A Narrow Win, Not a Verdict
On July 24, 2026, Justice Amit Bansal of the Delhi High Court declined Asian News International's request for an interim injunction against OpenAI, holding that training ChatGPT on ANI's news content is prima facie fair dealing under Section 52(1)(a) of the Copyright Act, 1957. The court also found that ChatGPT's Retrieval-Augmented Generation (RAG) outputs were not substantially similar to ANI's original reporting, and that ANI had "failed to satisfy this Court that any memorisation or regurgitation" of its copyrighted work had occurred (Bar and Bench).
This is an interim order in a case that is far from over. ANI Media Pvt Ltd v OpenAI Inc & Anr, filed in November 2024 as India's first major generative-AI copyright suit, continues on the merits (The Star/Reuters; Indian Kanoon, CS(COMM) 1028/2024). The Federation of Indian Publishers and the Digital News Publishers Association have since intervened, extending the dispute from one wire service's grievance into an industry-wide test of whether Indian copyright law lets AI firms train on news and books without a licence (Tech Policy Press).
The Steelman: Why Publishers Want a Licence, Not a Carve-Out
ANI's case deserves to be stated on its own terms before it's rebutted. Indian newsrooms operate on thin margins, and wire services in particular sell the thing OpenAI ingested for free: freshly verified, exclusively reported news. If a chatbot can answer a user's question using the substance of a paywalled or licensed report without ANI seeing a rupee of it, ANI's argument runs, the commercial incentive to keep reporters on the ground quietly erodes even where no single output is a verbatim copy. That is a real market-substitution concern, not a hypothetical one — and it is the same concern that has driven licensing deals between AI labs and outlets like the Associated Press and Axel Springer elsewhere in the world. A blanket "it's just training, it's fair use" answer doesn't fully engage with that.
Why the Court's Reading Is Still the Better Default
Section 52(1)(a) protects fair dealing for "private or personal use, including research." Justice Bansal's order extends that language to machine learning without holding that every model output escapes scrutiny — the court separately tested the ChatGPT outputs for substantial similarity and found none, which is the correct two-step analysis: training is not infringement per se, but an output that reproduces protected expression still can be. That distinction matters enormously for how this precedent should be read. It is not a blanket immunity for AI firms; it is a finding that ingestion for statistical training, standing alone, is not the same act as republication.
Getting this distinction right matters because the alternative — treating training itself as an infringing act requiring per-work clearance — would make foundation-model development in India legally impossible for any firm without the balance sheet to negotiate hundreds of individual licences, entrenching the handful of companies (mostly American) that can already afford U.S. and EU licensing deals. A narrow, output-focused infringement standard, paired with real consequences when a model does regurgitate protected text, disciplines the actual harm without pricing Indian and smaller global AI developers out of the market entirely.
The Real Fight Has Moved to DPIIT
What the ruling cannot do is settle India's copyright-AI policy question, because that question has already migrated from the judiciary to the executive. In December 2025, the Department for Promotion of Industry and Internal Trade published a working paper proposing a mandatory blanket-licence system: AI firms would get automatic access to lawfully available copyrighted works but would pay royalties into a central collecting body for distribution to rights-holders. The proposal drew a 30-day public comment window that closed in early February 2026 (TechCrunch). Nasscom, representing Google, Microsoft and other tech firms, formally dissented, arguing for a text-and-data-mining exception instead of mandatory royalties.
That is the fork that actually decides the shape of India's AI-copyright regime — not Justice Bansal's interim order. A judge applying 1957-vintage statutory language to a technology Parliament never contemplated can only go so far; DPIIT's working paper is explicitly framed as filling that gap through amendment, not litigation. If the government adopts the mandatory-licensing model, it would sit awkwardly beside a judicial finding that the same conduct is already lawful fair dealing — Parliament, not the bench, would need to resolve that tension.
What to Watch
Three threads now run in parallel: whether ANI's suit survives to a final judgment that might narrow or reverse today's interim finding; whether the Federation of Indian Publishers' broader claims over book content fare differently from ANI's news-wire claims; and whether DPIIT's committee, after reviewing the Nasscom dissent, finalises the licensing model or steps back toward a TDM exception closer to what the court has now effectively endorsed. Until the DPIIT process concludes, India's AI-copyright rules remain genuinely unsettled — this ruling is a data point in that process, not its answer.