower grounds. The training cut-off ruled out memorisation, and the Court found no substantial reproduction even under adversarial prompting (ANI, ¶ 119). The ordinary considerations governing interim relief were also sufficient (ANI, ¶¶ 265, 269). ANI’s own willingness to license its content (ANI, ¶ 265) bore on the adequacy of damages, while an injunction would have disrupted an already functioning product. Even so, the Court proceeded to articulate a broader view of section 52 on an interim record, before the suit had gone to trial.

If AI training is to be accommodated within Indian copyright law, that choice belongs to Parliament. Comparative experience points in the same direction. Where jurisdictions with closed, enumerated copyright exceptions have chosen to accommodate AI training, they have done so through legislation. Japan enacted a broad information analysis exception in Article 30-4 of its Copyright Act, in force since 2019, which extends even to commercial use. Singapore introduced a “computational data analysis” exception in section 244 of its Copyright Act 2021, likewise permitting commercial use where the user has lawful access. The European Union created its text and data mining exceptions through Articles 3 and 4 of the 2019 Digital Single Market Directive. Though it may be argued that the United States reached a similar result through judicial development, the comparison is unpersuasive. The U.S. Code’s open-ended fair use doctrine under section 107 is framed as a general standard, and allows courts to assess new technologies on a case-by-case basis. India’s fair dealing regime differs as it is built around a closed and enumerated list of exceptions, and as a result leaves less room for judicial expansion.

Where a new technology sits uneasily within a statute enacted for a different era, the judicial task is to identify the limits of that statute rather than to redraw them. It is this task, we argue, that the Court failed to perform. The Court’s repeated clarification that its conclusions are only prima facie is its only saving grace. It leaves the question open for fuller consideration at the final hearing, where these issues deserve a more complete examination before any lasting interpretation of section 52 is settled.