Delhi High Court Calls AI Training Fair Dealing, For Now

The Delhi High Court has become the third major jurisdiction to put a judicial opinion on the record about AI training on copyrighted news content — and the framing matters more than the headline. On July 24, 2026, Justice Amit Bansal refused ANI Media’s application for an interim injunction against OpenAI, holding on a prima facie basis that using ANI’s content to train large language models can fall within the fair-dealing exception in Section 52(1)(a)(i) of India’s Copyright Act, 1957.

Almost every summary of the ruling has flattened it into a verdict. It is not one. This was an application for a temporary injunction, decided on a prima facie standard, in a suit that is still alive. Business Standard reported that Justice Bansal sent the case back to the Delhi High Court’s Roster Bench to determine which bench will hear the full suit — meaning the merits are still ahead, with discovery, evidence, and a trial record that does not exist yet. The court itself said its findings would not prejudice the final adjudication of the suit.

This guide walks through what the court actually held, what the prima facie standard does and does not settle, how the reasoning compares with the two US rulings that came before it and with the EU’s structurally different opt-out model, and what content owners can sensibly take from an interim order. It is commentary for publishers and marketers, not legal advice — the sourcing behind each claim is flagged as we go.

The decision came from a single-judge bench of the Delhi High Court on Friday, July 24, 2026, in ANI Media (P) Ltd. v. Open AI OpCo LLC. Legal-tech trade outlet thelegalwire.ai reports the case as CS(COMM) 1028 of 2024 with neutral citation 2026 DHC 5900. The case number is corroborated by the business daily Business Standard; the neutral citation rests on thelegalwire.ai alone. We have not independently checked either identifier against a cause list or the judgment cover page, so treat them as reported rather than verified.

What ANI asked for was a temporary injunction — an order restraining OpenAI while the suit proceeded. That application was dismissed. Business Standard reported the court found ANI had failed to establish a prima facie case of copyright infringement as to either the training process itself or the responses ChatGPT generates. On that prima facie footing, the court held that OpenAI’s use of ANI’s copyrighted content to train its models fell within private or personal use, including research, under Section 52(1)(a)(i), and could therefore constitute fair dealing.

The qualifier is not decorative. It is the operative limit on everything above. A prima facie finding is a first-look assessment made on affidavits and submissions, without the evidentiary record a trial produces. The same court that made it also directed the suit onward, which is the clearest possible signal that the question is still open.

An interim injunction application is a triage exercise. The court asks whether the applicant has shown enough, on a first look, to justify restraining the other side before the case is tried. It weighs that against the balance of convenience and the risk of irreparable harm. What it does not do is resolve the underlying dispute — that is what the trial is for.

So a refusal has a narrow meaning: ANI did not clear the threshold for emergency relief on the material before the court at that moment. It does not mean ANI loses. It does not mean training on news archives has been declared lawful in India. It means one applicant, on one record, at one procedural stage, did not persuade one judge that a restraint was warranted while the case runs.

This distinction gets lost because interim orders in high-profile technology disputes are the only judicial text available for months at a time, so they get read as verdicts. The honest reading is narrower and more useful: a well-reasoned data point has entered the record, and it will shape argument in the next matter without binding the outcome of this one.

What is still undecided

• Whether the fair-dealing finding survives trial. A prima facie view formed on submissions can shift once discovery produces evidence about what was ingested, how, and with what effect on the market for the original.
• Whether training requires a licence in India. Nothing in an interim refusal settles a licensing regime. That question sits with the trial court, appellate courts, and ultimately Parliament.
• How the framework applies to other content types. This dispute concerned news reporting and alleged reproduction of news text. A different corpus, a different ingestion method, and a different evidence record can produce a different answer.
• What remedies would follow if ANI later succeeds. Damages, accounts of profits, and injunctive relief at final hearing are all untouched by a refusal of interim relief.

India does not have an open-ended fair use doctrine. It has fair dealing: a closed list of permitted purposes in Section 52 of the Copyright Act, 1957. Widely reported compilations of the statutory text describe Section 52(1)(a) as excluding from infringement a fair dealing with any work, other than a computer programme, for purposes including private or personal use and research — we were unable to pull that wording from a government-owned source directly, so we paraphrase rather than quote it.

That closed-list structure is why the drafting of the clause carried so much weight here. According to Business Standard’s report, the court rejected ANI’s argument that commercial entities are automatically excluded from the Section 52 defence, observing that Parliament expressly limited some copyright exceptions to non-commercial use but did not do so in Section 52(1)(a). SCC Online’s summary frames the same point as a deliberate legislative omission of a non-commercial restriction in that provision.

The harder move was semantic. A term written in 1957 has to be stretched to cover gradient descent over a news archive, or it does not reach the conduct at all. SCC Online’s summary describes the court applying an updating construction — reading research in light of present-day technology rather than confining it to human research as understood when the Act was passed. We are paraphrasing a secondary summary of judicial reasoning here, not quoting the judgment, and the same caveat applies to the three-part framing described below.

Per SCC Online’s summary, the court worked through three questions: whether the use was confined to training the model, whether it caused economic competition or commercial prejudice to ANI, and whether ChatGPT’s functions served a broader public interest. All three were reported to favour OpenAI on the material before the court. Business Standard also quoted the judgment as reasoning that LLM development depends on access to information in the public domain and that requiring licences from multiple sources would make developing an LLM economically unviable — a line we are describing rather than quoting, because we have not seen it in the judgment itself.

Strip out the legal machinery and the structure is familiar. Every jurisdiction wrestling with this question is asking some version of the same three things: is the copying instrumental rather than expressive, does it substitute for the original in the market, and is there a public-facing benefit that justifies the intrusion. The labels differ — transformativeness and market harm in the US, purpose and prejudice in India — but the analytical shape converges.

ANI’s case had two halves, and the second one is the half most publishers care about. Training is abstract; what a model says about your reporting is not. Secondary reporting on the original filing describes ANI alleging that ChatGPT could reproduce parts of its articles verbatim and, in some instances, generated fabricated stories falsely attributed to ANI. Those allegations reach us through aggregated coverage of the November 2024 filing rather than through a primary document, so we present them as alleged and secondary-sourced.

On the output side, Business Standard reported that the court held — again at the prima facie stage — that responses produced through retrieval-augmented generation did not infringe under Section 51 because they were not substantially similar to ANI’s original literary works. The report also indicates ANI failed to establish memorisation or regurgitation of its content, and failed to show substantial similarity or market substitution.

That is a distinction worth sitting with, because it is where the practical exposure lives for any content business. A model that retrieves, summarises, and attributes is doing something different from a model that emits your paragraphs. The first is hard to attack on similarity grounds and is better fought on market-substitution evidence. The second is the strongest claim a publisher can bring — and it is exactly the claim that requires you to have preserved the evidence.

The one part of the order that cut against OpenAI is the part that will travel furthest. Business Standard reported that the court rejected OpenAI’s argument that Indian courts could not examine the training claim because its servers sit in the United States. The reasoning, as reported, was blunt: accepting that position would let infringers evade Indian law simply by locating servers abroad.

For anyone building AI products with an international user base, that is the more consequential holding of the two. A prima facie fair-dealing finding is provisional and confined to this record. A court asserting that it can examine training conduct despite foreign infrastructure is a structural position that other Indian benches can adopt without needing to agree on fair dealing at all.

It also reframes what a favourable interim order is worth to a model developer. Winning the injunction stage while conceding that the forum can hear the claim means the dispute stays in India, on Indian statutory ground, through trial. That is a slower and more expensive posture than having the claim dismissed for want of jurisdiction — and it is the posture the ruling leaves in place.

Coverage of AI training law tends to stay inside one regulatory silo. Indian legal press covers the Delhi ruling without reference to the US cases; US coverage of Bartz and Kadrey rarely mentions India; EU AI Act commentary touches neither. Put the three side by side and the interesting difference is not the outcome — it is the machinery. Two of these systems decide the question after the fact, in court, one defendant at a time. The third decides it in advance, by registry.

The table below is our own synthesis across the three regimes. It compares mechanism, decision-maker, status as of July 27, 2026, and what the rightsholder actually has to do to be protected — which is where the systems diverge most sharply and where publishers consistently underestimate the workload.

The asymmetry in the last column is the point. In India and the United States, protection is something you litigate for: expensive, slow, uncertain per defendant, and — as ANI’s twenty-month path to an interim order shows — capable of consuming years before producing anything final. In the EU, protection is something you configure: cheap to invoke, uniform in application, but entirely dependent on the publisher having implemented a machine-readable reservation before anyone crawled the site.

That inverts the usual instinct. Publishers reflexively treat the litigation jurisdictions as the strong ones because the remedies are larger. In practice the opt-out regime is the one where a small publisher can act unilaterally and immediately, while the litigation regimes reward whoever can fund a multi-year case. If you operate across all three, the EU configuration work is the cheapest protective step available — and it is the one most often left undone. Anthropic’s own resolution of the US book-piracy claims, which we covered in our breakdown of the $1.5 billion copyright settlement, is the counter-example: a payout, but only after litigation most rightsholders could never sustain alone.

For the EU side of the picture, including the transparency duties that have applied to general-purpose model providers since August 2025 and what arrives next, see our agency checklist for the EU AI Act transparency obligations.

The commentary quoted around the ruling is notably unenthusiastic about treating it as settled. Ronil Goger, managing partner at Blaze Legal, told Business Standard the ruling is India’s first substantive judicial engagement with AI and copyright but leaves fundamental questions unresolved — among them whether AI training requires licences, how fair dealing applies to machine learning, and how creators should be compensated. That is three of the four questions a publisher would actually want answered, still open.

The tension the suit sits inside has not moved either. Publishers have been trying to convert traffic loss and content ingestion into either payment or access control, with mixed results across search, licensing deals, and crawler policy — the same terrain we mapped in our look at publishers blocking Google Search and the Reddit deal. An interim order that declines to restrain one developer does not change the economics driving any of that.

Our own reading, offered as commentary rather than advice: the most durable signal in this order is not the fair-dealing framing at all. It is the pairing of a jurisdictional holding that keeps the claim in India with an evidentiary standard that ANI could not meet. Together those say the forum is available but the burden is real. For content owners, that shifts the useful work upstream — from filing faster to documenting better, and from arguing principle to producing market-substitution evidence a court can weigh.

Three separate voices quoted in Business Standard’s coverage land in roughly the same place without stating it as a shared conclusion. Ankit Sahni frames the order as influential but interim. Ronil Goger catalogues what remains unresolved. Jameela Sahiba of The Dialogue argues content companies need a broader strategy that does not rely exclusively on litigation — pointing to commercial licensing models for AI use, stronger contractual restrictions on access, technical measures to manage automated scraping and ingestion, and engagement with the legislative process. Assembled, that is a playbook nobody published as one.

The technical leg of it is the one most within a publisher’s immediate control and the one most often deferred. Crawler policy, robots directives, llms.txt conventions, and edge-level access controls all function regardless of how any court eventually rules, and they are what make a machine-readable rights reservation meaningful in the EU. We keep the decision framework current in our AI crawler access-control decision matrix, which is the practical companion to everything in this section.

The commercial leg deserves more attention than it gets. Licensing reframes the relationship from one where you are trying to stop ingestion to one where you are pricing it — and it produces a contract with terms you wrote, rather than a judgment written by someone else. Whether that is worth doing depends on the value and uniqueness of your archive, which is a strategy question rather than a legal one. Structuring content so it retains value under AI retrieval is the same problem we work on inside content engine engagements and, on the discovery side, through agentic SEO.

The next twelve months should produce more signal than the last twenty. The immediate item is procedural: which bench takes the ANI suit, and how quickly it moves to a hearing on the merits. A trial record — discovery into what was ingested, in what volume, and with what measurable effect on ANI’s market — is the first thing that could genuinely test the prima facie framing rather than restate it.

The second thing to watch is whether the reasoning travels. Interim orders in Indian technology disputes tend to be cited heavily in subsequent applications, and a 135-page analytical framework, as Business Standard described this one, is unusually citable. If other benches adopt the updating-construction approach to Section 52, the practical effect could outrun the order’s formal weight well before any appellate court weighs in.

Our projection, held loosely: the durable divergence will not be between countries that permit training and countries that forbid it. Every regime surveyed here is converging on permitting it in some form. The divergence will be in what a rightsholder must do to exclude themselves and what evidence they must bring to be paid. The EU has already answered both questions prospectively. India and the United States are answering them one lawsuit at a time, which means the practical rules will be set by whoever can afford to litigate — and that is a distributional outcome, not a doctrinal one. Building an operating posture that survives all three answers is the realistic goal, and it is the kind of cross-jurisdiction planning we handle inside AI and digital transformation programs.