Why the case landed in court

ANI sued after noticing that ChatGPT’s answers to queries about recent Indian news resembled its own articles published in August and September 2024. The agency argued the large language model was reproducing its copyrighted text, a claim that, if upheld, would force OpenAI to halt or heavily restrict its models in India.

OpenAI replied that the two models cited—GPT-4 and the newer GPT-4o—were trained on data with cut-off dates of April 2022 and April 2024. ANI’s articles appeared after those dates, so they could not have been in the original training set. Judge Amit Bansal said the overlap most likely came from Retrieval-Augmented Generation (RAG), which pulls current web content into the response in real time, not from the model “remembering” the articles.

The case matters because the court read India’s copyright exception for “private or personal use, including research” broadly. Both sides agreed OpenAI used ANI material during the initial training phase, but the judge treated that use as “transformative.” In other words, the model extracted only grammar, syntax and statistical patterns, leaving the expressive content untouched.

The court set two strict conditions:

  • Copies used for training must come from lawful sources, not pirated archives or paywalls the user cannot access.
  • The material must stay inside the developer’s own environment; it cannot be published or distributed.

Applying a three-part fairness test, the judge found OpenAI’s use limited to training, saw no evidence the model memorized verbatim passages, and noted ANI suffered no direct economic loss because the two firms operate in different market segments.

How this fits into a patchwork of global rulings

India’s stance now mirrors several U.S. cases that favor a transformative view of AI outputs. In Kadrey v. Meta, a court held that generated text that does not copy exact wording is non-infringing, while the long-standing Google Books decision recognized a limited “search-and-display” exception for digitisation projects. Other U.S. suits, such as the Raw Story case, were dismissed for lack of demonstrable harm, but the Anthropic controversy reminded judges that using pirated data can still trigger liability.

Across Europe, opinions diverge sharply. Courts in Munich have ruled that reproducing lyrics embedded in model weights amounts to infringement, whereas a London tribunal recently dismissed a claim against Stability AI on similar grounds. The Delhi High Court’s ruling adds another data point: if training is confined to lawful sources and the output is not a verbatim copy, the activity may fall under a research exception.

What developers should watch

  • RAG vs. training – The court’s distinction between “memorization” (training) and real-time retrieval (RAG) suggests future disputes will focus on whether an answer comes from a static knowledge base or is fetched live from the web. Developers may need to make that line clearer in product documentation.
  • Source provenance – The “lawful sources” requirement could push firms to tighten data-curation pipelines, using contracts or licenses that prove each text slice is legitimate.
  • Internal-only use – Companies planning to commercialise model outputs must keep training data strictly internal. Sharing raw corpora with partners or the public could erode the research defense.
  • Economic-harm test – Because the court emphasized the absence of direct financial injury, claimants will need to show concrete loss, not just perceived brand dilution.
  • Legislative response – Indian lawmakers are monitoring AI-related copyright debates. Any amendment that narrows the research exception could retroactively affect models already deployed, prompting a wave of compliance audits.

The counter-argument that still haunts AI

ANI’s complaint highlighted a genuine concern: as LLMs become more adept at echoing specific phrasing, the line between “transformative” use and “copy” can blur. Critics argue that RAG, while technically a search function, still surfaces copyrighted content without permission, potentially violating the “communication to the public” right.

A precedent with worldwide ripple effects

En classant l'entraînement des modèles comme une activité de recherche autorisée, la Haute Cour de Delhi a signalé que l'Inde ne bloquera pas automatiquement le développement de grands modèles de langage pour des motifs de droit d'auteur, à condition que les développeurs respectent la légalité des sources et maintiennent l'entraînement en interne. Cette interprétation pourrait encourager les entreprises à étendre leurs activités en Inde, sachant qu'un obstacle juridique majeur s'est assoupli.

Pour les régulateurs d'ailleurs, cette décision offre un modèle : définir une exception de recherche étroite et bien documentée, imposer des normes de sourçage claires et exiger que le traitement des données d'entraînement reste strictement interne. Les nations qui adopteraient un langage similaire pourraient offrir à leurs écosystèmes d'IA nationaux la certitude nécessaire pour attirer les investissements.

L'essentiel : La décision de la Haute Cour de Delhi ne donne pas carte blanche pour le moissonnage de n'importe quel texte à des fins d'entraînement de l'IA, mais elle établit qu'en vertu du droit indien, un entraînement discipliné et conforme à la loi est qualifié de recherche. Cette interprétation pourrait devenir un point de référence pour les tribunaux du monde entier alors qu'ils s'interrogent sur la question de savoir si l'apprentissage du langage par les machines est une activité académique protégée ou une infraction en puissance.