Can AI Train on Your Work? What India's Proposed AI–Copyright Licensing Model Means for Creators and Businesses

Somewhere in a data centre, a model has probably read your blog, your product manual or your photographs. Whether that was lawful in India is, as of September 2026, genuinely unsettled. The government has proposed one answer. A court has, for now, leaned towards another.

If you create content, you want to know whether you can stop it or get paid for it. If you build or buy AI, you want to know what you might owe. This guide explains AI training and copyright in India as it stands: what the law says today, what DPIIT has proposed, what the Delhi High Court said in July, and what is sensible to do while the question is open.

Quick answer

India's Copyright Act, 1957 has no specific exception for text and data mining, so AI training on copyrighted work is not expressly permitted. In December 2025 a DPIIT committee proposed a "hybrid model": AI developers would get a mandatory blanket licence to train on lawfully accessed works, with no opt-out, and would pay royalties into a central body once their AI is commercialised. This is not law. Meanwhile, the Delhi High Court on 24 July 2026 refused an interim injunction against OpenAI, taking a prima facie view that storing news content for training was fair dealing. That ruling is under appeal.

What Indian copyright law says today

Three provisions frame the question.

Section 14 gives the copyright owner the exclusive right to reproduce a work "in any material form including the storing of it in any medium by electronic means". Copying a work into a training dataset is, on its face, storage by electronic means.

Section 51 says copyright is infringed when someone does any of those exclusive acts without a licence.

Section 52 lists what does not infringe. The relevant part is fair dealing for "private or personal use, including research", criticism or review, and reporting current events. There is also an exception for transient or incidental storage during electronic transmission. There is nothing on text and data mining or machine learning.

So the honest summary of the statute is this: training is a reproduction, and whether it is excused depends on how far fair dealing stretches. Until recently, nobody in India had tested that.

The DPIIT working paper: "One Nation, One Licence, One Payment"

On 28 April 2025, the Department for Promotion of Industry and Internal Trade (DPIIT) set up an eight-member committee, chaired by Additional Secretary Himani Pande, to assess whether existing law could cope with generative AI. It published Part 1 of its working paper on 8–9 December 2025, titled Working Paper on Generative AI and Copyright Part 1: One Nation One License One Payment. Comments closed on 6 February 2026 after an extension.

The proposal, in the committee's own terms:

• A mandatory blanket licence in favour of AI developers to use "all lawfully accessed copyright-protected works" for training.
• A statutory remuneration right for copyright holders, so the licence is not free.
• No opt-out. Rights holders "will not have the option to withhold their works for use in the training of AI Systems."
• A central collecting body, the Copyright Royalties Collective for AI Training (CRCAT), made up of copyright societies and collective management organisations, one for each class of work, and designated by the Central Government.
• Royalties as a percentage of revenue from AI systems trained on copyrighted content, payable only once the AI is commercialised, at rates fixed by a government-appointed committee and open to judicial review.
• Payment to non-members too, if they register their works to receive AI training royalties.

The paper sets no royalty percentage. Some commentators have read it as reaching back to models already trained; others read it the opposite way. Until the text is clarified, treat that question as open.

What the committee rejected, and why

This part explains the design.

A general text and data mining exception was rejected because it would "leave human creators powerless to seek compensation".

An exception with an opt-out, the European model, was rejected as unfair to small creators. The paper's reasons are practical: most creators don't know how to opt out, can't tell whether their opt-out was respected, and lose control permanently once content is stripped of metadata and absorbed into a model. An opt-out also shifts the burden onto creators.

Voluntary licensing, which the content industry favoured, was not adopted either. The logic of the paper as a whole suggests why: with millions of rights holders and a handful of developers, private deals would reach the big publishers and labels and leave almost everyone else out.

Nasscom dissented. Its position, recorded in the paper, is for a text and data mining exception where access is lawful, with a machine-readable opt-out for public web content and contractual reservation for content behind a paywall.

Where the proposal stands now

As of 24 September 2026:

• Part 2, on whether AI outputs can be protected and who is the author, has not been published, though DPIIT indicated in December 2025 that it would follow within about two months.
• No bill implementing the hybrid model has been tabled. In July 2026, a group of IP academics wrote to DPIIT noting reports that a Copyright Act amendment was at an advanced stage and asking for the draft to be published before it goes to Parliament. What that amendment contains is not public.
• Nothing has changed in the Copyright Act. The model is a proposal and has no legal force.

The ANI v. OpenAI ruling of 24 July 2026

While the policy debate ran, the Delhi High Court heard India's first major AI training case. ANI Media sued OpenAI in November 2024, alleging that ChatGPT was trained on, and reproduced, its news content. Publishers, music labels and technology groups joined on both sides.

On 24 July 2026, Justice Amit Bansal decided ANI's application for an interim injunction. The court:

• held that it had territorial jurisdiction, even though OpenAI's servers are in the United States, treating storage abroad as the last step in a chain that began with access from India;
• took the prima facie view that storing ANI's works for training fell within fair dealing under Section 52(1)(a);
• found the ChatGPT outputs ANI relied on were not substantially similar, and in some cases could not have come from training at all, because the articles were published after training ended;
• concluded that an injunction would harm OpenAI and the public more than refusing one would harm ANI.

The application was dismissed. ANI appealed, and on 15 September 2026 a Division Bench issued notice to OpenAI. The appeal is listed for early December 2026.

Two cautions. First, this was an interim ruling on a prima facie view, not a final judgment that AI training is lawful in India. The suit itself is still to be tried. Second, it pulls against the DPIIT paper, which concluded that tweaking fair dealing would not resolve AI developers' legal exposure. The court and the committee are, for now, looking at the same statute and reaching different instincts.

How other countries have answered the same question

Indian businesses serve global markets, so the foreign rules matter.

India's proposal is the most interventionist of these: no opt-out, but guaranteed payment.

What this means for creators

If you write, photograph, compose, design or publish, here is what you can do now.

• Register the works that matter. Registration is optional, but under Section 48 the Register is prima facie evidence of its entries. If the DPIIT model is enacted, registration with a collecting body is also how non-members would claim royalties. Keep dated originals as well.
• Set clear terms of use. State on your website and in your licences that content may not be used for AI training without permission. Under current law that won't stop everyone, but it records that you did not consent, and it is effective for EU-facing use.
• Use crawler controls with your eyes open. OpenAI's GPTBot and Google's Google-Extended can be blocked through robots.txt, and Google says blocking Google-Extended does not affect Search. Blocking OpenAI's separate search crawler, though, removes you from ChatGPT's search results. Decide which trade-off you want.
• Watch the collecting-society question. If CRCAT happens, the society that represents your class of work will decide a great deal about what you receive. It is worth knowing now which one that is.

What this means for businesses using or building AI

• If you buy AI tools, ask your vendor for an intellectual property indemnity, warranties about how training data was obtained, and clear terms on who owns and may use outputs. These are ordinary commercial protections, and they matter more while the law is unsettled.
• If you train or fine-tune models, keep a record of what data you used and on what basis. A record of lawful access is the entry condition under the DPIIT model and the core of any fair dealing defence under today's law.
• If you fine-tune on client or partner content, get a written licence. The ANI ruling concerned public news content at an interim stage; it is not a safe harbour for proprietary material.
• Model the royalty risk. A revenue-based royalty on commercialised AI systems would land on your margin, not your research budget.

How MYCrave can help

MYCrave Consultancy & Services works with creators, publishers, software companies and institutions on protecting and managing their works. On this question, practically:

• Copyright registration for the works that carry commercial value, including literary works, artistic works and software;
• a content-rights audit that sorts what you own, what you license in, and what your website and contracts currently say about AI use;
• licence and terms-of-use drafting that records your position on AI training;
• monitoring for when Part 2 is published, a bill is tabled, or the ANI appeal is decided, with a note on what it changes for you.

An open question, so decide what you can control

The law on AI training in India may be settled by Parliament, by the courts, or by both, and not necessarily in the same direction. You can't control that. You can control whether your key works are registered and documented, whether your terms say what you want, and whether your AI contracts protect you.

Frequently asked questions

Keep reading