Fair Dealing for Machines: The Delhi High Court's First Reading of AI Training Under Copyright Law
The court declined to restrain OpenAI, holding that training on ANI's content fell within fair dealing — an outcome opposite to Anthropic's costly US settlement days earlier
What happened
An aspirant who read the earlier piece on the Anthropic settlement should read this as its mirror image. The same question — may an AI model be trained on copyrighted material without permission — produced a large payout in one jurisdiction and a refusal of interim relief in another, and the divergence is not arbitrary: it traces to the different structure of copyright exceptions in Indian and American law, which is the examinable core.
Same question, two systems: AI training on copyrighted works
| Feature | India (OpenAI-ANI) | United States (Anthropic) |
|---|---|---|
| Copyright exception | Fair dealing — closed list (Sec 52) | Fair use — open four-factor test |
| Decisive concept | Enumerated purpose (research/private use) | Whether use is transformative |
| Outcome | Interim relief refused | ~$1.5 billion settlement |
| Stage | Interim — suit continues | Settlement — case resolved |
Source: Delhi High Court order, July 2026; reported terms of the Anthropic US settlement
The provision at the centre is Section 52 of the Copyright Act, 1957, which lists acts that do not constitute infringement — India's fair dealing regime.
●Unlike the open-ended fair use doctrine of United States law, which lets a court weigh four factors for any use, Indian fair dealing is a closed list of enumerated purposes, and Section 52(1)(a) covers fair dealing for private or personal use, including research, and for criticism or review.
●The court's reasoning had two limbs: that training on the works fell within this exception, and that ANI had not shown a prima facie case that ChatGPT reproduced or retrieved its original literary works in its outputs — the reproduction question being distinct from the training question.
●The order is at the interim stage, so it refuses an injunction rather than deciding the suit finally.
●The contrast with the United States matters because American fair use is more flexible and its outcomes turn heavily on whether the use is transformative, whereas Indian fair dealing is bounded by its enumerated purposes.
Indian fair dealing is a closed list of permitted purposes, not the open four-factor fair use test of US law — which is why the same AI-training question can resolve differently across the two systems.
◎ In Simple Words
A news agency accused the makers of ChatGPT of using its news articles, without permission, to teach the AI system. The Delhi High Court has said, for now, that this training probably does not break copyright law, because Indian law allows certain uses of copyrighted work, and because the news agency could not show that ChatGPT was actually reproducing its articles in answers. Interestingly, in the United States a similar dispute was recently settled for a very large sum — so the same question got very different treatment in the two countries.
Factual Pointers
Practice · 2 questions
Which one of the following correctly describes the copyright exception the Delhi High Court applied in the OpenAI-ANI matter?
The Delhi High Court's reasoning distinguished two questions. Which one of the following pairs correctly states them?
Mains Practice Questions
The Delhi High Court's order in the OpenAI-ANI matter distinguished training on copyrighted inputs from reproduction in outputs. Examine the significance of this distinction for the governance of generative AI.
India's fair dealing and the United States' fair use produced opposite outcomes on the same question of AI training. Discuss how the structure of copyright exceptions shapes such divergence.
Copyright law was written for human copying, not machine learning. Examine whether the courts or Parliament should settle the rules for training AI on copyrighted works in India.
MCQ Practice
3 questions on this article
With trap analysis, approach guide, and UPSC angle
Frequently Asked
· People also askWhat did the Delhi High Court decide in the OpenAI-ANI case?
It declined to restrain OpenAI, holding that at the interim stage, using ANI's content to train ChatGPT does not amount to copyright infringement. Justice Amit Bansal held that the training fell within fair dealing under Section 52(1)(a) of the Copyright Act, 1957, and that ANI had not shown ChatGPT reproduced its original works.
GS2 · Judiciary · IPRIt is the first detailed attempt by an Indian court to place AI training within the Copyright Act. Because the order is interim, it refuses an injunction rather than deciding the suit finally.
SOURCE Business Standard · Delhi High Court order, July 2026
What is the difference between fair dealing and fair use?
Fair dealing, in Indian and UK law, is a closed list of enumerated permitted purposes — such as research, private use, criticism and review under Section 52. Fair use, in United States law, is an open-ended doctrine under which a court weighs four factors for any use, with transformative use often decisive.
GS2 · Intellectual propertyThis structural difference is why the same AI-training question produced a refusal of relief in India and a large settlement in the United States: a closed list and an open standard can diverge on identical facts.
SOURCE Copyright Act, 1957, Section 52
Why did the Indian and US outcomes differ on the same question?
Because of different legal architecture, not inconsistency. American fair use is an open standard where transformative use can justify almost any use; Indian fair dealing is a closed list where a use must fit an enumerated purpose. A closed list and an open standard can reach opposite results on identical facts.
GS2 · Comparative lawThe Anthropic settlement resolved a US class action through payment, while the Delhi High Court refused interim relief under Section 52(1)(a) — two systems, one question, two outcomes.
SOURCE Delhi High Court order; reported Anthropic settlement
Does the order mean AI training on copyrighted works is legal in India?
Not finally. The order is at the interim stage and refuses to restrain OpenAI while the suit is heard; it is a provisional assessment that ANI had not made out a strong enough case for an injunction. The final determination in the suit could differ.
GS2 · Judicial processTreating an interim refusal of relief as a settled ruling that AI training is lawful overstates the order. The reasoning is influential but not the last word.
SOURCE Delhi High Court order, July 2026
What was ANI's concern beyond copying?
ANI argued that ChatGPT could falsely attribute statements to it, harming its reputation and contributing to misinformation. This is a harm to the integrity of information rather than a classic reproduction of protected expression, which points to the limits of copyright as the instrument for governing generative AI.
GS3 · AI governanceSome genuine harms of generative AI — false attribution, dilution of the value of original journalism — sit outside the reproduction-centred logic of copyright, suggesting the need for governance instruments beyond it.
SOURCE Reported pleadings, ANI Media v. OpenAI
Should courts or Parliament settle the rules for AI training?
The order shows courts adapting a 1957 statute to a technology its drafters could not foresee, which gives case-by-case guidance but leaves uncertainty until precedent forms. A legislative solution — a purpose-built exception or a text-and-data-mining provision — would offer clarity but requires Parliament to legislate ahead of a fast-moving technology.
GS2 · Governance · GS3 · Technology policyWhich institution should settle the rule is itself a governance choice, and jurisdictions such as the European Union have opted for an explicit statutory text-and-data-mining exception rather than leaving it to the courts.
SOURCE Copyright Act, 1957; comparative TDM frameworks