Month: July 2026

Finally, the cat is out of the bag – ANI v. Open AI judgment pronounced!

Let me start with a disclosure. I represented one of the intervenors in this matter, Flux Labs, an AI enabled commerce start up.

This post however is on principles, and the jurisprudential significance of the decision in  ANI Media v. OpenAI (2026 DHC 5900), because the principles are what will outlive the parties.

1. The first thing the court does is close a door that should have been closed long ago. ANI leaned on Walter v. Lane, the House of Lords decision from 1900 that gave a reporter copyright in his transcription of a speech. The court holds that Walter v. Lane was decided under an English statute, originality under which was determined by a “sweat of the brow” standard, and that India has rejected the sweat of the brow standard on which it rests. What governs in India is the “skill and judgment” test from Eastern Book Company v. D.B. Modak, which people keep confusing with the American modicum of creativity standard from Feist. They are different tests. And the skill and judgment involved must be more than trivial. Effort alone, however honest, earns no copyright here.

2. From there the court goes somewhere that matters enormously for anyone who works with facts. It applies the Division Bench in B.D. Bhandari and the merger doctrine applied there. Where a fact can be expressed in only a limited number of ways, the fact and its expression merge, they become so intrinsically connected that you cannot pull one apart from the other, and the expression itself falls out of protection. Copyright lives only in the form and manner of expression (or what one may think of as the packaging of expression), never in the underlying facts, however hard someone worked to gather them. And because news is exactly the kind of writing where the ways of expressing an event are limited, the court holds that the threshold for showing substantial similarity in a news article is higher than for other work. This is the thin protection doctrine.

3. On how you compare two works, the court restates R.G. Anand with some force. You compare the works as a whole. You do not dissect the plaintiff’s work into fragments and hold each fragment against the defendant’s text (in other words – the work as an atom). Infringement is made out only where the defendant’s work is a transparent rephrasing of the original, or takes its form, manner, arrangement and expression with minor changes. Applied to the record, this sank ANI’s output claim.

    Now to the part I care about most, and the part where I both agree and disagree.

    4. The court holds that the reproduction right in Section 14(a)(i) includes storing the work in any medium by electronic means, that the provision makes no distinction between temporary and permanent storage, and that the purpose behind the storage is irrelevant at the infringement stage. So, storage per se sits inside the copyright holder’s rights. I will be honest, I think this is wrong, and I have argued the contrary at length in my own writing. Reproduction, to my mind, is concerned with copying that carries the work to someone. Storage that exposes the work to nobody, storage that no human being ever reads, is non-expressive use and ought to fall outside the right itself, and if that is so the court never needed to travel to Section 52 at all. But I will also be honest about this, the destination is the same either way, because everything the court goes on to say about private use turns on the very fact that nobody is exposed to the stored work. My discomfort is with the route, since Section 52 is a more fluid instrument than the boundary of the right, and fluid instruments are easier to bend (or as colloquially called – indeterminate or conceptually unclear) in the next case. A difference of method, then.

    5. What the court does then with Section 52, however, is, to my mind, the heart of the judgment. It holds that Section 52 is not an exception. It is not a proviso hanging off Section 51. Section 52 independently defines the rights and privileges of users of copyrighted works, it is an integral part of the Act, and it must receive a broad and liberal interpretation. This is now only the second declaration of its kind by an Indian court, after the single judge of the Delhi High Court in Rameshwari Photocopy. I cannot overstate what this means. Once fair dealing is a user’s right rather than a grudging carve out, every future fight about education, research, access to information and access to technology starts from different ground. Think about the implications for access to paywalled articles!!? How do I exercise my (now!) right to research a paper behind a paywall? Scihub is gone!! Public libraries (with legal deposit requirement) are dysfunctional! So what? How to?

    6. Working through Section 52(1)(a) itself, the court applies a two part examination, a purpose test (whether the purpose falls within one of the subjects stated in Section 52(1)(a) and a fairness test (or the test of fair dealing), and under the purpose head it does three things that carry massive implication.

    (i) It holds that commercial use is not excluded from private or personal use including research. Parliament knew how to say non-commercial when it wanted to, and it did say so in several neighboring clauses of Section 52, and it chose not to say so here. A journalist reports the news for money and still enjoys the fair dealing defense for reporting. A reviewer is paid and still enjoys the defense for review. The court refuses to read into sub-clause (a) of Section 52(1) a limitation the legislature left out. It draws support from the Notes on Clauses to the 1994 amendment, which said in so many words that an unduly narrow reading of private study would result in harassment to the public.

    (ii) The Court then holds that private use is not confined to human beings but extends to juristic entities. If private and personal meant the same thing, one of the two words would be doing nothing, and the 2012 amendment deliberately put both into the provision. Private reaches a closed sphere, and a company can occupy a closed sphere. The training data here is accessible only to the models themselves, no human being can reach it, access it or download it, and the court holds the use is therefore purely private. B. Malini Mallya, where the Supreme Court extended Section 52(1)(a)(i) to a dance performance at an educational institution, supports the expansive reading. Interestingly the court held the use by Open AI to be private because no one beyond the organization would be exposed to the contents of the work being used or copied – essentially non-expressive use. Thus, instead of reading non-expressive use as an inherent limitation to the scope of rights, the Court read it as a part of private use.

    (iii). And then the boldest move (and honestly vulnerable from a literal interpretation point of view – which I hope is not applied), the court reads research separately, refusing to treat it as swallowed by private use, and gives it what the interpretation using the doctrine of updating construction. According to the Court, Research is no longer confined to humans, though it remains at the behest of humans and for the benefit of humans, and thus research ought to be interpreted more liberally to include use for training LLMs. It will be interesting to see how this reading of the research provision is interpreted when it comes to accessing paywalled articles for the purposes of both human and machine learning.

    7. Importantly, ANI argued that the fair dealing defense is available only where the stored copy is itself a non-infringing copy. The court reads the Explanation to Section 52(1)(a) and holds that the non-infringing copy condition attaches only to the incidental storage of a computer programme, and to no other class of work. Otherwise, the court says, a researcher who photocopies a page in a library would be protected while the same researcher who scans and stores the same page would not be, and that distinction between paper and electronic storage cannot be the intent of the provision.

    8. On fairness, the court does something I found genuinely refreshing. It declines to adopt the American four factor test, notes that Indian courts have never actually settled on a single test, and builds three factors for the case at hand. It asks (i) whether the use stays confined to training, (ii) whether it sets up economic competition that prejudices ANI’s legitimate interests and causes it actual or potential damage, and (iii) whether the functions performed serve the overall public interest. On the first, nothing on record showed OpenAI using ANI’s works for anything except training. On the second, ChatGPT is a general purpose tool, and ANI is a news business, their markets are different, ChatGPT’s responses do not substitute for ANI’s syndication feed, and ANI placed nothing on record beyond bare averments to show lost subscribers or lost revenue. On the third, the court found the public benefits of the technology considerable, in research, education, translation, accessibility and access to information.

    9. So, this leaves us with a judgment that provides a lot of “hope” at least on the jurisprudential side. Is this a tangible outcome of a multi-billionaire defendant, which has until now (maybe except Radio), been a rarity in Copyright disputes?

    P.S: What happens to the Part 1 of the DPIIT Report, the entire premise of which is, now, entirely shaken?

    Submissions of Flux AI Labs in the matter (prepared by Shivani Vij, Shrutanjaya Bharadwaj and myself)

    A compendium of extracts from precedents (bookmarked for anyone interested to look at) on what all we relied upon in principle.

    My writing on AI and copyright: