Fair dealing cannot be presumed: Why AI training fails the purpose test under Indian copyright law

As generative artificial intelligence (AI) systems become increasingly embedded in creative and commercial ecosystems, a growing claim has emerged: that training AI models on copyrighted works qualifies as “fair dealing”.

This argument, however, rests on a fundamental misreading of Indian copyright law and is often borrowed uncritically from US-centric fair use discourse.

Under the Copyright Act, 1957, fair dealing is not a flexible or open-ended defence. It is a narrow, purpose-specific statutory exception. Before any court can examine whether a use is “fair”, it must first determine whether the use falls within one of the expressly enumerated purposes under Section 52. If that threshold is not met, the enquiry ends.

When this statutory structure is applied faithfully, AI training struggles to clear the initial purpose test itself. As a result, the fair dealing defence may be unavailable at the outset.

Unlike the United States’ open-textured fair use doctrine, India follows a closed-list model of copyright exceptions. Section 52(1)(a) permits fair dealing only for the following purposes:

• Criticism or review;

• Private or personal use, including research; and

• Reporting of current affairs.

Indian courts have repeatedly emphasised that these categories are exhaustive. In Super Cassettes Industries Ltd v. Hamar Television Network Pvt Ltd (2011), the Delhi High Court made it clear that fair dealing requires both fairness in the manner of use and conformity with one of the listed statutory purposes. The latter operates as a threshold condition. This sequencing is crucial. The question of “fairness” arises only after the court is satisfied that the impugned use fits within a permitted purpose. Any argument that AI training qualifies as fair dealing must, therefore, begin and succeed at this first gate.

Training a generative AI model requires the systematic copying of vast quantities of protected material - including text, images, music and source code - into training datasets. This copying is neither incidental nor accidental; it is a deliberate and technically necessary step in building commercially valuable systems.

Importantly, the works are not engaged with for their expressive or communicative content in the human sense. They are converted into numerical representations to enable algorithmic pattern recognition. While the end product may appear innovative, the training process itself depends on extensive copying.

Under Section 14(a)(i) of the Copyright Act, the right to reproduce a work “in any material form” is exclusively vested in the copyright owner. The legal question is not whether AI development is socially useful, but whether this reproduction falls within any statutory exception. Absent such an exception, unauthorised copying remains prima facie infringing.

Proponents of fair dealing often rely on the “research” limb of Section 52(1)(a). This reliance is doctrinally problematic. In Indian copyright jurisprudence, research has traditionally been understood as a human-centred intellectual activity aimed at studying or understanding the copyrighted work itself - its content, themes, arguments or underlying ideas. The work is the object of inquiry.

AI training reverses this relationship. The copyrighted work is not studied for its meaning; it functions as a statistical input used to optimise model parameters. The objective is not to understand the work, but to improve a prediction engine capable of generating new outputs.

Interpreting “research” broadly enough to encompass large scale, automated data ingestion for commercial product development would effectively collapse the purpose based limits of Section 52. It would amount to judicially creating a text and data mining (TDM) exception - a step Parliament has not taken.

This legislative silence is significant. Jurisdictions such as the European Union and Japan have enacted explicit TDM provisions after detailed policy deliberation. Indian law contains no comparable exception and courts are ill-equipped to supply one through interpretation alone.

The remaining fair dealing purposes criticism, review and reporting of current affairs are intrinsically expressive. They involve engaging with a work’s meaning to generate commentary, opinion or information for public consumption. These purposes align closely with the free speech values protected under Article 19(1)(a) of the Constitution.

The copying that occurs during AI training, by contrast, is non-expressive. It does not comment on, critique, or report anything to a human audience. Expression is consumed as raw material to build a functional tool.

Some argue that AI training should be protected because it is “transformative.” This argument conflates distinct legal concepts. In Eastern Book Company v. DB Modak (2008), the Supreme Court discussed transformation in the context of originality for copyright subsistence, not fair dealing. Indian courts have not adopted the US.doctrine of transformative use, articulated in Campbell v. Acuff-Rose Music, Inc (1994), as an independent determinant of fairness.

While Indian courts have occasionally referred to transformation when assessing fairness, such references operate only after the statutory purpose test is satisfied. Technical transformation cannot substitute compliance with a listed purpose in India’s purpose-driven framework.

This structural point is often overlooked. If a court concludes that a use does not fall within Section 52(1)(a), it need not and should not examine questions of fairness at all. Considerations such as the amount copied, commercial intent or market harm become legally irrelevant if the threshold purpose requirement is unmet. Arguments that leap directly to balancing factors import a US analytical model that sits uneasily with Indian statutory design.

If AI training fails at the purpose stage, the fair dealing defence collapses before any further analysis begins.

The uncertainty surrounding AI training and copyright has come into sharper focus in recent years.

In November 2024, Asian News International initiated proceedings against OpenAI before the Delhi High Court, challenging the unauthorised use of copyrighted content for AI training and contesting the availability of fair dealing as a defence. While the case remains pending, it may offer the first judicial engagement with these questions in India.

At the policy level, the government has initiated expert consultations to examine whether the Copyright Act adequately addresses AI-related uses, including the possible need for a tailored TDM exception. These discussions implicitly acknowledge that existing provisions may not comfortably accommodate industrial-scale machine learning.

Government statements have generally maintained that current law is sufficient, while also emphasising that authorisation is required where fair dealing does not apply. Read together, this suggests that fair dealing cannot be assumed to cover AI training by default.

A careful reading of the Copyright Act demonstrates that AI training does not easily fit within India’s fair dealing framework. Claims to the contrary conflate India’s purpose-specific regime with the far more flexible US fair use doctrine. Copyright exceptions designed for human creative engagement are poorly equipped to regulate industrial-scale machine learning.

If India wishes to promote AI innovation while safeguarding creators’ rights, the solution lies in legislative action whether through licensing mechanisms, opt-out systems or a carefully calibrated TDM exception with appropriate safeguards. Stretching existing provisions beyond their textual and doctrinal limits risks undermining both legal certainty and democratic law-making.

Shubhi Priyadarshi is a recent B.A. LL.B. graduate with a keen interest in research and writing on contemporary legal issues.