The court that reserved judgment when TNPS covered this case in April has now delivered it — in OpenAI’s favour, on almost every point that matters, in the world’s most populous country.
The Verdict, Finally
On April 1 this year, TNPS reported that India’s Delhi High Court had reserved judgment in ANI Media Pvt Ltd v. OpenAI, calling it potentially “the most consequential AI copyright case outside the United States.”
On Friday, July 24, Justice Amit Bansal delivered it. News agency ANI’s request for an interim injunction against OpenAI has been dismissed. The main suit continues, but at the interim stage, the court has found – in a 135-page order following 32 hearings – that training ChatGPT on ANI’s news archive is lawful under Indian copyright law.
The core holding, in the judge’s own words, was that OpenAI’s storage of ANI’s material for training purposes “falls under Section 52(1)(a) of the Copyright Act and therefore does not amount to infringement.” That’s the fair-dealing exception for private use, research, criticism and news reporting – a closed statutory list, not the flexible multi-factor test American courts use.
Bansal went further on the output side too, holding that responses ChatGPT generates via retrieval-augmented generation were not “substantially similar” to ANI’s original reporting, and that ANI had failed to show the model had memorised or regurgitated its work.
Three findings inside that order, one in ANI’s favour, two in OpenA’s favour, warrant close attention:
First, the court rejected ANI’s argument that commercial AI companies should be automatically barred from claiming a fair-dealing defence – Parliament, Bansal noted, chose to limit certain exceptions expressly elsewhere in the Act, and didn’t do so here. That’s a deliberate closing-off of the “OpenAI is a business, not a researcher” argument publishers have leaned on.
Second, and less noticed: ANI actually won on jurisdiction. OpenAI argued Indian courts had no business hearing a claim about training that happens on US servers; Bansal disagreed, reasoning that accepting that logic would let any infringer dodge Indian law by hosting abroad.
That’s hugely important. It’s the thing that keeps this case, and future ones like it, inside Indian courts rather than being jurisdictionally waved away. If we are tracking whether Global South courts can actually assert authority over US AI labs, that finding matters as much as the fair-dealing holding.
Third, the “balance of convenience” reasoning, fell in OpenAI’s favour. Bansal held that an injunction would cause “irreparable injury” not just to OpenAI but to the public – India’s roughly 100 million weekly ChatGPT users among them – and warned that requiring licences from every data source before training would make LLM development “economically unviable.”
That’s a policy judgment as much as a legal one, and one publishers need to heed.
Not the First, Not the Second – the Third
TNPS readers will recall the two American rulings from June last year: Bartz v. Anthropic (Judge William Alsup, June 23, 2025) and Kadrey v. Meta (Judge Vince Chhabria, June 25, 2025), both out of the Northern District of California – same court, different judges, notably different reasoning.
Alsup called training on lawfully acquired books “exceedingly transformative” fair use, while drawing a sharp separate line against Anthropic’s use of pirated copies to build its library.
Chhabria reached the same fair-use outcome for Meta but on narrower, more fact-specific grounds, explicitly warning that his ruling shouldn’t be read as blessing AI training generally – the authors in that case, he said, simply hadn’t proven market harm.
So it’s more accurate to call Delhi the third ruling, not the third court – two came from one district court in California, applying the same US fair-use statute under different judicial temperaments.
What makes Delhi new territory, rather than an echo, is that it reached the same destination through a completely different legal road. Section 52 of India’s Copyright Act is an exhaustive list with no AI carve-out and none of the flexible “transformative purpose” weighing that Section 107 invites in the US.
Bansal had to reason his way into fair dealing rather than balance his way into fair use – and the judgment’s citations of the Google Books litigation and the American rulings suggest he was aware he was building a bridge, not crossing one that already existed.
And that’s the real significance of three rulings hitting home the same way inside thirteen months, across two entirely different legal systems: it starts to look less like coincidence and more like a converging judicial instinct – that training itself is the kind of activity courts are, for now, inclined to protect, regardless of whether their statute uses “fair use” or “fair dealing” language to get there.
Where TNPS Was Already Standing
TNPS flagged the Delhi case’s global stakes back in April, months before the ruling – at a point when general publishing trade press attention was mostly elsewhere.
Legal trade outlets – Mondaq, Bar and Bench, LiveLaw – tracked the case closely on the law, but TNPS was pretty much alone in connecting it to the Bartz/Kadrey pattern unfolding. The India judgement of course has no direct impact on US or UK law, so in that respect this was not a US or UK trade story. But the India judgement on jurisdiction, while against OpenAI, matters to US publishers and AI companies alike.
The Licensing Question, Recalculated
April’s TNPS piece argued that a loss for OpenAI would make licensing “unavoidable and legally enforceable.” That specific mechanism is now off the table – no Indian court is going to compel a licence regime via injunction on this evidence.
But the practical incentive to license hasn’t disappeared; it’s been reshaped. Coverage of the ruling in AI-industry newsletters has already drawn the sharper conclusion: the burden now sits with publishers to prove memorisation or reproduction, not merely that their work was ingested – a considerably harder evidentiary bar to clear than the one ANI approached this case with.
For any Indian newsroom weighing whether to negotiate a licence or hold out for a courtroom win, that shift in the burden of proof is a real loss of leverage, even though the door to a negotiated deal remains open. The Digital News Publishers Association and the Federation of Indian Publishers, both intervenors in this case, now have a much narrower legal argument to build on than they did in April.
The Hallucination Claim: A Separate, and Possibly Dated, Question
ANI’s suit always contained two distinct claims, and Friday’s ruling only resolved one of them. The training-data claim – was storing and learning from ANI’s articles infringement – is settled, at the interim stage, in OpenAI’s favour.
The second claim, that ChatGPT fabricated stories and wrongly attributed them to ANI, is untouched by this order and remains for trial.
Here the timeline is important. ANI filed in November 2024, and its hallucination complaint was already part of the original filing – meaning any fabricated-attribution incidents ANI can point to necessarily involve a ChatGPT model generation from that period or earlier, not the current one.
OpenAI itself told ANI, in correspondence disclosed during the case, that it had already placed ani.in on an internal blocklist and stopped using the site in training since around that time. Two full model generations and multiple safety and retrieval-augmentation iterations have shipped since.
That doesn’t make the underlying legal claim disappear – a wrong is still a wrong regardless of which model version committed it, and OpenAI can still be held liable for output from a retired model if the claim is proven.
But as a live product-safety concern about today’s ChatGPT, it’s closer to an inherited liability from an earlier era of the technology than a warning about current behaviour. Any trial coverage of that half of the case should make that distinction explicit rather than let a two-year-old failure mode be read as evidence of a present one.
The Tally From The Beach
Three rulings, two legal systems, one direction of travel. That’s a harder pattern to dismiss than any single decision would be, and it’s the story for global publishing, wherever we are.
That’s the thing: this isn’t an American doctrine spreading outward, it’s multiple independent judicial traditions arriving at a similar accommodation with the same technology, for overlapping but not identical reasons.
For publishers – in India and everywhere else watching this case as the Global South’s test of the question – the realistic path to being paid for AI training data was never going to run through injunctions. It runs through licensing markets maturing faster than litigation does.
Friday’s ruling didn’t close that path. It just made clear, again, that courts aren’t going to build it for them. Publishers need to be looking licencing options now, while they still have leverage.
This post first appeared in the TNPS LinkedIn Analysis Newsletter.
