What the court did on 15 September

The bench of Justices Avneesh Jhingan and Manmeet Pritam Singh Arora directed OpenAI to file its response to ANI Media's appeal, which challenges the refusal of an interim injunction restraining the company from using ANI's content to train and operate ChatGPT. The court also indicated it would take up the Broadband India Forum's application to intervene at the next hearing. The matter was listed for 5 December 2026.

Procedurally this is a modest step: a notice, not a stay. But its significance is in what it keeps alive. The July ruling had effectively removed the interim threat to OpenAI's Indian operations and set a reference point that every AI developer in the region began citing. An appellate bench willing to hear the challenge — and to entertain industry intervention — signals that the reference point is not settled.

How the case got here

ANI, one of India's largest news agencies, sued OpenAI in the Delhi High Court in November 2024, alleging that its articles had been used without licence to train ChatGPT and were being reproduced in outputs. The litigation has run for almost two years and has drawn in a striking range of participants:

  • November 2024: The court appointed two amici curiae to assist it — an early acknowledgement that the technical questions exceeded the ordinary adversarial record. OpenAI's central position was stated at the outset: copyright protects expression, not ideas or facts.
  • January 2025: OpenAI told the court that an order to delete training data would conflict with its obligations under US law, raising the cross-border enforcement problem that recurs in every one of these cases.
  • March 2025: ANI alleged that OpenAI continued to access its content through subscribers notwithstanding assurances, pointing to prompts that produced ANI headlines. One amicus argued that the storage of copyrighted material is itself infringement.
  • April–May 2025: ANI sought an injunction covering scraping of subscriber-facing content; OpenAI countered that ChatGPT's search function sends traffic to publishers.
  • August 2025: The Digital News Publishers Association — whose members include the country's largest newspaper groups — argued that OpenAI scrapes, stores and reproduces news content without licence or attribution.
  • October 2025: The Broadband India Forum, representing technology and telecom interests, argued that large language models do not infringe copyright at all.
  • July 2026: The single judge declined interim relief, finding that ANI had not put forward the evidence needed to establish infringement at that stage, with the reasoning running in favour of fair dealing.

The legal core: "fair dealing" is not "fair use"

The distinction ANI presses on appeal is the whole case, and it is one Turkish lawyers will recognise immediately, because Turkish copyright law is built the same way.

American fair use is an open-ended standard. A court weighs four factors — purpose and character of the use, nature of the work, amount taken, and market effect — and can find almost any use fair if the balance falls that way. It was under this standard that a US court accepted that training on lawfully acquired books was transformative, while the acquisition of pirated copies was not, in the litigation that ended in the settlement we covered in our report on Anthropic's $1.5 billion copyright settlement.

Indian fair dealing is a closed list. Section 52 of the Copyright Act enumerates specific permitted purposes — private or personal use including research, criticism or review, reporting current events, and so on. If a use does not fit one of the listed purposes, no amount of transformation or public benefit makes it lawful. ANI's argument follows directly: training a commercial general-purpose model is neither private nor personal use nor research in the statutory sense, and the copying and storage involved is reproduction.

The counter-argument, which persuaded the single judge at the interim stage, runs along two tracks. First, evidentiary: a claimant seeking an injunction must show what was copied and how it surfaces in outputs, and general assertions about training corpora will not do. Second, doctrinal: if the model learns statistical relationships rather than storing expression, the act complained of may not be reproduction of protected expression at all.

Why the appeal matters beyond India

Three reasons, in ascending order of importance.

Market size. India is among the largest user bases for generative AI products in the world. An injunction regime that bites in Delhi changes deployment economics in a way that a ruling in a smaller jurisdiction does not.

Doctrinal transferability. Most of the world does not have fair use. Closed-list exception systems are the norm across the common-law countries that inherited the British model and across most civil-law jurisdictions, including Türkiye. A reasoned appellate judgment on whether model training can fit inside a closed list would be the first authority of its kind, and it would be cited far outside India.

It cannot be settled away. The dominant pattern in the United States is now resolution by payment: content licences with publishers, or class settlements. That pattern answers the commercial question and leaves the legal one open. A judgment on the merits in a closed-list system does the opposite — which is precisely why both sides have fought the interim stage so hard.

The contrast with Europe is instructive. The EU dealt with this by legislating rather than litigating: a text-and-data-mining exception with a machine-readable opt-out for rights holders, and transparency duties for general-purpose model providers under the EU AI Act. Germany's courts, meanwhile, have shown that memorisation can be treated as reproduction, as we reported in our piece on the German AI music copyright ruling. Three regions, three methods: legislation in Europe, settlement in America, adjudication in India.

The jurisdiction problem nobody has solved

OpenAI's January 2025 submission — that deleting training data on an Indian court's order would conflict with its obligations under US law, where litigation holds required preservation — is the most under-examined thread in this case, and it is not a technicality.

It states the structural problem of AI copyright litigation plainly: the training happened once, somewhere, years ago, on servers in another jurisdiction, and produced a set of model weights that are now deployed everywhere. A national court asked to remedy that has three imperfect options. It can order deletion of the data, which may be impossible or unlawful abroad. It can order the model withdrawn from the local market, which is disproportionate in almost every case and commercially catastrophic in the few where it is not. Or it can award damages, which requires quantifying a harm nobody knows how to price.

Indian courts have been more willing than most to assert jurisdiction over foreign platforms where content is accessible and commercially exploited locally, and they have a developed practice of global takedown orders that has already produced friction with US companies. If the Division Bench engages with that line, the appeal becomes about far more than Section 52. The same question is waiting for Türkiye: our courts can block access under Law No. 5651, but blocking a website is a remedy designed for content, not for a model that has already absorbed the content and no longer needs the source.

What the intervener line-up tells you

It is rare for a copyright dispute between two companies to attract this cast. Two amici curiae appointed by the court. The Digital News Publishers Association on one side, arguing that scraping, storage and reproduction without licence or attribution is straightforward infringement. The Broadband India Forum on the other, arguing that large language models do not infringe at all. Now an application by the latter to intervene formally in the appeal.

That line-up is the industry structure made visible. Publishers are litigating for a licensing market; the technology sector is litigating for a doctrinal ruling that removes the need for one. Neither is arguing about ANI's particular articles. Whatever the bench says about Section 52 will be read as an allocation of value between two industries, in a market where domestic Indic-language content is the scarce input that every model developer now needs.

That last point deserves emphasis, because it is the part that translates directly to Türkiye. High-quality Indian-language journalism, like high-quality Turkish-language journalism, is not abundant on the open web. Models that perform well in these languages need exactly the archives that news agencies control. The bargaining position of a Turkish publisher is therefore stronger than the global discourse suggests — provided the legal baseline is clear enough to make a licence worth buying rather than a risk worth running.

What it means for Türkiye

Turkish law sits on the same side of the divide as India, and the resemblance is close enough to be useful rather than merely interesting.

  • No general fair-use principle. Law No. 5846 on Intellectual and Artistic Works sets out a closed list of exceptions — private use, quotation, reporting of current events, education — with no residual balancing power. The transformative-use reasoning that carried the day in the American training cases has no doorway into Turkish law.
  • No TDM exception. Unlike the EU, Türkiye has not enacted a text-and-data-mining exception. A developer training on Turkish-language press archives is therefore in a weaker position than a European counterpart doing the same thing, not a stronger one.
  • Personal use will not stretch. The private-use exception in Turkish law is explicitly non-commercial and does not extend to reproduction for the purpose of building a product. This is the identical argument ANI makes about Section 52.
  • Criminal exposure. Turkish copyright law carries criminal sanctions alongside civil liability, which changes the risk calculus for acquisition of training corpora from unlawful sources in a way that a purely civil regime does not.
  • The personal-data layer sits on top. News archives contain personal data, so the KVKK analysis runs in parallel with the copyright one; we set that out in training data: the KVKK and GDPR questions to settle first, and the cross-border dimension in training AI across borders.

The practical upshot: if the Delhi Division Bench eventually holds that commercial model training cannot be squeezed into a closed list, that reasoning maps onto Article 38 of Law No. 5846 with very little adjustment. Turkish publishers and collecting societies should be reading the Indian pleadings, not waiting for a domestic test case.

What publishers and developers should do now

For publishers and rights holders

  • Fix the evidentiary problem first. ANI lost the interim round on evidence, not on doctrine. Systematic, timestamped records of outputs reproducing protected expression are worth more than any argument about corpora.
  • Reserve rights in machine-readable form. Even without a domestic TDM regime, an express reservation in terms of use and in robots-style signals strengthens a claim in jurisdictions where opt-outs have legal effect.
  • Act collectively. The Indian case shows the weight that publisher associations add; an individual outlet rarely has the resources to sustain two years of interlocutory litigation.
  • Price the licence. Every one of these disputes has ended in a licensing conversation. Rights holders with organised archives and clean metadata negotiate from strength.

For AI developers

  • Know your exception regime per market. A compliance posture built on fair use does not survive translation into India, Türkiye or most of Asia.
  • Document provenance. The single most consequential fact in the Anthropic litigation was where the books came from. Acquisition records are the first thing a court asks for.
  • Control outputs, not just inputs. Regurgitation of protected expression converts a contestable training claim into a straightforward reproduction claim.
  • Expect jurisdictional conflict. OpenAI's January 2025 position — that deletion orders would conflict with US obligations — will be raised again, and courts are becoming less patient with it.

Three ways the appeal can end

  • Dismissed on discretion. The bench holds that refusing an interim injunction was a reasonable exercise of discretion on the material before the single judge, and declines to reach Section 52. This is the most likely outcome procedurally, and it would leave the doctrinal question to the trial — years away.
  • Remitted with guidance. The bench sets out how Section 52 applies to model training while sending the interim question back for fresh consideration. This is the outcome with the greatest international effect: reasoning without a disruptive remedy.
  • Injunction granted in part. A narrow order — for instance, restraining reproduction of ANI content in outputs rather than restraining training — would be the most sophisticated response available, and the one that maps most closely onto how these disputes are actually being settled commercially.

What to watch next

The 5 December hearing will show whether the bench treats this as a narrow review of an interim discretion or as an opportunity to rule on the scope of Section 52. Watch for three things: whether the Broadband India Forum is admitted, which would turn the appeal into a de facto industry reference; whether the court presses ANI on output evidence, which would signal that the evidentiary route remains decisive; and whether any parallel licensing deal is announced, which in every comparable case has quietly preceded a withdrawal.

Frequently asked questions

Did the Delhi High Court rule that AI training is lawful?

No. In July 2026 a single judge refused an interim injunction, finding the evidence insufficient at that stage, with reasoning favourable to fair dealing. That is an interlocutory decision, not a final ruling on the merits, and it is now under appeal.

What happened on 15 September 2026?

A Division Bench issued notice to OpenAI requiring a response to ANI's appeal, indicated it would consider the Broadband India Forum's intervention application, and listed the matter for 5 December 2026.

What is the difference between fair use and fair dealing?

Fair use is an open-ended balancing standard found in US law. Fair dealing, in India and in most closed-list systems including Türkiye, permits only the specific purposes enumerated in the statute. A use outside the list cannot be saved by its transformative character.

Could a Turkish publisher bring the same claim?

In principle yes, under Law No. 5846, and the doctrinal posture would be similar. The practical obstacles are the same as ANI's: proving what was copied, and establishing jurisdiction over a foreign developer.

Does the Anthropic settlement affect this case?

Not as precedent. It was a settlement in a different jurisdiction under a different doctrine. Its relevance is economic: it established a per-work value that informs every licensing negotiation now under way.

Is OpenAI required to stop using ANI content in the meantime?

No. The notice does not stay anything; the position established by the July refusal of interim relief continues until the appeal is decided.

Burhan Doğuş Ayparlar's View

This section sets out my personal assessment as the founder of this site and an AI ethics & compliance counsel.

In my view this appeal is more important than any single American judgment delivered this year, and it is being under-covered for the obvious reason that it is happening in Delhi rather than San Francisco. The American cases have been resolved by money. That is a perfectly rational outcome for the parties and a useless one for the rest of us, because it leaves the central legal question — is training an exception, or is it an unlicensed reproduction that the market simply tolerated for a decade? — permanently unanswered. India is one of the few places where the question can now actually be decided, because there is no open-ended standard available to absorb it.

I also think ANI's core argument is doctrinally strong and was procedurally premature. A closed list is a closed list: you cannot read a commercial training purpose into "private or personal use" without rewriting the provision, and courts in closed-list systems are generally unwilling to do that. But interim injunctions are not granted on doctrinal strength alone, and ANI went in without the output evidence that would have made the harm concrete. The lesson for every rights holder reading this is unglamorous: the case is won in the evidence file, not in the skeleton argument.

For Türkiye the implication is direct. Our Article 38 problem is India's Section 52 problem. If a developer trains a Turkish-language model on newspaper archives without a licence, I do not see a route through the exceptions in Law No. 5846, and the criminal provisions make that exposure sharper than it would be in a civil-only system. I have said before that our priority should be a licensing market rather than a litigation wave, and I hold to that — but a licensing market only forms once the alternative is understood to be risky. A reasoned judgment from Delhi may do more to create that understanding in Ankara than any domestic circular.

My advice to Turkish publishers is to start capturing evidence now, in a disciplined way, and to organise their catalogues so that a licence is something a developer can actually buy. My advice to developers building on Turkish content is to stop reasoning from American cases. They do not apply here, and the day someone runs the ANI argument in a Turkish court, the answer is unlikely to be the comfortable one.

This article is for information only and does not constitute legal advice. The procedural facts are based on press reporting of the 15 September 2026 order, principally Business Standard and MediaNama; the order text was not available at the time of writing. Descriptions of the July 2026 ruling reflect reported summaries rather than the full judgment. Statements about Turkish law are general in nature. The analysis and assessments are the author's own.