Two decisions landed in India this year, months apart, in different forums, on facts that have nothing to do with each other. Nobody has read them together. Read together, they describe the position an Indian writer now occupies with some precision, and it is a narrower position than it was eighteen months ago.
The first is the Delhi High Court's judgment in the ANI matter against OpenAI, delivered by Justice Amit Bansal on 24 July 2026 after thirty-two hearings. The second is a decision of the Indian Copyright Office on whether a work generated by an algorithm can be registered, and who may be named as its author.
Take them one at a time, then put them side by side.
What the court actually held
The Delhi High Court held three things, and it is worth being exact about each, because the reporting has blurred them.
Storage of the copied material was held to fall within the fair dealing exception in Section 52(1)(a). Training on that material was held not to infringe. And the outputs were held not to be substantially similar to the originals.
Then a fourth holding, which is the one that will matter longest: the court declined to read a commercial-use bar into Section 52. It was invited to say that fair dealing is unavailable to a party making money from the use, and it refused.
The practical consequence is a shift in who has to prove what. The burden now sits with the claimant to demonstrate memorisation or reproduction — to show that the model has retained and can emit the protected expression itself, not merely that it was trained on it. That is an evidentiary burden of a completely different order. Proving your book was in a training corpus is difficult. Proving the model has memorised your sentences and will reproduce them is difficult in a way that requires access, expertise and money that no individual author has.
So the first door closes. An author cannot stop the training.
What the Copyright Office decided
The second decision runs the other way and arrives at a similarly closed door.
The Copyright Office considered an algorithmically generated work and found — this is the interesting half — that such a work can meet the originality threshold. It did not hold that machine-made output is automatically outside copyright. That was the outcome many expected and it is not what happened.
The application was nonetheless rejected, on a different ground: it named the AI itself as the author. Authorship, on this reasoning, attaches to a person. A system is not a person, and a work whose declared author is a system has no author the register can record.
I am deliberately not giving you a date or a file number for this one. Both are circulating and I have not been able to verify either against a primary source, so I am not going to repeat them and add to the noise. The holding is what matters and the holding is sound.
The two doors, together
Here is the pairing nobody in India has written, and it is the whole point of this piece.
An author cannot prevent a model being trained on their work — the burden of proof has moved somewhere they cannot reach. And an author cannot claim a work the model produces — because authorship requires a person, and if the person's contribution was a prompt, the register is unlikely to accept that as authorship of the resulting text.
Both propositions are individually defensible. The court's reasoning on fair dealing is careful and the Copyright Office's reasoning on authorship is orthodox and matches what most jurisdictions have concluded. Neither is a scandal.
But stand them next to each other and you see the shape of the space a working writer now occupies. The input side is open to the machine and the output side is closed to the writer. Whatever value accrues from the combination accrues to whoever owns the model.
That is not a conspiracy. It is just what happens when two sensible decisions from two different forums are laid over one another, and nobody is looking at the overlap because it is nobody's job to look at the overlap.
What this actually changes for someone writing books in India
Not as much as the alarm suggests, and more than the reassurance suggests.
It does not change your copyright in what you have written. Your manuscript is yours. Someone who copies your chapters and sells them is still infringing and you can still stop them. Nothing in either decision touches ordinary infringement.
What it changes is the theory some writers were quietly relying on — that training was going to turn out to be a licensable act, that a market would emerge in which Indian publishers and authors were paid for corpus inclusion, and that the courts would build the tollbooth. In India, on this judgment, the courts have declined to build it. If a licensing market appears it will now have to be built commercially, by publishers negotiating from a weaker position than they had in July.
It fits a pattern I have written about elsewhere. India's AI Governance Guidelines create no new statute and no new regulator; they bind through procurement clauses and sectoral circulars instead. The copyright question has now been settled the same quiet way — by interpretation of an existing provision, not by a new one built for the purpose. India is regulating AI, repeatedly, without appearing to.
And it changes what is worth spending your own effort on. Watermarking your prose against training is not a strategy. Registering an AI as co-author of anything is not a strategy. The things that still work are the unglamorous ones: your name, your particular way of seeing a subject, the readers who came for you specifically, and direct channels to them that do not run through a platform's recommendation engine.
Which is a slightly bleak way of arriving at an ordinary conclusion. The defensible asset was never the text as a commodity. It was the relationship with the person reading it.
The one lever that survives
There is a third thing neither decision touched, and it is the only one an individual writer still controls: the contract.
Nothing in the judgment stops a publisher and an author agreeing, between themselves, what may be done with a manuscript. If the courts will not create a licensing market by implication, it can still be created by express term. An Indian author signing a publishing agreement this year should be reading the grant-of-rights clause for language about machine learning, text and data mining, and derivative or downstream training use — and should notice that most standard Indian contracts are silent, which after July means the silence resolves in favour of whoever wants the corpus.
Ask for the clause. Ask for it to be carved out, or reserved, or licensed separately for a fee. You may not get it. But the ask costs nothing, and a term you negotiated is worth more now than a doctrine you were hoping a court would find for you.
The number worth holding on to
I wrote recently about what Amazon actually pays an Indian author on a KDP sale — the royalty line nobody reads, the one that survives every conversation about the size and growth of the Indian book market. That piece and this one are the same argument approached from two sides. I write both non-fiction and fiction under Indian publishing contracts myself, so the grant-of-rights clause above is not a hypothetical for me either.
One says: the market is enormous and growing and almost none of the published economics describes what reaches the writer.
This one says: the legal architecture around the writer's raw material has just been settled in a direction that does not favour the writer either.
Put them together and the conclusion is not despair, it is redirection. Stop optimising for a share of somebody else's platform economics. Build the thing that is yours.
Read next: The Royalty Line Nobody Reads: What Amazon Actually Pays an Indian Author — the same argument from the commercial side.
