They shouldn't.

Nobody in that room asks the obvious question: what if the code was written by GitHub Copilot? What if the marketing copy came from ChatGPT, or the product designs from Midjourney? If the answer is yes, the assignment clause may have assigned nothing at all. The chain of title on which the whole valuation depends could be legally hollow, and no one in the room would know it.

That's not a hypothetical. It's the current state of most standard corporate agreements in India, and the profession still hasn't caught up. The problem becomes clearer when you look at the law those agreements rely on.

I. A Statute Built for Human Authors

The Copyright Act, 1957, was built on an assumption it never had to spell out: a human made this. Section 2(d) defines "author" separately for each category of work, and every version points to a person. For literary works, whoever creates them. For musical works, the composer. For artistic works, the artist. Section 17 then gives the author first ownership, with the employer stepping in for anything made in the course of employment. That's the provision every IP assignment clause in an Indian employment contract is quietly built on.

Once the "author" is an AI tool, that logic has nowhere to land.

There's one provision that might still cover this, but it's a stretch. Section 2(d)(vi) deals with what the Act calls "computer-generated works," and it says the author is "the person who causes the work to be created." That language is from 1957, long before anyone imagined generative AI. What Parliament might have had in mind was a spreadsheet, a CAD drawing, or a program that follows a fixed set of rules and requires a person to set the inputs. Whether that same idea stretches to cover something like ChatGPT or Claude is genuinely unclear. The uncertainty matters because it leads directly to the next question: who, if anyone, caused the work?

Here's where it gets messy. If someone types one quick prompt and just takes whatever comes out, did they really "cause" that output, in the sense the law means? Probably not, but no court has ever said so. If someone spends hours prompting, rejecting drafts, and reshaping the result into something specific, their claim to have caused it looks much stronger. Either way, no Indian court has actually ruled on this yet, for either kind of user.

That gap matters more than it sounds like it should. If no one counts as the "author" under the Act, there's no copyright in the work at all. And if there's no copyright, there's nothing for an IP assignment clause to transfer. It doesn't matter how well the clause is drafted. It can't assign a right that was never created in the first place.

II. What the Courts Are Working Through

The most useful current source for understanding the current state of Indian law is ANI Media Pvt. Ltd. v. OpenAI Inc. [CS(COMM) 1028/2024, Delhi High Court]. It started small: ANI accused OpenAI of scraping its news articles to train ChatGPT without permission. But the case has grown well past that original complaint. Justice Amit Bansal now has to decide two things. First, does storing copyrighted material to train an AI model count as infringement under Section 14? Second, even if it does, is it protected anyway under the fair dealing exception in Section 52? Those questions sit at the center of the dispute, and neither has yet been answered by any court in India.

The court brought in two outside experts, Dr. Arul George Scaria and Adarsh Ramanujan, to help work through it, and they disagreed with each other. One argued that storing text for training purposes falls within the fair dealing carve-out in Section 52(1)(a). The other argued that copying at this scale, for a commercial product, was never what fair dealing was meant to protect.

Hearings finally wrapped on 27 March 2026, after 32 separate sittings. The judgment is reserved now, meaning the court has heard everything and is just deciding. It could come down any day. That ruling will be the first real judicial word in India on how copyright law treats AI training, and it will also shape how courts think about AI-generated output.

A second matter is putting forth the human-authorship question from a different angle. Back in April 2026, the Delhi High Court gave the Copyright Office eight weeks to decide Stephen Thaler's long-pending application to register "A Recent Entrance to Paradise," an artwork his DABUS system produced, a case that had been sitting untouched since 2022. That window closed in late June. No order has surfaced yet. Whenever it does land, it'll be one of the first direct tests in India of the human-authorship requirement. The Copyright Office hasn't been friendly to these claims before: in 2020, in the Ankit Sahni "RAGHAV" matter, it registered an artwork listing an AI tool as co-author, then tried to withdraw the registration once the implications sank in.

There's movement on the legislative side too. The Ministry of Commerce and Industry set up an eight-member expert panel in May 2025 to assess whether the Copyright Act still works in a generative AI world. Word is the panel is looking at a new Chapter XII-A for AI-generated works, possibly borrowing from Section 9(3) of the UK's Copyright, Designs and Patents Act, 1988, which hands authorship to "the person by whom the arrangements necessary for the creation of the work are undertaken." Sensible enough. But it means nothing to a client whose deal closes next month. So the draft reform remains separate from the immediate contract problem.

III. Where Standard Agreements Fall Apart

Picture a tech company where developers use GitHub Copilot day-to-day. Their employment contracts carry the usual language: all IP conceived or developed during employment vests absolutely in the company. An investor asks for ownership warranties. The company signs off without thinking twice. Nobody brings up Copilot. That silence is where the contract starts to break down.

The problem sits at two levels. If the AI-generated code doesn't clear the Section 2(d)(vi) causation bar, say the developer's input was a vague comment and a stub function, there may be no copyright at all, and nothing to vest. And even where copyright does exist, who holds it isn't obvious: the developer, the AI vendor under its own terms of service, or some split between them. The employment contract wasn't written for any of these actors. It assumes a human made the thing and a human owns what they made. That assumption is exactly what fails once AI enters the chain.

Consulting agreements, joint development deals, technology services contracts, all the same story. The clause hands over the creator's rights. When the creator is a machine, there's nothing there to hand over.

IV. What a Properly Drafted Clause Needs to Do

Banning AI tools isn't the answer. It's not realistic, and no court would enforce it anyway. The real fix is drafting agreements that hold up whether or not the output ever clears the copyright bar in the first place. That means rethinking four specific parts of the standard clause before the agreement has to deal with them.

a) Redefine the categories. Most IP clauses list copyright, patents, trademarks, and trade secrets, and stop there, never asking how the work actually came into being. Split output into two buckets instead: "AI-Assisted Work," where a person drove real creative decisions, and the tool did the typing, and "AI-Generated Work," where the human barely touched it. Which bucket a deliverable lands in should decide how it's owned, whether it's worth registering, and how much warranty risk it carries.

b) Require contemporaneous documentation. Any deliverable that's AI-assisted, or could plausibly be argued that way later, needs a paper trail while it's being made, not reconstructed after the fact: what was prompted, how many passes it took, what a human actually decided to keep or change. Skip this, and the Section 2(d)(vi) causation argument has nothing to stand on the moment someone tests it in court. India's 2025 AI Governance Guidelines already nudge developers toward exactly this kind of record-keeping; a contract just needs to make that nudge mandatory instead of aspirational.

c) Add a fallback ownership clause. Almost nobody has one. If a court finds a deliverable doesn't qualify for IP protection, the assignment clause has nothing to operate on, and the company that paid for the work is left with nothing while the person who built it walks away holding it. A fallback closes that hole directly: the creating party holds the deliverable in trust, doesn't exploit it, and hands over its practical benefit regardless of what the copyright position turns out to be. Where the statute ends, the contract can continue.

d) Tie disclosure to a narrower warranty. Every major AI platform, OpenAI, Adobe, Microsoft, buries its own IP terms inside its terms of service, and those terms reach straight into your chain of title without anyone noticing. A company that can't say which tools its people used has no way to price that risk. The standard non-infringement warranty needs a carve-out for anything arising purely from AI tool use, because with the infringement question still sitting unresolved before the Delhi High Court in ANI Media Pvt. Ltd. v. OpenAI Inc., giving that warranty without qualification is a bet no client should be making blind.

Get these four in place, and the clause stops depending on how a court eventually rules on human authorship. It works either way. That is what makes the drafting resilient.

The Copyright Act, 1957, runs on a theory of authorship: creativity comes from a human mind, deserves protection because it reflects human effort, and can be owned and passed on because it traces back to a person. Generative AI hasn't broken that theory. It's just moved faster than it. The ANI case will eventually settle some of this. The expert panel will eventually propose amendments. But the agreements being signed this month will govern disputes that come up long before either of those lands.

The clause that quietly assumes a human wrote it and hands over that human's rights needs to stop being invisible. Not because the statute requires it yet. Because the exposure is already sitting there, and the lawyers who deal with it now will be the ones whose clients are covered when the law finally catches up.

Author is an Advocate based in Surat. Views are personal.