From Libraries to Large Language Models: The New Meaning of "Research"

The training claim turned on Section 52(1)(a)(i), which permits fair dealing with a work “for the purpose of private or personal use, including research”. The issue, therefore, was whether OpenAI’s storage of ANI’s articles for model training fell within this exception. This issue raised three questions: a) whether a provision enacted in 1957 could apply to modern technology, b) whether machine learning can be treated as “research”, and c) whether a commercial developer can rely on a defence framed around “private or personal use”. Before addressing these questions, the Court considered the clause’s history. As enacted in 1957, it referred only to “research or private study”. In 1994, Parliament broadened it to “private use, including research”, describing the change as a correction to the “unduly narrow interpretation” given to “private study”. In 2012, the wording became “private or personal use, including research”, and an Explanation was added to clarify that storing a work electronically for those purposes does not amount to infringement. The Court treated this steady broadening as significant.

2.1 THE "ALWAYS SPEAKING" STATUTE: UPDATING CONSTRUCTION IN ACTION

The issue was whether a provision last amended in 2012 could apply to a technology that had not yet taken its present form. The Court answered this through the doctrine of updating construction. Relying on the Supreme Court’s decision in State v. S.J. Choudhary [2] and Francis Bennion’s [3]  formulation of an “always speaking” statute, the Court treated the provision as current law whose language must be interpreted in light of developments its framers could not have foreseen. In S.J. Choudhary, the Evidence Act’s reference to “science” was interpreted broadly to admit evidence from a typewriting expert, although that use was unknown when the provision was drafted. Applying the same approach to Section 52(1)(a)(i), the Court read “research” in the context of machine learning. 

2.2     CAN MACHINES CONDUCT "RESEARCH"? JUDICIAL RECOGNITION OF AI TRAINING

The next question was whether model training qualifies as “research”. Since the Act does not define the term, the Court adopted its ordinary meaning, to study or investigate carefully, consistent with the Madras High Court’s approach in Blackwood v. A.N. Parasuraman [4] . It then considered how training works in practice, and how literary works are screened, organised, converted into machine-readable inputs, and processed iteratively so the model can generate statistically probable outputs. On that basis, the Court held that training may be treated as research directed at producing new knowledge and advancing AI systems. It added a normative point: research is no longer exclusively human, but is increasingly conducted through AI, while still being undertaken at human direction and for human benefit. The Court reinforced this reasoning by analogy to the teaching exception in Section 52(1)(i), observing that it would be regressive to deny an AI system an exception a human teacher could claim if the system performed the same function. Foreign authority also supported this view. In CCH Canadian Ltd. v. Law Society of Upper Canada [5] , the Supreme Court of Canada gave “research” a broad and liberal meaning and refused to confine it to non-commercial activity, holding that lawyers conducting research for profit were still engaged in research

2.3     COMMERCIAL AI AND THE MEANING OF "PRIVATE OR PERSONAL USE"

That brought the Court to ANI's central objection, that a defence built around “private or personal use” cannot extend to a global enterprise training a commercial product. The Court rejected the view that commercial purpose is, by itself, fatal to the defence, relying on two textual points. First, Section 52(1)(a)(i) contains no express non-commercial limitation, even though the Act uses such limits elsewhere when Parliament intends them (for instance, Sections 52(1)(k)(ii), 52(1)(l) and 52(1)(o)). Second, the adjacent sub-clauses on criticism, review and reporting current events cover activities often carried out for profit, making it inconsistent to construe sub-clause(i) more narrowly than its neighbours. The Court also distinguished Rupendra Kashyap v. Jiwan Publishing House [6] , on which ANI relied, as a case where the defendant claimed the exception on the strength of its customers' use rather than its own. OpenAI's training, by contrast, was its own internal activity and, therefore, within the ambit of “private”. 

The Court held that the word “private” is not limited to individual human beings but also covers private entities, including companies. It gave two reasons: reading 'private' as a mere synonym for 'personal' would leave it redundant in the provision, and the Supreme Court [7] had already read 'private use' to cover an institutional actor. Since the training data was accessible only to the models and was never exposed to the public, the use was, in the relevant sense, interpreted to be ‘private’. Of the three questions, this is where the Court travelled furthest, for the extension of a private-use defense to an enterprise of OpenAI's scale could be a considerable step away from the individual user’s rights.