Copyright in AI generated content in India vests in the human who caused the work to be created, not in the AI system: section 2(d)(vi) of the Copyright Act, 1957 makes the author of a computer-generated literary, dramatic, musical or artistic work "the person who causes the work to be created". But identifying an author does not by itself create a copyright. The work must still be original under section 13, and where the human contribution is a single short prompt there may be no protectable work at all. Anyone researching AI generated content copyright India rules will find no new statute on the point; the position rests on a 1957 Act, a 2007 Supreme Court test for originality, and a government working paper that is still only half written.
That gap between authorship and subsistence is where the disputes will be fought. A company can put its name on a thousand AI-generated product images and still find, when it sues a copycat, that it owns nothing enforceable.
AI Generated Content Copyright India: What the Act Actually Says
Four provisions of the Copyright Act, 1957 do the heavy lifting.
Section 2(d)(vi), inserted by the Copyright (Amendment) Act, 1994, is the only clause in the Act that contemplates a machine producing a work. It says that in relation to any literary, dramatic, musical or artistic work "which is computer-generated", the author is "the person who causes the work to be created". Note what it does not say. It does not say the programmer, and it does not say the owner of the machine. It uses a causal test, and any future dispute will turn on who, on the facts, caused this particular output to exist.
Section 13 restricts copyright to "original literary, dramatic, musical and artistic works", cinematograph films and sound recordings. Originality is a condition of subsistence. If the work is not original, section 2(d)(vi) never comes into play, because there is no copyright to be authored.
Section 17 makes the author the first owner, subject to the employment and commissioning provisos. An employee who generates marketing copy in the course of employment under a contract of service normally leaves the employer as first owner under section 17(c), absent an agreement to the contrary. A freelancer is not an employee, so ownership there has to come from a written assignment.
Section 57 confers moral rights on the author: the right to claim authorship and to restrain distortion prejudicial to honour or reputation. Moral rights are a strong textual signal that the Act contemplates a human author. A model has no reputation to injure.
Key takeaway. Section 2(d)(vi) answers "who is the author". It never answers "is there a copyright at all". Those are two separate questions, and businesses routinely collapse them into one.
These are the four provisions of the Copyright Act, 1957 that decide who owns AI-assisted work and whether there is anything to own.
Section 2(d)(vi)
Inserted in 1994, it makes the author of a computer-generated work the person who causes the work to be created. A causal test, not the programmer.
Section 13
Copyright subsists only in original works. Originality is a condition of subsistence, so if the work is not original the authorship clause never comes into play.
Section 17
The author is first owner, subject to the employment and commissioning provisos. Work made in the course of employment normally leaves the employer as first owner.
Section 57
Moral rights let the author claim authorship and restrain distortion prejudicial to honour or reputation. A strong signal that the Act contemplates a human author.
The Originality Problem: Why Some AI Output Has No Copyright At All
The governing Indian test comes from Eastern Book Company v. D.B. Modak, decided by the Supreme Court on 12 December 2007. The Court rejected the "sweat of the brow" doctrine, under which mere labour and capital earned copyright, and settled on a middle position: the author must have exercised skill and judgment that is more than trivial, producing something with a distinguishable flavour of its own, even if it is not novel or non-obvious in the patent sense.
Apply that to a generative model. Typing "a photorealistic image of a Bengaluru street at dusk" and accepting the first output involves almost no skill and judgment on the human side. The expressive choices are the model's. On the Eastern Book Company standard that is a weak claim.
Change the facts. A designer photographs the street herself, feeds that photograph in as the base, selects a specific style reference, tunes parameters over forty iterations, discards thirty-nine, then composites the survivor with hand-drawn elements in a layered file. Human selection, arrangement and refinement are now substantial and documented. That is a materially stronger claim, and the copyright is properly described as subsisting in the human contribution and in the composite as a whole.
The Copyright Office has already had to grapple with this. In November 2020 it registered an artwork titled Suryast, made by combining an original photograph with a style transfer through an application called RAGHAV, listing the human applicant and the AI application as co-authors. The Office later issued a withdrawal notice dated 25 November 2021 asking the applicant to explain the legal status of the AI application, expressly citing sections 2(d)(iii) and 2(d)(vi). No further action followed and, on reported accounts, the entry stayed on the register. The United States Copyright Office Review Board reached the opposite conclusion on the same work, holding that creations of non-humans may not be protected.
Common mistake. Treating a registration certificate as proof that copyright exists. Section 48 makes the Register only prima facie evidence of the particulars entered. A defendant can still lead evidence that the work was never original, and the Suryast episode shows the Office itself is not settled on AI entries.
Who Is the Person Who Causes the Work to Be Created?
Several people can plausibly claim to have caused an output to exist. The table sets out how each claim looks under the present Act. It is an analysis of the statute, not a prediction about any particular case.
| Candidate claimant | Basis relied on | Strength of the claim today |
|---|---|---|
| The user who wrote the prompt and curated the output | Section 2(d)(vi) with section 13 | Strongest, if the human skill and judgment clears the Eastern Book Company threshold |
| The employer of that user | Section 17(c), contract of service | Strong for works made in the course of employment, absent contrary agreement |
| The client who commissioned an agency | Section 17(b) for photographs, paintings, portraits, engravings and films; otherwise a written assignment | Depends on the category of work and on the assignment clause |
| The AI developer or platform | Terms of service, sometimes a licence back | Weak as authorship; the developer did not cause this specific work |
| The AI system itself | None | Not sustainable; the Act contemplates a person and section 57 presupposes a human |
| Nobody, because the work is not original | Section 13 | A live outcome for thin single-prompt output with no human curation |
Terms of Service Are Not a Substitute for Copyright
Most commercial AI platforms say that as between the platform and the user, the user owns the output. That clause is worth having, but it is not a title deed. A contract can allocate whatever rights exist; it cannot manufacture a right the statute never created. If the output is not original under section 13, the assignment transfers an empty box and the user cannot restrain a third party who copies the same image.
There is a second trap. Some platform terms grant the provider a broad licence to use inputs and outputs for further training. For anyone handling confidential drawings or client material, that clause matters far more than the ownership sentence above it.
Training Data: Where the Delhi High Court Has Reached
Ownership of output is only half the exposure. The other half is whether the model was lawfully trained. India's leading dispute is the suit brought by Asian News International against OpenAI in the Delhi High Court, the first Indian case of its kind. On 24 July 2026 Justice Amit Bansal declined ANI's application for an interim injunction, holding that OpenAI's act of storing ANI's works did not, prima facie, amount to copyright infringement under the Copyright Act, 1957.
Three cautions follow. It was an order on an interim application, not a final adjudication of the suit. Refusal of an interim injunction is not a declaration that training on copyrighted material is lawful in all circumstances. And the position on outputs that reproduce protected expression verbatim remains distinct from the position on inputs. Businesses building on third-party models should not read the order as a clearance.
Key takeaway. Owning your AI output and not infringing somebody else's copyright are separate risks. A clean assignment from your AI vendor does not protect you against a claim that the output reproduces a third party's protected expression.
Where the Law Is Headed: The DPIIT Working Paper
The Department for Promotion of Industry and Internal Trade constituted a committee on 28 April 2025 to examine generative AI and copyright. In December 2025 the committee published Part 1 of its Working Paper on Generative AI and Copyright, subtitled "One Nation One License One Payment", and opened it for public comment.
Part 1 deals only with the input side, the use of copyright-protected works as training data. It records that technology industry stakeholders largely pressed for a broad text and data mining exception while content industry stakeholders unanimously favoured voluntary licensing.
The committee expressly deferred the output questions to Part 2: copyrightability of AI-generated works, identification of authorship, applicability of moral rights, and attribution of liability for infringing outputs. Until Part 2 lands and Parliament acts on it, section 2(d)(vi) and Eastern Book Company remain the operative law.
How to Keep AI-Assisted Work Protectable
The objective is to build an evidentiary record that a human exercised real skill and judgment, and to secure ownership by contract from everyone who touched the file.
- Decide at the outset whether the deliverable needs enforceable copyright or whether confidentiality and contract are enough. A social media post rarely needs a copyright strategy; a product interface or a character design does.
- Keep a creation log recording the base assets you supplied, the prompts and parameters, the number of iterations, what was rejected, and the post-generation edits. This is the evidence that section 13 originality will turn on.
- Preserve layered and intermediate files, not just the exported image or the final document. Layers show human arrangement.
- Put the human contribution beyond argument: supply your own photographs, sketches, data or text as inputs, and rework the output rather than shipping it raw.
- Fix ownership in writing. Employment contracts should confirm section 17(c) ownership and add an assignment for anything outside it. Every freelancer and agency agreement needs an express assignment covering AI-assisted deliverables.
- Read the vendor terms for both the output ownership clause and the input licence, and record which tool and version produced which asset.
- Consider registration for commercially significant assets. Registration is optional, but the Register is prima facie evidence under section 48 and a certificate shortens arguments at the interim stage. Our note on copyright registration in India sets out the process.
- Clear high-value output before launch through reverse image search and a manual review for recognisable characters, logos and brand elements.
- If infringement occurs, weigh the civil route under section 55, which allows injunction, damages and accounts, before a criminal complaint under section 63. Criminal complaints are now processed under the Bharatiya Nagarik Suraksha Sanhita, 2023, which replaced the Code of Criminal Procedure, 1973. Our note on remedies for copyright infringement compares the routes.
Four habits from that list do most of the work in keeping AI-assisted output protectable.
Keep a creation log
Record the base assets you supplied, the prompts and parameters, the iterations and rejects, and the edits made after generation. This is the evidence originality turns on.
Preserve layered files
Keep the intermediate and layered files, not only the exported image or the final document. Layers are what show human arrangement rather than raw machine output.
Fix ownership in writing
Employment contracts should confirm section 17(c) ownership and add an assignment for anything outside it. Freelancer and agency agreements need an express assignment for AI-assisted deliverables.
Clear before launch
Run reverse image search and a manual review of high-value output for recognisable characters, logos and brand elements before anything is published.
Indicative Costs and Timelines
The figures below are broad indications only. They are not quotations, they vary widely with complexity and forum, and nothing here is a promise about any matter.
| Step | Indicative timeline | Indicative cost |
|---|---|---|
| AI use and IP ownership policy for a company | 2 to 4 weeks | Usually a modest fixed fee |
| Copyright registration for one work, uncontested | Commonly several months from filing to certificate, longer if objected | Statutory fee per work under the Copyright Rules plus professional fees |
| Cease and desist notice with negotiated resolution | 2 to 8 weeks | Lowest of the options where the infringer is reachable |
| Civil suit with an interim injunction application | Interim orders in weeks to months; final disposal typically years | Substantially higher, driven by hearings and evidence |
Deadline warning. Each infringing act starts its own limitation clock, and sitting on a known infringement weakens an interim injunction, because delay and acquiescence are standard defences at that stage. If you find a copy, act quickly rather than waiting to see whether it grows.
A Practitioner's Observation
In practice the failure is almost never at the statute. It is at the paperwork. Founders arrive with a full brand system, a mascot and a set of illustrations, all generated with AI assistance by a contractor paid through a UPI transfer against a WhatsApp brief, and there is no assignment, no record of what the human supplied, and no version history. The legal analysis at that point is not difficult, it is simply unhelpful, because nothing can be proved. The teams that end up in a strong position are the ones whose designers kept their layered files and whose contracts said, in one plain sentence, that all copyright in the deliverables vests in the company. Anyone building a brand on generative tools may also find our note on IP protection for startups and our intellectual property law page useful background.
Related guides and where to get help
- Who Owns a Bollywood Song? Copyright, Producers, Music Labels and Legacy Rights Explained
- Copyright Infringement Remedies in India
- Copyright Registration in India: Process, Works, Term and Benefits
Frequently Asked Questions
Can an AI system be named as the author of a work in India?
The Act contemplates a person as author, and section 57 moral rights presuppose a human. The Copyright Office did once register an artwork listing an AI application as co-author, then issued a withdrawal notice citing sections 2(d)(iii) and 2(d)(vi). Naming an AI as author is not a safe course.
If I use a chatbot to write my website copy, do I own the copyright?
As between you and the platform, the terms usually say you do. Whether copyright exists at all depends on originality under section 13. Copy produced from one generic prompt with no editing is a weak claim; copy you have researched, restructured and rewritten is much stronger.
Does registration make my AI-generated work safe?
No. Registration is optional and the Register is only prima facie evidence under section 48. It helps at the interim stage, but a defendant can still show the work was never original.
Who owns AI-generated work made by my employee?
If it was made in the course of employment under a contract of service and there is no agreement to the contrary, the employer is generally first owner under section 17(c). For contractors and freelancers you need a written assignment, because section 17 will usually not help.
Can I be sued because the AI tool was trained on copyrighted material?
The training claim normally lies against the model developer, not against you. Your exposure arises if the output you publish reproduces a third party's protected expression, a recognisable character or a logo. Clearance checks before launch address the real risk.
What did the Delhi High Court decide in the ANI case against OpenAI?
On 24 July 2026 it declined ANI's interim injunction, holding that OpenAI's storage of ANI's works did not prima facie amount to infringement. That was an interlocutory order in a pending suit, not a final ruling that training is lawful in every case.
Is India about to change the law on AI and copyright?
A DPIIT committee constituted on 28 April 2025 published Part 1 of a working paper in December 2025 on training data, and has said Part 2 will address copyrightability of outputs, authorship, moral rights and liability. Nothing has been enacted.
Can I copyright a prompt?
A short instruction is unlikely to clear the originality threshold. An extensive structured prompt library with substantial expressive content stands a better chance as a literary work, and in most commercial settings such a library is better protected as a trade secret under confidentiality obligations.
This article is general information on Indian copyright law as it stands in August 2026 and is not legal advice. The position on AI-generated works is developing and every matter turns on its own facts.
