Who Owns the Copyright in AI-Generated Work in India

Take the question at its hardest. A model produces a novel of ninety thousand words. No person outlined it, wrote a sentence of it, or decided how it ends. Or a picture arrives from one line of instruction, and the person who typed the line could not have said in advance what would appear on the screen. Nobody collaborated with the machine in any sense a court would recognise. Is there copyright in the novel or the picture, and if there is, whose is it?

Half of that question was answered in Delhi on 31 August, when the Registrar of Copyrights decided the only application India has ever had on the subject.¹ The answer is that the work is not disqualified from copyright merely because a machine made it. The other half was not answered, and the same order is the reason it cannot be.

The work is not the problem

Section 13 of the Copyright Act, 1957 says copyright subsists in original literary, dramatic, musical and artistic works. The Act does not define original, so the content of the word comes from the case law, and in India that means Eastern Book Company v. D.B. Modak: not novelty, not inventiveness, but a minimal degree of creativity and independent creation rather than copying.²

The Copyright Office had put the objection in the strongest form available to it. Its letter of 7 May required the applicant to show how “the algorithmic, statistical pattern recognition processes of a neural network” could satisfy a threshold of creative expression and intellectual effort, “given the admitted absence of human aesthetic judgment or conscious intention”.³

The Registrar rejected his own office's objection:

The fact that the work was generated through algorithmic or computational processes cannot, by itself, render the resulting expression unoriginal. Statistical processing and pattern recognition describe aspects of the mechanism by which the output was produced; they do not necessarily determine whether the resulting expression is copied, commonplace or devoid of creative character. A work is not deprived of originality merely because technology played an indispensable role in its production.

He went further, and dealt directly with the absence of a human at the moment of creation. Where a person supplies and curates the inputs, configures and initiates a generative process that is not trivial, “the absence of real-time human intervention at the final computational stage does not, by itself, defeat originality”. The two enquiries are separate: section 13 asks whether the resulting work has independently generated expressive character, and section 2(d)(vi) separately identifies the person to whom authorship is attributed. It would be wrong, he said, either to treat the machine as the bearer of creativity or to insist that the person named as author manually selected every expressive element of the output.⁴

This is not the American position and it was not an accident. In Thaler v. Perlmutter the District of Columbia Circuit held in March 2025 that human authorship is a statutory requirement, and the consequence there is that a work with no human author belongs to nobody from the instant it exists.⁵ India has not gone there. Here the output of a machine is capable of being a protected work. Someone in Mumbai running a studio on these tools has subject matter that exists in law. Their competitor in New York does not.

Which makes the rest of the Act the problem.

Copyright in India is issued to a person

A book and a painting are both inside section 2(d)(vi), which was inserted in 1995 and says that for any literary, dramatic, musical or artistic work which is computer-generated, the author is “the person who causes the work to be created”.⁶ There is no fifth option in that clause. Either a person caused the work, or the clause supplies no author for it.

Everything downstream runs through that person. Section 17 makes the author the first owner of the copyright, subject to exceptions that all identify some other person. Section 22 measures the term of protection as sixty years from the beginning of the calendar year after the one in which the author dies. Section 45 lets the author, publisher or owner apply to enter particulars in the Register. Section 57 gives the author the right to claim authorship and to restrain distortion of the work. And section 16 closes the system: nobody is entitled to copyright in any work otherwise than under and in accordance with the provisions of the Act.⁷

The Copyright Office saw this in May, before any of it was argued. The first ground of objection reads, in full:

The Act's framework, including moral rights under Section 57 and the term of copyright under Section 22, presupposes a human author possessing legal personality and juridical will.

That objection was never answered. It did not have to be, because the Registrar found a human author on the facts in front of him.

The test, and the sentence that decides your case

Stephen Thaler's argument was that his system, DABUS, had produced the picture by itself and should be recorded as the author. The Registrar held that a machine is not a natural or juristic person and cannot be entered in that column. The reported line stopped there. The reasoning did not.

Reading “causes the work to be created” as pointing to the originator rather than to the last step in the process, the Registrar adopted what he called the effective cause or mastermind enquiry: not who performed the final act, but who conceived the work, determined its direction and is responsible for its coming into existence. Autonomy in execution, he held, is not conception. A system that runs by itself after activation is still running inside an architecture, a set of objectives and a training method chosen by somebody. To hold otherwise “would elevate every autonomous technological process into an independent source of legal authorship”.⁸

Read quickly, that sounds like an answer to the question this essay asks. However autonomously the machine ran, there is always a person further back, and the provision finds them.

It does not, and the Registrar said so twice. The first time, in the paragraph setting out the test:

A person does not become the author of every output merely because that person owns, designed or developed the computer system. There must be a legally sufficient connection between the acts of the identified person and the origination of a particular work.

The second time is the passage that decides the ordinary Indian case, and it has gone almost unnoticed:

The inquiry is not satisfied merely by locating a human being somewhere in the causal history of the work … What must be identified is: who is the person who, having regard to the making of the particular work, exercised such creative control and made such proximate and material arrangements that the work can fairly be regarded as having been caused by that person. The provision requires a legally meaningful nexus between the identified person and the creation or origination of a specific work, not simply the existence of a person upstream from an autonomous chain of generation.⁹

Two requirements, then. The nexus must be to a particular work, and the arrangements must be proximate and material to that work. Building the machine does not do it. Being somewhere in the causal history does not do it.

Why Thaler passed and the ordinary case does not

Thaler satisfied the test because he had personally done a list of things. On his own account he conceived and built the system, configured how it ran, supplied the visual inputs from photographs he had taken himself, curated the linguistic material, wrote the textual descriptions linking the two, and started the process from which this particular picture emerged. Then the sentence the finding rests on: no other natural or juristic person is identified as having undertaken these acts.¹⁰

That sentence is true of Stephen Thaler and of very few other people. He is at once the builder of the model, the curator of what it learned from, the supplier of the specific inputs and the person who pressed start. In ordinary use those acts are split, and mostly they are not performed by anybody in relation to any particular work.

Put the novel back on the table. The company that trained the model has a claim to having designed and developed the system, and paragraph 98 rules that out by itself. It made no arrangements proximate to this novel; it does not know the novel exists. The person at the other end typed an instruction, which is not the supply and curation of inputs, not configuration, and on any honest description not creative control over ninety thousand words of expression they could not have predicted. The machine, which did originate the expression, is disqualified for not being a person.

Nobody in that chain has a legally meaningful nexus to the particular work. Not because the Registrar overlooked them, but because he wrote the test that excludes them, and the facts in front of him gave him no reason to notice.

The order also cannot tell you what a prompt is worth, and it was careful not to pretend otherwise. Every finding is expressly confined to the material placed in that proceeding. DABUS was operated with no text prompt at all, so the question of whether writing an instruction is a proximate and material arrangement simply did not arise. India therefore has a test for authorship of machine-made work, applied once, to the one fact pattern in which it produces an easy answer.

What the Act does with a work that has no author

Nothing. That is the honest answer, and it is worth walking through slowly, because the consequences are not symmetrical with the American rule even though they look similar from a distance.

The work is original, so it is the kind of thing in which copyright subsists under section 13. But section 17 vests first ownership in the author, and there is no author to vest it in. The exceptions in section 17 do not help: each of them moves ownership from the author to some other identified person — a proprietor of a newspaper, a person who commissioned a photograph or painting for valuable consideration, an employer under a contract of service — and each presupposes an author whose rights are being redirected. There is no clause that creates ownership where authorship failed.

Term is worse. Section 22 gives sixty years from the beginning of the calendar year following the year in which the author dies. An entity that cannot die cannot start that clock, and a person who was never the author cannot either. Section 23 does not rescue it. Anonymous and pseudonymous works get sixty years from publication, but the section is about a work “published anonymously or pseudonymously”, which describes an author whose identity is withheld, not the absence of an author; and the proviso, which switches back to sixty years from the author's death if the identity is later disclosed, confirms that there is a person there waiting to be named.¹¹

Registration is closed off by the same gap. Section 45 requires particulars, and the Registrar's own order holds that particulars found to be legally inconsistent cannot be entered, that an application naming a machine as author cannot be registered, and that no descriptive notation recording the machine's role can cure it. The applicant in this case was offered the chance to put a human name in the author column and refused. Someone whose novel genuinely has no human author has nothing to put there and no refusal to blame.¹²

And section 57, the right to claim authorship and to object to mutilation, belongs to the author and is exercisable after death by the author's legal representatives. It has nobody to attach to.

So India arrives at the same practical destination as the United States by an entirely different route, and the difference in route matters. The American rule says the work is not copyrightable subject matter at all, which is at least a clean answer that anyone can plan around. India says the work is perfectly good subject matter and then has no mechanism to issue the right to anyone. The output of a genuinely autonomous system is, on the current state of Indian law, an original work that nobody owns: free for anyone to copy, with no one holding standing to complain about it.

Autonomy makes the claim weaker, not stronger

The direction of travel runs against the person hoping to own this material, and the reason is structural.

Where a human closely directs the expression, authorship is easy — and section 2(d)(vi) is not needed, because someone using a tool is the author of the result under the ordinary clauses of section 2(d) anyway. A photographer using a camera is the author of the photograph. The special provision earns its place only in the cases where the human did least, which are precisely the cases in which the proximate and material arrangements are hardest to find.

Systems that plan their own steps sharpen this. A person who states a goal and leaves the machine to decide what to produce, in what order and in what form, is further from the expression than a person writing prompts, not closer. Each increase in autonomy strengthens the argument that the output is original expression copied from nothing, and weakens the argument that any identifiable person made proximate arrangements for the particular work. The two findings move in opposite directions, and the Act only works when they arrive together.

India chose these words

Sub-clause (vi) was modelled on section 9(3) of the United Kingdom's Copyright, Designs and Patents Act 1988, which makes the author of a computer-generated work “the person by whom the arrangements necessary for the creation of the work are undertaken”. Parliament did not copy that wording. It wrote “the person who causes the work to be created” instead.

The difference is the whole of the present problem. The British formula asks who made the arrangements, which in commercial use points at whoever deployed the system, and a company that operates a service makes arrangements for its outputs whether or not it conceived any of them. The Indian formula asks who caused the work, which the Registrar has now read as the originator, the mind behind it. On identical facts those provisions name different people, and in the case of a work nobody conceived, the British provision still names somebody and the Indian one does not.

The United Kingdom has decided to delete its version. In the Government's report on copyright and artificial intelligence of 18 March this year, section 9(3) was found to be unclear and little used and recommended for repeal, leaving works assisted by these tools protected as ordinary works on ordinary principles, and works made solely by a machine protected not at all.¹³ India borrowed the idea in 1994 and interpreted it for the first time five months after the country that drafted it resolved to abolish it.

Nobody in government has answered this

Not for want of being asked. The Parliamentary Standing Committee on Commerce recommended in its 161st report in July 2021 that a separate category of rights be created for artificial intelligence and related innovations.¹⁴ In February 2024 the Minister of State for Commerce and Industry told the Rajya Sabha the opposite: that India grants adequate protection for works created by legal persons, and that there is no requirement to create a separate category of rights for AI in the Indian regime.¹⁵ Thaler relied on that statement as an official confirmation that machines could be authors. The Registrar held it did no such thing, and read it as confirming that protection runs to legal persons, which is the point at issue.

Meanwhile the department that houses the Copyright Office had set up a committee. In April 2025 it constituted an eight-member group to examine whether the Copyright Act is adequate for generative artificial intelligence. On 8 December 2025 the committee published a working paper, Part I, dealing with whether protected material may lawfully be used to train models, and expressly set aside copyright in AI outputs for a Part II. Ten days later the Minister of State told the Lok Sabha that authorship, ownership and copyrightability of AI-generated works were under examination. Part II was expected within two months. Nine months on, it has not appeared.¹⁶

The Registrar, for his part, sent the question where it belongs. Whether authorship should ever be extended to autonomous systems, he held, “remains a policy decision strictly reserved for Parliament, and cannot be introduced via administrative reinterpretation”.

What you actually have to prove

Most arguments about owning this material run the wrong case, and the order is unusually clear about why. There are two requirements. They attach to different things, and only one of them is about you.

Originality under section 13 is a question about the work. Authorship under section 2(d)(vi) is a question about a person. The creativity requirement sits on the first of those, which means that the thing people most often try to prove — that they were creative — is not what the first limb asks and not what the second limb accepts.

Look at the standard the Registrar actually applied. The work must not be copied from an identifiable pre-existing work, and it must carry “at least a minimal degree of non-trivial expressive character”. He is expressly not assessing “artistic merit, aesthetic quality, novelty or inventive ingenuity”. The examination is confined to whether the expression is independently generated and is not “so commonplace, inevitable or mechanically predetermined” as to fall below the modest threshold in Eastern Book Company.¹⁸

Then see what satisfied it. The picture qualified because it “was not supplied to the system in its final form” and was not “the predetermined or inevitable reproduction of any particular item in the training material”, and because it contained a particular arrangement of forms, colours, tonal variations and spatial relationships. Every one of those is a fact about the machine's process and the resulting image. None is a fact about Thaler's mind. The office had objected that there was an “admitted absence of human aesthetic judgment”, and the Registrar answered that this does not matter to section 13.

So the answer to what creativity you must show to get past originality is: none of your own. You are not proving that you were creative. You are proving two negatives about the output — that it is not a copy of something identifiable, and that it is not the inevitable result of a trivial or mechanical operation. Those are proved by the character of the work and a technical account of how it was produced, not by an account of your intentions.

What you must prove about yourself is causation. Causation is not creativity, and the difference is the whole practical point.

The acts that count

The list is short and the Registrar gives it twice. A person who supplies and curates the relevant inputs, configures a generative process that is not trivial, and initiates it, thereby causing an independently generated expressive work to come into existence, is the author. In Thaler's case the acts relied on were that the inputs were photographs he had taken himself, the linguistic material was curated by him, he wrote the descriptions linking the two, and he started the run from which this particular work emerged.

Notice what is not on that list. Nothing requires you to have chosen a single element of the final expression. Thaler chose none of them, and did not see the picture until it existed. The qualifying acts are preparatory and directive rather than expressive, which is exactly why the provision was enacted: works under section 2(d)(vi) are by definition works whose expression a person did not compose.

Against that sit the two disqualifications, and they are where ordinary users lose. Owning, designing or developing the system is not enough by itself. Being somewhere in the causal history is not enough. What is required is “such creative control” and “such proximate and material arrangements” that this particular work can fairly be regarded as caused by you.

Which gives a workable rule of thumb. Ask whether your acts were directed at this work or at your capability in general. Choosing a model, paying for a subscription and learning to use it well are directed at your capability. Assembling the specific reference material, setting the specific parameters, and running the specific process that produced this specific output are directed at the work. The first set is worth nothing under section 2(d)(vi) however much it cost. The second set is the case.

Four grounds, strongest first

One: stay out of section 2(d)(vi) altogether. The provision applies to works that are computer-generated. Where a person shapes the final expression — editing the draft, redrawing parts of the image, selecting and rewriting across many outputs until the result is their own composition — the work is authored under the ordinary clauses of section 2(d) by a person using a tool, as a photographer is the author of a photograph the camera physically produced. This ground needs no novel provision, no administrative order and no litigation about machines. Whenever the facts allow it, this is the ground to be on, and the way to make the facts allow it is to do the work and keep the drafts.

Two: selection and arrangement. Section 2(o) defines a literary work to include compilations. Eastern Book Company is itself a case about editorial input into material the editors had not written, and it is the authority for the proposition that skill and judgment in selection and arrangement can be original even where the underlying content is not.¹⁹ A collection of generated images with a chosen sequence, a curated dataset, a catalogue, an edited anthology: the compilation is a separate work with an ordinary human author, and it holds even if no individual component has one. For anyone sitting on a large body of machine output, this is the most underused route in the Act.

Three: what you added. Where you take an output and work on it, your contribution is protected expression in its own right, and your copyright runs to that contribution rather than to the raw output. The claim is narrower than people want it to be, but it is a claim that does not depend on any of the contested questions.

Four: causation under section 2(d)(vi). The Registrar's checklist, set out above. It is last not because it is wrong but because it rests entirely on one administrative order that no court has tested, and because it is the only one of the four that fails outright if a court later reads “causes the work to be created” more narrowly.

How you prove it

Three things, and the first two cost nothing.

Put your name on it. Section 55(2) provides that where a name purporting to be that of the author appears on copies of the work as published, that person is presumed in infringement proceedings to be the author unless the contrary is proved. That presumption does not create authorship, but it decides who has to prove what, and in a field where the underlying law is unsettled the burden is most of the argument. Then register, because section 48 makes the Register prima facie evidence of the particulars entered in it, and a certified extract is admissible without further proof.²⁰

Then keep the record. Per work rather than per project, and made as you go rather than reconstructed once there is a dispute. What is worth keeping depends on which of the four grounds your claim runs on, and most claims run on more than one:

• Your drafts and your edit history. This is the evidence for the first ground and the third, and it is the most valuable thing on the list, because it is the only thing that shows a person shaping the expression rather than receiving it. Keep the output as it first arrived, keep the version you published, and keep the states in between. Version control, a document with history enabled, or dated exports all do the job; what matters is that the sequence is contemporaneous and that the first state survives. Someone who can produce the raw generation alongside four rounds of their own revision is not arguing about section 2(d)(vi) at all — they are an author using a tool, which is a much older and safer argument.
• Where the inputs came from. The photographs, text, data or reference material you supplied, with enough provenance to show they were yours to supply. Supplying and curating the inputs is the first of the qualifying acts, and it was the first thing the Registrar listed in Thaler's favour.
• What you configured. The model, the settings and parameters, any fine-tuning or reference material, and the instructions as actually issued rather than as remembered. Configuration is a qualifying act in its own right.
• What you rejected. Discarded outputs are the proof of selection, and selection is what carries the second ground. Forty generations and a note of why one was chosen is a better record than the one you kept.
• When. Dates on all of it. The statutory question is whether your acts were directed at this particular work, and a record assembled afterwards is worth a fraction of one made at the time.

This is the material that turns an abstract question about who caused the work into a set of facts, and the order shows it cutting both ways. Thaler won the authorship point because he could point to acts no other person had performed. His file succeeded on the facts and failed only on what he insisted on declaring.

Which is the last practical lesson, and the cheapest. The verified Statement of Particulars is the case. Thaler was offered the chance to name himself as author and refused, and the Registrar held he could not enter particulars found to be legally incorrect nor compel an applicant to register on a basis he had declined to adopt. If you want an accurate record of how a work was made without conceding it away, the order leaves the door open: paragraph 183 expressly permits an applicant to seek an entry identifying the system solely as the technology through which the work was generated, without attributing authorship or legal personality to it, and says such a request would be considered on its merits. Name yourself in the author column and put the machine in the remarks.²¹

One more thing, because a record you cannot get admitted is not a record. Electronic evidence in India now goes in under section 63 of the Bharatiya Sakshya Adhiniyam, 2023, which replaced section 65B of the Evidence Act in July 2024, and secondary electronic evidence requires a certificate under section 63(4) in the form set out in the Schedule — Part A signed by the party, Part B by an expert, both asking for the hash value of the record. The Supreme Court held the predecessor certificate mandatory in Arjun Panditrao Khotkar.²¹ Two things follow for anybody keeping the material described above. Hash the files as you make them, because the certificate will ask for it and a hash computed years later, from a file that may have been touched since, proves less than one computed on the day. And keep the material on a system you control and can give evidence about, because the certificate asks who was in charge of it. A record that lives only inside a service you cannot produce is the weakest possible form of the thing you went to the trouble of keeping.

Where that leaves you

Your material is probably protectable in principle, which is more than it would be in the United States. Whether any person is its author depends on what a human actually did in relation to that specific work. The further your facts are from someone who supplied the inputs, configured the process and started it for this particular output, the closer you are to a work with no author, and a work with no author is a work with no owner, no term, no registration and no moral rights.

The provider's terms of service will not save you. An assignment carries only what the assignor has, and if the provider is not the author of your particular output it has nothing to assign. Nor should the failure of copyright be treated as the end of the question: section 16, which shuts out every right not granted by the Act, expressly preserves any jurisdiction to restrain a breach of trust or confidence, and unpublished machine output held in confidence, protected by contract, or registered as a design or used as a mark, can be defended on grounds that never ask who the author was.²²

None of it is settled. All of it rests on one administrative order, made by an officer rather than a judge, in a proceeding the applicant chose to lose, construing a borrowed provision that the country which drafted it has resolved to repeal. An appeal from the Registrar now lies to a court rather than a tribunal, the appellate board having been abolished in 2021, and nothing in the order binds one.¹⁷

Thaler went to the Delhi High Court to force India to answer whether a machine could be an author. India said no, and told him he was the author himself. He is one of the few people on earth about whom that is true. For everybody else the order raises a harder question than the one it was asked, and leaves it open: India will protect a book that no person wrote, and has nobody to give it to.

Filed under Cyber Law