AI and Copyright in India: Key Takeaways from ANI v. OpenAI

On 24 July 2026, the Delhi High Court (Justice Amit Bansal) refused to grant an interim injunction to the news agency ANI against OpenAI, the maker of ChatGPT, in one of the first major Indian cases on whether training an AI model on copyrighted content infringes copyright.^[1] The ruling is a prima facie (first-look) decision on an interim application. It does not decide the suit, and it is under appeal. But the reasoning is, as far as we are aware, the most detailed judicial treatment in India of AI training, “memorisation”, retrieval-augmented generation and the fair dealing exception, and it will shape how publishers, AI developers and their lawyers approach these disputes until a higher court says otherwise.

Read the judgment: Official judgment of the Delhi High Court in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC (PDF, 135 pages). A mirror copy is also available from the Supreme Court Observer. Paragraph numbers used below refer to this judgment.^[2]

This article sets out what the Court decided and why, what it left open, where the appeal stands, how the ruling compares with decisions abroad and the Government’s licensing proposal, and what it means in practice. It is written to be read alongside the judgment and is general information, not legal advice.

The case at a glance

Case. ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, CS(COMM) 1028/2024, on I.A. 45300/2024 (an application for an interim injunction under Order XXXIX Rules 1 and 2 of the Code of Civil Procedure).^[1]

Bench and dates. Justice Amit Bansal, Delhi High Court. Judgment reserved on 27 March 2026 and pronounced on 24 July 2026. A corrigendum order of 27 July 2026 corrected a case name in paragraph 80 (discussed below).^{[1, 3]}

Parties and intervenors. ANI is the plaintiff; OpenAI OpCo LLC is the defendant. The Court appointed two amici curiae, Mr Adarsh Ramanujan and Professor Arul George Scaria, and allowed six interventions: the Federation of Indian Publishers, the Digital News Publishers Association and the Indian Music Industry broadly supported ANI, while Flux Labs AI, IGAP Project LLP and the Broadband India Forum broadly supported OpenAI (paras 28 to 31).^[1]

Outcome. The application was dismissed. The Court held, prima facie, that (1) it has jurisdiction, (2) ANI had not shown that ChatGPT’s outputs reproduce its articles substantially or that the model memorises and regurgitates them, (3) storing ANI’s works to train the model falls within the fair dealing exception in Section 52(1)(a) of the Copyright Act, 1957, and (4) the balance of convenience and public interest weigh against an injunction (paras 55, 125, 256 and 270 to 272).^[1]

Status. The Court said its observations are made only for deciding the application and have no bearing on the final outcome of the suit (para 274).^[1] ANI has appealed, and the Division Bench issued notice to OpenAI on 15 September 2026.^[4]

How the dispute arose and what ANI claimed

ANI alleged that OpenAI infringed its copyright in two ways: a training claim (copying and storing ANI’s data to train the large language models, or LLMs, behind ChatGPT) and a reproduction or output claim (reproducing ANI’s works in ChatGPT’s responses) (para 6).^[1] The Court framed four issues on 19 November 2024: whether storage for training infringes (Issue 1); whether the use of ANI’s data to generate responses infringes (Issue 2); whether the use is “fair dealing” under Section 52 (Issue 3); and whether Indian courts have jurisdiction given that OpenAI’s servers are in the United States (Issue 4). The Court noted that “fair use” in Issue 3 must be read as “fair dealing”, the expression used in Section 52(1)(a) (para 9).^[1]

The Court decided Issue 4 first, then Issue 2, and then Issues 1 and 3 together, since the parties and both amici agreed that storage and the fair dealing defence are intertwined (para 10).^[1] The Court also recorded that ANI had, by a letter of 3 October 2024, offered OpenAI a licence to its content for USD 7.5 million, a fact that later mattered to the balance of convenience (para 258.2 and para 265).^[1]

How an LLM works, as the Court understood it

The judgment begins with a technical primer (paras 12 to 27), because the outcome depends on the difference between training and retrieval. In outline: an LLM predicts the next word from the preceding words; it is trained on a very large text corpus, much of it scraped from the web; the text is broken into “tokens”, converted into numerical “vectors”, and the model repeatedly adjusts itself by predicting masked text and comparing it with the original; it is then fine-tuned to follow instructions and avoid harmful output.^[1]

Retrieval-augmented generation (RAG) is different. Instead of relying only on what it learned in training, the system uses the user’s prompt to pull information from an external source at the time of the query, and generates a response grounded in what it retrieved (para 26).^[1] This distinction is the key to the output claim: if a response draws on an article published after training ended, it cannot be the product of memorisation. It must have come from retrieval.

Issue 4: can an Indian court hear a claim about training done abroad?

OpenAI argued that training happens on servers in the United States, that the Copyright Act has no extra-territorial reach, and that the ChatGPT terms of use select the courts of San Francisco (paras 36.1 to 36.6). ANI relied on Section 62(2) of the Copyright Act, which lets a copyright owner sue where it carries on business, and on Section 20 of the Code of Civil Procedure (paras 37.1 to 37.4).^{[1, 5]}

The Court held, prima facie, that it has territorial jurisdiction under Section 62(2) and Section 20 CPC: ANI’s principal place of business and registered office are in Delhi, OpenAI admittedly offers its services to users in India, and the alleged infringing responses were generated in India (paras 45 to 48).^[1] On the territoriality argument, the Court called the storage of ANI’s works on US servers “a terminal step in the chain of events” that begins with access to works in India and their transmission abroad, and held that the Act “does not require this Court to sever the chain of events and examine only the last step”. Otherwise, infringers could evade Indian law by moving the last link offshore (para 51). It relied on the Delhi High Court’s earlier decision in Neetu Singh v. Telegram and held that the training claim cannot be “entirely divorced” from the output claim, which is reproduced in India (paras 52 to 54).^[1]

Takeaway. At the interlocutory stage, an AI developer with Indian users cannot expect to defeat a suit in India merely because model training takes place outside it. The point was decided prima facie, however, and Professor Scaria had cautioned that the training itself, if wholly overseas, may fall outside the Indian Act (para 42.4).^[1]

Issue 2: the output claim, memorisation and substantial similarity

What copyright protects in news. The Court accepted that ANI’s articles and interviews are “original literary works” under Section 13(1)(a) and, prima facie, that ANI owns copyright in those created by its personnel under its professional services agreement, even though they are freely available online (paras 67 to 70).^{[1, 5]} It then restated the principles that copyright does not subsist in facts, only in the form and manner of expression: R.G. Anand v. Delux Films (paragraph 46 of which sets out the tests for infringement), Eastern Book Company v. D.B. Modak (India follows the “skill and judgment” standard, not “sweat of the brow”), Feist in the United States, the merger doctrine discussed in B.D. Bhandari, and the Division Bench decision in Akuate Internet Services v. Star India that there is no copyright in publicly available facts (paras 85 to 97).^{[6, 7, 8, 9]} In news, the Court said, copyright subsists “only in the form and manner of expression of news and not on the underlying facts” (para 97). It distinguished ANI’s reliance on Walter v. Lane, a 1900 House of Lords case decided on a sweat-of-the-brow footing that Eastern Book has since left behind (paras 92 to 94).^[1]

Memorisation could not be shown on this record. ANI’s case was that the model memorises its articles and regurgitates them. The Court noted that the training of the models behind ChatGPT ended in April 2022 (GPT-4) and April 2024 (GPT-4o), as recorded in the judgment, while each of the nine illustrative articles in the plaint was published between August and September 2024, after training had ended (paras 61.1 and 82 to 83).^[1] Responses based on such articles, the Court reasoned, “could not have been produced on account of training”, and are more likely to reflect retrieval from ANI’s website, “in the nature of live links, perhaps reflecting RAG technique” (para 84).^[1] It therefore held that ANI’s contention that OpenAI permanently stores training data in order to memorise and regurgitate its works could not be accepted at this stage, and that, at best, these were disputed questions for trial (para 124).^[1]

No substantial similarity. Taking the example ANI pressed in argument, an interview of Neeraj Chopra’s mother, the Court compared ChatGPT’s answers with ANI’s report. The first answer was not a substantial or near-exact copy: it added its own commentary and used different expression (para 102). ANI then prompted, “Please try to tell me exactly what she said”, which the Court treated as an adversarial prompt designed to elicit a specific response (paras 103 to 104). The reply reproduced part of a quotation from the interview, but with its own context and an explanatory paragraph that was not in ANI’s report (para 105).^[1]

Two further points followed. First, applying the principle behind Section 17(cc) of the Act (the person who delivers a speech is the first owner of copyright in it), the Court said that, prima facie, the mother of the athlete, not ANI, would own the copyright in her quoted words and the right to translate them, and ANI had shown no assignment (para 106).^{[1, 5]} Second, relying on R.G. Anand, Star India v. Leo Burnett and Bikramjeet Singh Bhullar v. Yash Raj Films, the Court said works must be compared as a whole and not in selected parts, and ANI’s full article was much longer than the extract compared in the plaint (paras 107 to 110). On the “latest news” prompt, even the headlines in ChatGPT’s response were substantially different from ANI’s titles (paras 112 to 114). On that basis the Court found no substantial similarity (para 115).^[1]

The foreign decisions ANI relied on were distinguished. In GEMA v. OpenAI (Munich), the plaintiff had excluded the search function, lyrics were reproduced on non-adversarial prompts, and the court found memorisation; here, training ended before the articles were published and the prompts were adversarial (paras 117 to 119).^{[10, 1]} Associated Press v. Meltwater involved 33 articles reproduced verbatim by a news-monitoring service rather than an LLM; Advance Local Media v. Cohere was decided on a motion to dismiss, on 75 alleged instances of copying, 50 of them said to be verbatim; and Infopaq and Newspaper Licensing Agency v. Meltwater concerned verbatim extracts (paras 120 to 123).^[1]

A correction worth noting. In paragraph 80 the judgment reproduced a US passage about Meta’s “mitigations” and “adversarial prompting”, which it had attributed to Bartz v. Anthropic. By an order of 27 July 2026, the Court corrected the reference to Richard Kadrey v. Meta Platforms, Inc.^[3] The correction does not change the reasoning, but readers citing paragraph 80 should use the corrected name.

What was not decided. The Court expressly said that whether outputs produced through RAG infringe copyright “has not been pleaded in the plaint, though it was referred during the course of submissions” (para 84), and Mr Ramanujan had submitted that RAG should be tested at trial (para 65.5).^[1] In its conclusion the Court held, prima facie, that outputs “using RAG technique” did not infringe because they were not substantially similar to ANI’s works (para 271). That is a finding on the nine examples before the Court, not a ruling that RAG-based summaries are always lawful.^[1]

Issue 1: storing a work for training is “reproduction”

On the training claim, the Court first decided in ANI’s favour on the threshold question. Section 14(a)(i) gives the owner of a literary work the exclusive right “to reproduce the work in any material form including the storing of it in any medium by electronic means”. The words on electronic storage were added by the Copyright (Amendment) Act, 1994 (paras 145 to 147).^{[5, 1]} The Court held that storage of a work in any medium by electronic means amounts to reproduction; that Section 14(a)(i) draws no line between temporary and permanent storage; and that, under Section 51, the purpose of the storage is not relevant to whether there is infringement. Even if ANI’s works were deleted after training, “temporary storage occurred in the first place” (paras 148, 149 and 169).^[1]

This is an important holding for rights-holders. The Court did not accept the argument that training is merely an incidental “intermediate step” outside copyright. The question was whether the exception in Section 52 covers it.

Issue 3: fair dealing under Section 52(1)(a)

How Section 52 is read. Section 52(1)(a) provides that “a fair dealing with any work, not being a computer programme, for the purposes of (i) private or personal use, including research; (ii) criticism or review…; (iii) the reporting of current events and current affairs…” is not an infringement. The Court held that Section 52 is not a proviso or an exception to Section 51 but an integral part of the Act that defines the rights of users, and is to be given a broad and liberal reading, relying on Narendera Publishing, B.D. Bhandari and the Supreme Court of Canada’s decision in CCH Canadian (paras 150 to 167).^{[5, 9, 11, 1]} It then applied a two-step test: a purpose test and a fairness test (para 168).

Purpose test: three objections by ANI, all rejected.

(a) Commercial use. ANI argued that Section 52(1)(a)(i) cannot protect a commercial venture. The Court disagreed. Where Parliament wanted to confine an exception to non-commercial use it said so expressly (for example in Sections 52(1)(ad), (k)(ii), (l), (n) and (o)), and no such limit appears in clause (a). A book review or news reporting can be done commercially and remain protected (paras 173 to 179). The Court distinguished Jiwan Publishing House and TIPS Industries v. Wynk Music, where the defendants made copyrighted material available to third parties, whereas OpenAI’s use was “completely an internal process” and the training data is never made available to anyone in natural-language or tokenised form (paras 186 to 190).^[1]

(b) “Non-infringing copy”. The Explanation to Section 52(1)(a) says that storing a work in any electronic medium for the purposes mentioned in the clause, “including the incidental storage of any computer programme which is not itself an infringing copy”, is not infringement. ANI and the intervenors said the stored copy must itself be non-infringing. The Court read the limitation as applying only to the incidental storage of computer programmes, given the placement of the commas, and reasoned that a different reading would protect a researcher who photocopies a page but not one who scans it (paras 191 to 198, 201). It added that, in any event, ANI did not allege that OpenAI obtained its content from unauthorised sources or by breaking its paywall; the data was freely available on ANI’s website (para 200).^{[5, 1]}

(c) “Private or personal use, including research”. The Court held that “private” is wider than “personal” and is not confined to individuals; it can apply to a closed group or a company. OpenAI stores the works in a closed space, accessible only to the models themselves, so the use is private (paras 204 to 211). It also treated training as “research”: research is generally an intermediate, closed activity whose output, not its process, is communicated to the public, and the training of an LLM “can be considered as a form of research for generating new knowledge and advancing artificial intelligence systems” (paras 212 to 215). Applying the doctrine of updating construction, the Court held that research is no longer confined to human beings and “would extend to machine learning as well”; it gave the example of a teacher replaced by an AI bot, saying that to deny that teacher the benefit of Section 52(1)(i) would be a “regressive view” (paras 216 to 219).^[1]

The Court added a candid note: “there is no factual foundation in the plaint to reach a finding of infringement in respect of the training claim”, because ANI had not shown that its extracted data was used in training, but, since temporary storage was admitted, it examined whether the storage was protected (para 202).^[1]

The fairness test: three factors, not the US four-factor test

Having found the purpose satisfied, the Court asked whether the dealing was “fair”. It noted that Indian courts have used different tests, that there is no single uniform test, and that counsel generally agreed that the US four-factor test (Section 107 of the US Copyright Act) does not apply in India (paras 221 to 234). It then formulated three factors for this case, which it described as “broadly in line with” the principles emerging from Article 9 of the Berne Convention, whose paragraph (2) permits exceptions that do not conflict with a normal exploitation of the work or unreasonably prejudice the author’s legitimate interests (paras 235 to 239):^{[1, 12]}

(a) Is the use limited to training? Yes. ANI had not established memorisation or regurgitation, and no instance was shown of ANI’s works being used for any purpose other than training (paras 241 to 243).^[1]

(b) Does it compete with ANI or prejudice its legitimate interests? No, prima facie. ChatGPT performs many functions beyond news; in news it provides summaries or snippets with references, in the manner of an AI-enabled search engine; its purpose and character are “fundamentally different” from ANI’s news articles; and ANI had placed nothing but bare averments on record to show lost market share or reduced subscription revenue (paras 244 to 251). The Court found support in the US decisions in Bartz v. Anthropic, Kadrey v. Meta and Authors Guild v. Google, which treated training or digitisation as transformative.^{[1, 13]}

(c) Does it serve the public interest? Yes. The Court pointed to access to information, education, scientific research, software development, translation and tools for persons with disabilities, and said that the societal benefits of scientific and technological research are a relevant consideration in judging fairness, while the rights of copyright owners remain important (paras 252 to 254).^[1]

The Court concluded that both the purpose test and the fairness test were met, so the storage of ANI’s works to train ChatGPT’s underlying models falls under Section 52(1)(a) and is not infringement (paras 255 and 256).^[1]

Balance of convenience, irreparable injury and the USD 7.5 million offer

Even apart from the merits, the Court held that an injunction was not warranted. Its reasons (paras 262 to 269) were:^[1]

Opt-out was available. It was admitted that ANI can block web crawlers, including for search and RAG, but had not done so; OpenAI said it had in fact blocked ANI’s website for scraping and for ChatGPT search. ANI’s paying subscribers could also block their own sites but had not, “possibly for their own commercial reasons” (paras 262 and 263).

No loss shown. There was no averment or material that OpenAI’s activities had cost ANI subscribers or hurt its syndication business (para 264).

The claim is quantifiable. ANI itself had offered OpenAI a licence for USD 7.5 million by letter of 3 October 2024, so its claim can be compensated in money if it succeeds, whereas an injunction would significantly affect OpenAI in ways that money could not repair (para 265).

Public interest. The Court noted the use of AI in education, healthcare, financial services, agriculture and skills, observed that it would be “economically unviable” to develop an LLM if training needed licences from many sources, and held that an injunction would harm the growth of AI, particularly LLMs being developed in India, and millions of ChatGPT users in India, many of them not paid subscribers (paras 266 to 269).

The application was dismissed (para 273).^[1]

What the judgment does not decide

It is easy to over-read an interim ruling. The following points were left open, and the Court said so itself:

It is not a final decision. The suit remains pending. The Court said the observations are for the application only (para 274).^[1]

Memorisation was not ruled out. The Court held that ANI had not shown it on the examples before it, and said that whether LLMs memorise and regurgitate is a matter for trial, with evidence (para 124). Professor Scaria and Mr Ramanujan had both recognised that memorisation can occur, particularly through extraction attacks or with unique or repeated content (paras 78 and 65.2).^[1]

RAG and search were not decided as a category. The Court said the RAG point was not pleaded and decided only that the nine examples were not substantially similar (paras 84 and 271).^[1]

Ownership of individual works was left to trial, apart from the prima facie view that ANI owns copyright in the works its personnel create (para 70).^[1]

The legal findings are prima facie. The holdings that storage is reproduction, that “private” can include a company, and that training is “research” are interpretations at an interim stage, in a case where the Court recorded no factual foundation for the training claim (para 202), and they will be tested on appeal.^[1]

The appeal: where matters stand

ANI appealed to the Division Bench of the Delhi High Court. On 15 September 2026, a bench of Justice Avneesh Jhingan and Justice Manmeet Pritam Singh Arora issued notice to OpenAI and did not grant ANI interim relief at that stage, declining to pass any direction without first hearing OpenAI.^{[14, 4, 15]} Reports differ on the next date, with one giving 8 December 2026 and others 5 December 2026, so the cause list should be checked.^{[4, 15]}

As reported, ANI’s grounds include that India follows fair dealing, not fair use, and Section 52 is a closed list of purposes that cannot stretch to large-scale commercial exploitation; that the public-interest finding conflicts with OpenAI’s own case that ANI’s content is a minuscule share of its training data; that a licensing market exists, since OpenAI licenses news content from publishers such as the Financial Times; and that public accessibility does not give anyone a right to make and store copies.^[15] These are ANI’s contentions; the Division Bench has not ruled on them, and no outcome should be assumed.

How India compares: the US, Germany and the DPIIT proposal

United States. US law uses an open-ended “fair use” test with four factors. In June 2025, two federal judges in California found that training on books was fair use on the records before them: Bartz v. Anthropic (while holding that building a central library from pirated copies was not protected, a case later reported to have settled for USD 1.5 billion) and Kadrey v. Meta (where the judge stressed that plaintiffs had presented no meaningful evidence of “market dilution” and indicated that such evidence could change the result). In Thomson Reuters v. Ross Intelligence, a 2025 decision involving a non-generative AI legal search tool, the court rejected fair use because the tool was a competing product.^[13] The Delhi High Court drew on the first two (and on Authors Guild v. Google) for support, but decided the case under India’s purpose-limited statute and its own three factors rather than adopting the US test (paras 234, 247 and 248).^[1]

Germany. In GEMA v. OpenAI (11 November 2025), the Munich Regional Court held that ChatGPT infringed copyright in song lyrics by memorising them within its models and reproducing them in outputs.^{[10, 1]} The Delhi High Court distinguished it on the facts, as discussed above. Between them, the cases suggest that outputs which closely reproduce protected text are where liability is most likely to arise, while courts are more receptive to defences for training itself.

The DPIIT working paper. On 8 December 2025, a committee of the Department for Promotion of Industry and Internal Trade issued Part I of a working paper on generative AI and copyright, proposing a “hybrid model”: a mandatory blanket licence for lawfully accessed copyrighted works used in AI training, with a statutory right to remuneration; a central body, the Copyright Royalties Collective for AI Training (CRCAT), to collect and distribute royalties; royalties calculated as a share of the global gross revenue of the commercialised AI system, with rates set by a government-appointed committee and subject to judicial review; and no right for creators to withhold their works from training. The committee rejected a blanket text-and-data-mining exception and judged an opt-out model unworkable. Part I leaves the question of AI outputs to a Part II.^{[16, 17]} This is a proposal for consultation and, to our knowledge, not law.

Why the two sit in tension. In our view, the DPIIT proposal and the ANI ruling point in different directions. The Delhi High Court, applying the Act as it stands, held at the interim stage that training can fall within fair dealing, even for a commercial developer, and that “lawful access” is not a condition of Section 52(1)(a). The DPIIT paper would, if enacted, make a paid licence the price of training on lawfully accessed content. Whether the courts or Parliament settle the question first will matter greatly for news publishers and AI developers alike.

Practical points for publishers, AI developers and businesses

For news and content owners. In our view, the judgment shows what a court will look for. (1) Technical controls matter. The Court gave weight to the fact that ANI could block crawlers and had not, so review robots.txt and crawler settings, paywalls and the terms on which syndication partners display your content (paras 262 to 263). (2) Evidence of harm matters. Bare averments of lost market share were not enough, so keep data on subscriptions, traffic and licensing revenue (paras 250 and 264). (3) Plead and test outputs carefully. Record dated, non-adversarial test prompts, compare outputs with the whole work and not extracts, and plead retrieval or search outputs specifically, since RAG was held not to have been pleaded (paras 84, 107 to 110). (4) Prove your chain of title. The Court relied on ANI’s professional services agreement for ownership and held that quotations from interviewees may belong to them, so check contracts with journalists, freelancers and contributors (paras 68 to 69 and 106). (5) Think about licence offers. A price put on a licence before suit was used to show that the loss can be compensated in money (para 265).^[1]

For AI developers. The ruling is favourable but narrow. Respecting crawler controls and paywalls, keeping the use of data to training, mitigating regurgitation and documenting the purpose of storage all helped the defence here (paras 200, 243 and 262). Retrieval-based features that display or summarise current news are the less settled area, and the appeal and any change in the law remain live risks.^[1]

For businesses and creators generally. Review the terms on which you publish, license or scrape content, and the rights you grant to AI services. If you are building or using AI tools in India, keep records of data sources and of what the tool outputs. Disputes of this kind turn on evidence, as much as on the law.

For related reading, see our article on intellectual property enforcement in the generative AI era. Our intellectual property practice advises on copyright, licensing and AI-related disputes. This article is general information, not legal advice.

Frequently asked questions

Did the Delhi High Court rule that training AI on copyrighted content is legal? Not finally. In an interim, prima facie decision, it held that storing ANI’s works to train ChatGPT’s underlying models falls within the fair dealing exception in Section 52(1)(a) of the Copyright Act. The suit is pending and the decision is under appeal.

Did ANI lose the case? ANI lost its application for an interim injunction. The suit itself has not been decided.

Is there copyright in news? Yes, in the original expression of a news report. There is no copyright in the underlying facts, and the Court said the threshold for showing substantial similarity in news is higher than for creative works such as song lyrics (paras 97 and 122).

Is storing a copyrighted work in a computer a “reproduction”? Yes. Section 14(a)(i) includes “the storing of it in any medium by electronic means”, and the Court held that this covers temporary as well as permanent storage.

Does a company count as “private” use under Section 52(1)(a)(i)? The Court held that “private” is not limited to individuals and can include a company, though that holding is prima facie and under appeal.

Where can I read the judgment? The official judgment is on the Delhi High Court’s website: ANI Media Pvt. Ltd. v. OpenAI OpCo LLC (PDF).

Conclusion

ANI v. OpenAI is an early and detailed Indian judgment on AI training and output under the law as it stands. Its core message is a balanced one: facts are free, copyright protects expression, and a court will ask for evidence of copying and of harm before it restrains a technology used by millions. At the same time, the Court accepted that storing a work electronically is reproduction, left memorisation and retrieval-based outputs open for trial, and recorded that the legislature never imagined AI when it last amended Section 52 in 2012 (para 216).^[1]

The final word has not been said. The Division Bench will hear the appeal, the suit will proceed to trial, and the Government’s licensing proposal may yet change the rules. Publishers and AI developers should plan on that basis: follow the reasoning, not just the result, and revisit it as the case develops.