Copyright in the Age of AI: How Indian Law Handles Training Data, Fair Dealing and Publishers’ Rights

What does the Copyright Act, 1957 protect?

The Act protects original literary, dramatic, musical and artistic works. It also covers films and sound recordings. Section 13 lists these classes of work.

Section 14 gives the owner exclusive rights. These include copying, adapting and communicating the work to the public.

What is the idea–expression split?

Copyright protects how an idea is expressed. It does not protect the idea itself. Anyone can write a new book on the same theme.

This split matters for AI. A model that learns patterns may use ideas. Copying exact text is a different matter.

What does section 52 say?

Section 52 lists acts that are not infringement. Examples are private use including research, criticism or review, and reporting of current events.

The list is closed. If a use is not on the list, it is not fair dealing, however fair it feels.

How is this different from US fair use?

The US has an open test of fair use. A judge weighs four factors. These include the purpose of the use and the effect on the market for the original.

India has no such open test. So Indian courts must fit AI training into one of the listed purposes.

Why is training the hard question?

To train, a developer first collects and stores works. Storing a copy can be infringement unless an exception applies. Rights holders say training is mass copying without permission.

AI firms say the model learns patterns, not the works themselves. They also say training is a form of research.

Why is output a separate question?

An output can copy a work if it is substantially similar to it. This is the classic test. A model that repeats a news article word for word may infringe.

A fresh answer that only uses facts or ideas does not. Courts so far look at the outputs actually shown to them.

Can an AI be an author in India?

The Act has no AI author. Section 2(d)(vi) says that for a computer-generated work, the author is the person who causes the work to be created.

That points to a human. Which human is still unclear: the user, the developer, or both.

What was the RAGHAV case?

An Indian registration for the artwork “Suryast” named an AI tool, RAGHAV, as co-author with its creator Ankit Sahni. The Copyright Office later questioned this. The final outcome is not clearly reported, so do not memorise a result.

The US Copyright Office refused to register the same work. Canada registered it. This shows how countries differ.

What are the main policy options?

There are three ideas. One is a free exception for text and data mining. Another is an opt-out, where creators say “do not use my work”. The third is a licence, where developers pay.

Each option balances AI growth against creators’ income in a different way.

What did the DPIIT committee propose?

The Department for Promotion of Industry and Internal Trade set up an eight-member committee on 28 April 2025. Part 1 of its working paper came out on 8 December 2025.

It proposed a hybrid model. Developers get a mandatory blanket licence for lawfully accessed works. Creators get a statutory right to payment. There is no opt-out.

Royalties are due only when the AI system is commercialised. A government-appointed committee would set the rates. A body called CRCAT would collect and share the money.

What does the EU do?

The EU has a text-and-data-mining exception in the DSM Directive. Rights holders can opt out in a machine-readable way, such as robots.txt.

The AI Act adds duties. Under Art. 53(1)(c), providers of general-purpose AI must follow a copyright policy that respects those opt-outs. Under Art. 53(1)(d), they must publish a summary of the training data. These duties apply from 2 August 2025.

What do US courts say?

US courts apply fair use case by case. In Kadrey v Meta (25 June 2025), the judge gave Meta partial summary judgment on fair use. He said the ruling was narrow, because the authors showed little evidence of market harm.

In Bartz v Anthropic, a class settlement of at least USD 1.5 billion followed a finding about pirated books. The New York Times case against Microsoft and OpenAI, filed in December 2023, is pending.

What is the UK position?

The UK ran a consultation from December 2024 to February 2025. Its report on copyright and AI, published on 18 March 2026, did not adopt a new training exception for now. The government said it would keep watching.

What happened in ANI v OpenAI?

ANI, a news agency, sued OpenAI in the Delhi High Court in November 2024. Publisher groups joined or intervened, including the DNPA (Indian Express, Hindustan Times, NDTV and others) and the Federation of Indian Publishers.

On 24 July 2026, Justice Amit Bansal refused an interim injunction. He prima facie held that storing material for training falls within s.52(1)(a) fair dealing. He found the outputs shown were not substantially similar to ANI’s works. He also held that Delhi has jurisdiction, though OpenAI says its servers are abroad.

This is an interim view. The suit continues. Treat it as an example, not settled law.

AI and copyright: where India stands

• [Precedent] On 15 September 2026 a Division Bench of the Delhi High Court issued notice to OpenAI on ANI’s appeal against the July order. The next hearing is listed for early December 2026 (reports give 5 or 8 December). So what: the appeal may decide how far s.52 covers AI training. (Source: LiveLaw) (Source: Bar & Bench)
• [Precedent] On 24 July 2026 the Delhi High Court refused ANI an interim injunction against OpenAI. It prima facie treated training as fair dealing and upheld Delhi jurisdiction. So what: it is the first Indian court view on AI training, though only interim. (Source: Supreme Today) (Source: LiveLaw)
• [Policy] On 8 December 2025 the DPIIT committee released Part 1 of its working paper. It proposed a mandatory blanket licence with statutory payment. So what: it could replace an exception or opt-out with a pay-per-use system. (Source: Storyboard18) (Source: IAM)