Trained Abroad, Read at Home: AI, Copyright and the Jurisdiction Gap

The Lift Line

When the machine is trained on servers abroad, the copyrighted text is Indian, the harm is felt in India, and yet the firm can plausibly say the act of copying never happened in India at all, which is the exact seam where the reach of national copyright law tears.

Why This Editorial Matters for Your Exam

This is a rare topic that sits squarely across GS2 (governance, law, the working of institutions) and GS3 (science and technology, intellectual property rights) and lets you write with both legal precision and policy imagination. The examiner is increasingly setting questions on AI regulation, and a candidate who can explain fair dealing and the territoriality of intellectual property in the same answer stands out.

It also teaches a transferable idea: that the internet routinely detaches an activity from any single territory, and that national statutes written for a physical world struggle to keep up. That structural insight applies far beyond copyright, to data, taxation and content regulation alike.

GS Paper 2: government policies and interventions; the functioning of the judiciary; issues arising from the design and implementation of laws.

GS Paper 3: developments in science and technology; issues relating to intellectual property rights.

For Prelims, hold the specifics: the Copyright Act, 1957 and its fair dealing exception under Section 52, including Section 52(1)(a) for private use, research, criticism and review; the concept of text and data mining (TDM) exceptions abroad (the EU’s approach, the UK debate, and the United States “fair use” doctrine); the safe harbour for intermediaries under Section 79 of the IT Act, 2000; and the ANI versus OpenAI proceedings in the Delhi High Court, where on 24 July 2026 the court (Justice Amit Bansal) declined interim relief, treating training storage as fair dealing at this stage while holding it had territorial jurisdiction. For Mains, argue how India should adapt copyright and regulation to global AI training.

Background and Context

News agency ANI sued OpenAI in the Delhi High Court, arguing that its copyrighted reporting had been used without permission to train ChatGPT and that outputs sometimes reproduced or falsely attributed its content. The case is India’s first major test of whether training a large language model on protected material is infringement or falls within a permitted exception. On 24 July 2026, the court declined to grant ANI an interim injunction. The reported reasoning was twofold: at the interim stage, storing published material to train a model looked like fair dealing under Section 52, and outputs produced through retrieval-augmented generation were not substantially similar to ANI’s originals. The larger substantive questions were kept open for full trial.

Running underneath the copyright question was a jurisdiction question. OpenAI’s line of argument, that its training and servers sit outside India, put the court to the test of whether an Indian statute can reach an act of copying performed abroad. The court held that it did have territorial jurisdiction to hear the suit, but the very fact that the point had to be argued at length exposes how easily a global AI firm can frame its core activity as happening nowhere in particular, and therefore nowhere it can be easily regulated.

The Core Argument / Issue

Fair dealing was written for humans, not models

Section 52’s fair dealing exceptions were drafted for a person quoting a passage for review or copying an article for research. Applying that to a model ingesting millions of works to learn statistical patterns stretches the concept in ways Parliament never contemplated. Reasonable judges can disagree on whether machine training is transformative use or industrial-scale copying, which is why the interim view (fair dealing at this stage) is a holding pattern, not a settled rule.

The territoriality problem is the deeper one

Intellectual property law is territorial: an Indian copyright is enforced by Indian courts against acts connected to India. But if the copying happens on foreign servers, the firm has no Indian establishment, and only the reading public is in India, the traditional hooks of jurisdiction loosen. A defendant can argue its activity is extraterritorial while still profiting from Indian users and Indian data. That is the jurisdiction gap.

Learning from how others closed the gap

Other jurisdictions have at least tried to be explicit. The European Union created a text and data mining exception with an opt-out for rights holders; the United Kingdom has debated a similar carve-out and its licensing terms; and the United States relies on an open-ended fair use test that courts are now applying to AI training case by case. India has neither a dedicated TDM exception nor a clear extraterritorial rule, so the burden of writing the law is falling on judges, one interim order at a time.

How to Think About This (Analytical Frame)

Use the frame territorial law versus borderless activity. National statutes assume an act happens somewhere identifiable and someone accountable is present there. AI training breaks both assumptions: the act is distributed across servers abroad, and the value is extracted from Indian data and Indian users who never leave home. The regulatory task is not only to decide whether training is fair, but to decide when Indian law attaches to a foreign act because its effects land in India, the “effects” principle. Separate the two questions cleanly: the substantive question (is training infringement or fair dealing) and the jurisdictional question (can India even reach it). Muddling them produces confusion; treating them in sequence produces a policy.

The Diagram in Words

Indian copyrighted news -> ingested to train a model on foreign servers -> output read by Indian users -> Question 1: fair dealing under Section 52 or infringement? (court: fair dealing at interim stage) -> Question 2: did the act happen "in India"? (firm says no; court holds it has jurisdiction) -> jurisdiction gap for global AI -> [resolution] clear TDM exception + effects-based jurisdiction + licensing framework

Way Forward

1. Legislate a text and data mining exception. Give India a purpose-built rule for AI training, with a rights-holder opt-out and licensing mechanism, rather than forcing judges to bend Section 52 case by case.
2. Adopt an effects-based jurisdiction rule. Clarify by statute that Indian copyright and regulatory law attaches to foreign AI activity when its effects and commercial benefit land in India.
3. Build a licensing market for training data. Encourage collective licensing so that news publishers and creators can be paid for training use, turning a zero-sum dispute into a market.
4. Fix intermediary and output liability. Align Section 79 safe harbour with AI output, distinguishing hosting third-party content from generating it, so responsibility is clear when a model reproduces protected work.

PYQ Linkage and Practice

UPSC has asked on the challenges emerging technologies pose to law and governance, on intellectual property rights and their significance, and on the regulation of the digital economy. This editorial gives you a live case that fuses IPR, technology and the limits of territorial jurisdiction.

Practice question: “Training generative AI on copyrighted content exposes both the limits of the fair dealing exception and a jurisdiction gap in enforcing national copyright law.” Discuss with reference to India, and suggest reforms. (250 words, 15 marks)

Sources: Adgully: Delhi HC denies interim injunction to ANI in OpenAI copyright suit, Outlook Business on the Delhi HC ruling

Source: Trained Abroad, Read at Home: AI, Copyright and the Jurisdiction Gap — Ujiyari.com | Free UPSC & State PCS Editorial Analysis