The Delhi High Court’s recent ruling in Asian News International (ANI) v. OpenAI has emerged as India’s first major judicial pronouncement on AI copyright, with legal experts split on whether the order signals a durable legal framework or merely a fact-specific, interim finding that could unravel on appeal.
The case, which saw ANI allege that OpenAI’s ChatGPT infringed its copyright by training on and reproducing its news content, has drawn commentary from general counsels, litigators, and IP specialists on four fronts including the scope of fair dealing, the commercial-use question, comparisons with the US fair-use regime, and the future of publisher business models.
Most legal experts agree the Court’s reliance on Section 52(1)(a)’s “private or personal use, including research” limb to shield AI training marks a significant expansion of India’s fair dealing doctrine.
Ankit Sahni, partner at Ajay Sahni Associates and counsel for intervenor Federation of Indian Publishers (FIP), said, “The Court has effectively created, through interpretation, a broad Text and Data Mining (TDM)-type exception which the statute does not expressly contain.” He argued legislative intervention is now necessary, noting the UK’s exception remains confined to non-commercial research and that India needs a “bespoke provision” rather than a stretched fair-dealing clause.
“We can think of training an Artificial Intelligence (AI) model like a student reading thousands of public news articles to learn grammar and factual concepts. However, existing laws were written for human students, experts can also argue that India needs explicit statutory rules for Text and Data Mining (TDM) so that technology developers have clear boundaries and publishers can opt out using machine-readable tags,” said Zameer Nathani, group general counsel at DNEG.
Deepank Singhal, advocate & IP attorney, argued that the judgment is “highly fact-specific” and should not be read as recognising a broad TDM exception, since the LLM was trained before ANI’s articles were published and the infringement claim centred on retrieval-augmented generation (RAG), not training itself. He opposed rushing into blanket TDM legislation without stakeholder consultation.
Commercial Use
The Delhi High Court’s move on commerciality was to treat OpenAI’s large‑scale training as capable of falling within “private or personal use, including research” under Section 52(1)(a)(i), while expressly rejecting the idea that revenue‑generating entities are automatically barred from invoking this defence. Justice Amit Bansal, in this case, noted that parliament has, in other parts of the Copyright Act, explicitly confined some exceptions to non‑commercial uses, but chose not to do so in Section 52(1)(a), which he read as a deliberate drafting choice. Legal experts debate on whether a global, revenue-generating platform can claim the “private or personal” defence.
“Section 52(1)(a)(i) was written for individual, non-commercial study, and calling a for-profit global platform’s industrial training “private or personal use, including research” strains the words heavily. An Indian court cannot paper over that with an American style transformative use analysis, because the exception is a fixed category, not a flexible standard. The saving grace is only that the finding is expressly prima facie and interim, decided on a low threshold, so it is provisional,” said Arjit Benjamin, associate partner, Prosoll Law.
Sahni called this the “most contested aspect” of the ruling, questioning whether “research” has been given “such an expansive meaning that the limiting words ‘private or personal use’ become largely redundant” when millions of copyrighted works train a subscription-based global product.
“Global commercial enterprises can use the “private research” defence under Section 52(1)(a) because internal model training is treated as an intermediate analytical step rather than directly selling the news stories. You will find that critics will argue that multi-billion dollar technology corporations are using free public data to build paid commercial software and shall stretch the originally intention of “personal research,” said Nathani.
India’s Fair Dealing vs US Fair Use
Legal experts contrasted India’s fair dealing exceptions with the US’s open-ended four-factor fair use test. US courts apply the open‑ended four‑factor fair use test under section 107, asking whether the use is transformative, how much of the work is copied, what kind of work is involved, and whether the use harms the relevant market.
In Bartz v. Anthropic, the court held that using lawfully acquired books to train the Claude model was “exceedingly” or “quintessentially” transformative, analogising training to human learning and finding no cognisable market harm where outputs did not reproduce the books.
Kadrey v. Meta accepted that training could be transformative but treated market effect as the decisive factor. Meta won summary judgment because the authors couldn’t prove Llama regurgitated their books or displaced any proven licensing market, though the court floated a “market dilution” theory that could matter in future cases.
Thomson Reuters v. Ross went the other way. A Delaware court rejected fair use where Ross copied Westlaw headnotes to build a directly competing legal research tool, finding the use non transformative and market substitutive, marking the first US ruling to deny fair use for AI training.
“New York Times v OpenAI remains undecided and is still in discovery, its central issue being “regurgitation”, namely whether ChatGPT reproduces Times material in substance. The Delhi Court addressed that same output question and held that the outputs generated through RAG were not substantially similar to ANI’s reporting,” said Benjamin. “The Delhi order aligns in outcome with the US rulings that favour developers, though it reaches that outcome by a narrower route.”
Sahni observed that OpenAI’s defence in the New York Times litigation rests on transformative purpose and lack of market substitution, whereas the Delhi High Court reached a “broadly pro-training result” through statutory interpretation of “research,” while still borrowing US-style concepts like transformative purpose and market substitution, making the Indian ruling “more categorical at the interim stage.”
Singhal emphasised the Delhi High Court “consciously declined to import the American doctrine of transformative use in its fullest sense,” concluding that “the destination may be similar, but the jurisprudential path is fundamentally different.”
Publishers Must Pivot to Licensing
Globally, leading publishers have already begun to treat their archives and live feeds as licensing assets in the AI economy. News Corp’s deal with OpenAI, announced in May 2024, is reportedly worth up to 250 million USD over five years. Axel Springer’s agreement with OpenAI is a multi-year partnership, while the Financial Times has struck a multi-year arrangement; the Associated Press signed an earlier, smaller licensing deal in 2023.
Nathani noted, “In terms of AI Content Licensing Deals (2026), no major India news publishers have secured direct and highly paid LLM licensing deals, whereas US publishers have signed commercial AI content licensing agreements exceeding USD 250 million combined.”
Sahni said the judgment “weakens the immediate ability of an individual publisher” to demand blanket licensing, since the Court found RAG outputs “not substantially similar” to ANI’s articles and saw no evidence of market substitution. But he argued licensing remains “commercially relevant,” with publishers gaining more leverage collectively, negotiating over archives, real-time feeds, attribution, and revenue-sharing rather than training access alone.
