On 24 July 2026, the Delhi High Court refused Asian News International's (ANI) application for an interim injunction against OpenAI OpCo, LLC (OpenAI), holding that storing ANI's news content to train the models behind ChatGPT falls, prima facie, within the fair dealing exception under Section 52(1)(a) of the Copyright Act, 1957 (Act). 

The significance of the ruling lies not in the fact that AI training may proceed without a licence in some cases but more in the Court's approach. While accepting that storing copyrighted works for training involves reproduction under Section 14(a)(i) (which expressly includes "the storing of it in any medium by electronic means"), the Court treated Section 52 not as a narrow carve-out from the copyright owner's monopoly, but as an independent statement of permitted uses to be construed liberally. That reframing gives the judgment significance beyond AI training.

An exception stops being an exception

The core of the ruling lies in the Court's reading of Section 52. Relying on The Chancellor, Masters and Scholars of the University of Oxford v. Rameshwari Photocopy Services and Canadian fair-dealing jurisprudence, it held that Section 52 independently defines permitted uses and should be read broadly. Legislative history supporting that approach: the 1994 amendment avoided an unduly restrictive reading of “private study”, while the 2012 amendment extended fair dealing to “any work”, other than computer programmes, and expressly contemplated electronic storage. The judgment is therefore less an AI-specific exception than an application of flexibility already built into Section 52.

Getting "Research" to Cover a Training Run

The key aspect of the judgment is its treatment of a language model’s ingestion of text as “research” under Section 52(1)(a)(i). 3 moves make that possible:

a.                       Commerciality is not an automatic bar: ANI argued that a profit-making chatbot could not invoke ‘private or personal use’ under Section 52(1)(a). The Court disagreed, noting that Parliament used the qualifier ‘non-commercial’ elsewhere in Section 52 but omitted it from Section 52(1)(a).[1] That omission, in the Court’s view, had to be given meaning. Commercial purpose may still weigh against fair dealing, but does not, by itself, exclude Section 52(1)(a).

b.                       “private use” is not confined to individuals: The Court declined to equate ‘private use’ with use by a natural person. OpenAI’s training corpus was maintained in a closed computational environment and not available for public access, browsing or download. On that basis, a corporate entity could, therefore, operate within a private or closed sphere.

c.                       “research” gets an updating construction: The Court held that research need not be undertaken directly by a human being. Machine learning carried out at human direction and for human benefit could fall within the statutory concept of research. By analogy with Section 52’s educational exceptions, protected activity should not lose its character merely because technology performs a function previously undertaken by a human.

Fairness, Market Harm, and the ‘Public Interest’ Problem

The Court assessed fairness by asking whether the use was confined to training, whether it competed with or economically prejudiced ANI, and whether it served a broader public interest. These factors favored OpenAI: no memorization was shown, ChatGPT was not found to substitute for ANI’s news service, and ANI produced no concrete evidence of commercial harm.

The public-interest limb is more contentious. Section 52 sets out specified permitted statutory purposes, not an open-ended fairness inquiry. The tension lies in using broad ideas of transformation and public benefit to expand “research”, despite the provision being a closed list of permitted purposes.

The market-harm analysis may also prove too narrow if future claims concern lost licensing opportunities, attribution, or the ability to develop competing products.

The judgment does not alter the settled position that copyrighted material does not enter the public domain because it is publicly accessible online. Copyright continues to subsist; the question is whether a particular use falls within a statutory limitation or exception. The ruling is better understood as recognising a form of ‘limited functional commons’ for certain computational uses: the work remains protected, but internal machine analysis or training may proceed without a licence where Section 52 is satisfied. That reasoning could affect research, archiving and search technologies beyond generative AI. The significance is not that Section 52 has become a source of user rights, but that the Court has extended the practical reach of those rights to a technologically novel form of large-scale computational processing.

What The Judgment Doesn't Decide

The judgment is significant, but its scope will depend on how courts apply it in subsequent litigation. Its findings are prima facie, and the dispute concerned news content, where copyright protection is comparatively thin and the underlying material was available on ANI’s public website. Different questions may arise for expression-heavy works such as music, photographs, films or books, or material obtained from pirated datasets, shadow libraries or by circumventing access controls. Those scenarios were not before the Court.

The reasoning may also be tested where the evidence differs: where a model reproduces substantial protected expression, a licensing market for AI training shows measurable displacement, or a retrieval-based product substitutes for the copyright owner’s service. ANI therefore establishes that some AI training can fall within Section 52; it does not settle the legality of AI training as a category.

There is also a policy tension. DPIIT’s working paper from December 2025 contemplated a mandatory licensing and remuneration regime for AI training, deeming that such use engages copyright owners’ exclusive rights. The Court has now held that at least some training may fall within Section 52 without consent or payment. If Parliament creates a compulsory remuneration right over uses courts regard as fair dealing, it would redraw rather than clarify the copyright bargain.

Beyond the Act of Copying

ANI v. OpenAI does not make the internet free training data, nor does it put copyrighted works in the public domain. Its significance is subtler: it treats copyright’s limits as rights in themselves, gives “research” enough elasticity to include machine learning, and shifts the dispute from whether copying occurred to what was copied, how it was obtained, what the model can reproduce, and what market harm actually follows.

For AI developers, that is meaningful legal ground, at least for now. For rights holders, it changes the terrain of the next fight: access, paywalls, licensing markets, retrieval systems and the point at which technological restrictions begin to shape statutory rights.

Whether the Court’s approach survives trial, appeal and Parliament remains open. After ANI v. OpenAI, the harder question is no longer simply whether machines copy; it is which acts of copying copyright law was meant to control.


[1] See Act: Section 52(1)(ad) using the phrase "non-commercial personal use"; Section 52(1)(k)(ii) restricting protection to clubs "not established or conducted for profit"; Section 52(1)(l) covering performances by amateur clubs "if the performance is given to a non-paying audience, or for the benefit of a religious institution"; Section 52(1)(n) limiting storage to "a non-commercial public library"; and Section 52(1)(o) covering "not more than three copies of a book" made by a "non-commercial public library".


The authors are Lawyers  at Bharucha & Partners. Views expressed are personal.