Summary: The Delhi High Court refused ANI’s interim injunction, finding that OpenAI’s model training prima facie qualified as fair dealing for “private or personal use, including research,” while ANI’s post-training examples failed to prove memorisation or substantial reproduction.
The refusal of immediate relief was defensible, but the broader reasoning is vulnerable because it stretches “private research,” underestimates licensing and market harm, conflates public accessibility with lawful copying, and selectively imports US fair-use principles into India’s narrower statutory framework.
The ruling gives AI developers significant—but provisional—freedom to train on publicly accessible works, while forcing rights owners to provide model-specific evidence of copying, provenance, regurgitation and economic harm; its wider international influence will remain limited by differing national laws.
ANI v OpenAI: A Bold Indian Safe Harbour Built on Uncertain Legal Foundations
by ChatGPT-5.6
Introduction
The Delhi High Court’s judgment in ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, pronounced on 24 July 2026, is the first substantive Indian judicial ruling on whether copyrighted material may be copied and stored for the training of a generative AI model without the permission of the rights owner. It is potentially one of the most consequential copyright decisions yet delivered by an Indian court.
Its immediate effect, however, must not be overstated. Justice Amit Bansal was deciding ANI’s application for an interim injunction, not finally determining the suit after discovery, expert evidence and trial. The judgment expressly states that its findings are prima facie and will not control the final outcome. Nevertheless, the court went considerably beyond finding that ANI had failed to justify emergency relief. It held, provisionally but unequivocally, that:
the Delhi High Court had jurisdiction despite OpenAI’s training infrastructure being located abroad;
the temporary electronic storage of ANI’s works during model training constituted reproduction under Section 14 of the Copyright Act 1957;
that reproduction was nevertheless protected as fair dealing for “private or personal use, including research” under Section 52(1)(a);
ANI’s examples did not establish memorisation or regurgitation;
the outputs shown were not substantially similar to ANI’s protected expression; and
the balance of convenience and public interest strongly opposed an injunction.
The court therefore dismissed ANI’s application, while preserving the main suit.
The refusal of an interim injunction was, in my assessment, defensible on the evidential record. The much broader proposition that commercial LLM training on publicly available copyrighted works is ordinarily “private research” and fair dealing is substantially more vulnerable. The judgment contains several persuasive insights, but it also conflates secrecy with private use, underestimates emerging licensing markets, relies selectively on foreign jurisprudence, stretches Delhi High Court precedents beyond their original contexts, and allows broad industrial-policy considerations to substitute for some of the statutory analysis.
1. The jurisdiction ruling: sensible result, uncertain reasoning
The court’s approach
The court correctly separated two questions in principle:
whether the Delhi High Court was an appropriate forum to hear the suit; and
whether Indian copyright law could govern acts of training and storage allegedly occurring in the United States.
ANI had its principal and registered offices in Delhi. OpenAI deliberately offered paid and free services to users in Delhi, and the allegedly infringing outputs were generated there. The court consequently held that both Section 62(2) of the Copyright Act and Section 20 of the Code of Civil Procedure supplied territorial jurisdiction.
That part is strong. A foreign digital platform that intentionally serves Indian customers and produces allegedly infringing outputs in India should not ordinarily be able to exclude Indian courts through server location or contractual forum clauses.
The more difficult issue was whether Indian law could govern the training copies. OpenAI contended that the relevant copies were made and stored on US servers. The court accepted that the Copyright Act does not operate extraterritorially, but reasoned that US storage was merely the “terminal step” in a chain beginning with access to ANI’s works in India and transmission abroad. It warned that otherwise an infringer could evade Indian copyright law simply by locating servers outside India.
The court relied heavily on Blueberry Books v Google India and Neetu Singh v Telegram. It also refused to separate the training claim from the output claim because the Indian output depended upon the trained model.
Does this reasoning hold up?
Only partially.
The court is right that server location must not mechanically determine jurisdiction. Cloud services should not become legally untouchable merely because data centres sit abroad. But forum jurisdiction, territorial infringement and the applicable substantive law are different inquiries.
Neetu Singh involved copyrighted course materials being unlawfully circulated to users in India through Telegram. The foreign server was infrastructure supporting an infringement whose dissemination and effects occurred in India. The quoted reasoning itself emphasised circulation within the Indian jurisdiction. That does not necessarily establish that an entirely foreign act of making a training copy becomes governed by Indian law merely because the source website or rights owner is Indian.
Similarly, the passage from Blueberry Books did not conclusively establish that Indian copyright law governs every foreign copy. It rejected reliance, at an interim stage, on Amazon’s contested factual assertion that the allegedly infringing conduct occurred exclusively in the United States. The Delhi court turned that procedural caution into a wider territorial principle.
Most importantly, the judgment assumed rather than demonstrated that OpenAI accessed ANI’s works “from India.” A web crawler operating on overseas infrastructure might retrieve a publicly accessible page hosted in India, outside India or through a distributed content-delivery network. The location of the rights owner is not necessarily the location of the act of reproduction. The court needed evidence about where the crawler, initial copy, dataset processing and training copies were made.
The better analysis would have been:
Indian courts plainly possess jurisdiction over the Indian output, RAG and service-related claims.
The law applicable to each training copy should depend on where the legally relevant reproduction occurred.
Any access, copying or transmission occurring in India may be governed by Indian law.
Copies made exclusively abroad remain subject primarily to the copyright law of the country where they were made, unless Parliament has provided otherwise.
Closely connected claims may be tried together for procedural efficiency, but that does not make the same national copyright law applicable to every act.
The result therefore protects Indian rights owners from being left without a forum, but the “unbroken chain” theory risks extending Indian law far beyond its accepted territorial scope.
2. The court’s treatment of news copyright and originality
The court correctly began with the idea–expression distinction. Facts, events and “news of the day” cannot be monopolised. Copyright can subsist in the author’s selection, arrangement, wording, translation and other original expressive contributions, but not in the underlying event.
Relying on Eastern Book Company v D.B. Modak, the court rejected the old “sweat of the brow” standard and applied the Indian requirement of skill and judgment with a non-trivial creative contribution. It also relied on Feist, the merger doctrine and Delhi authority involving public sporting information to conclude that factual works generally receive relatively “thin” copyright protection.
That is doctrinally sound. ANI cannot obtain a monopoly over a cabinet decision, a court order, a sporting result or words spoken by a public figure merely because ANI reported the fact first.
Nevertheless, the judgment sometimes moves too quickly from “news contains facts” to “news expression receives little protection.” News works vary enormously. A bare market update may contain little protectable expression. An investigative report, exclusive interview, original translation, curated chronology or deeply contextualised analysis may contain substantial authorship. The relevant inquiry should concern the originality and qualitative importance of the material copied in each work, rather than the general label “news.”
The court’s suggestion that the threshold for substantial similarity must necessarily be higher for all news articles is directionally understandable but too categorical. Thin copyright means protection against close copying of the protectable contribution; it does not mean that copying a short but distinctive and valuable passage becomes harmless.
3. Memorisation, RAG and the output claim
The strongest part of the judgment
ANI’s nine pleaded examples were published in August and September 2024. OpenAI’s evidence stated that training of GPT-4 ended in April 2022 and GPT-4o in April 2024. The court reasonably concluded that those articles could not have been memorised during the stated pre-training of those model versions.
That is a powerful factual answer to ANI’s particular memorisation theory. A work created after a model’s training cut-off cannot prove that the work was contained in that training dataset.
The court then inferred that the responses were likely produced through live retrieval or RAG rather than model memory. It correctly observed that post-cut-off output cannot be used to prove pre-training memorisation. It also acknowledged that ANI had not properly pleaded RAG infringement.
The court was therefore justified in rejecting the claim that these examples demonstrated permanent memorisation in GPT-4 or GPT-4o.
Where the analysis becomes incomplete
The temporal mismatch defeats one theory but does not resolve the copyright status of the outputs. If ChatGPT Search or another retrieval component copied ANI material after publication and supplied it to the model, copyright questions still arise at the retrieval, temporary-storage, summarisation and communication stages.
The court sometimes treats “not memorised” as moving the case substantially towards “not infringed.” That is too broad. An output can infringe because a live retrieval system copied and communicated protected expression, even where no model weight contains the work.
The judgment also assumes that an answer about a post-cut-off event “perhaps” reflected RAG. That may be a reasonable inference, but it is not a technical finding. Other possibilities could include updated model components, search indexes, cached material, third-party retrieval, tool use or later post-training. These issues require model-specific records and expert evidence.
The court’s own decision to use “LLM” and “ChatGPT” interchangeably compounds the difficulty. ChatGPT is a service containing models, system instructions, safety layers, search, retrieval tools, APIs and potentially third-party data flows. Model pre-training, fine-tuning, RAG indexing and real-time retrieval are legally and technically distinct processes.
4. Substantial reproduction: the court applied an excessively rigid “whole work” approach
The court found that ANI’s examples reproduced facts and, in places, quotations but did not reproduce ANI’s articles as a whole. It relied on R.G. Anand, Star India v Leo Burnett and Bikramjeet Singh Bhullar for the proposition that works should be compared as complete entities rather than dissected into selected passages.
This is one of the judgment’s most questionable doctrinal moves.
Section 14 protects a work or any substantial part thereof. Substantiality is qualitative as well as quantitative. A defendant cannot normally escape liability by taking the most distinctive, valuable or original portion of a work and pointing out that most of the article remains uncopied.
Even the Leo Burnett passage reproduced in the judgment expressly reserved the possibility that copying a particular scene with a striking resemblance could constitute substantial copying. The authority therefore does not support an absolute rule against examining protectable components.
The proper method is more nuanced:
identify the elements of the claimant’s work that are original and protectable;
exclude facts, ideas, commonplace formulations and merged expression;
compare the allegedly copied material with those protectable elements;
consider both its quantity and its qualitative importance; and
assess the overall impression without allowing unprotectable surrounding matter to obscure close copying of the original core.
For news articles, comparing the entirety of a short AI answer with a much longer article may systematically under-detect infringement. A service could reproduce the most commercially useful paragraphs while omitting background material and thereby appear dissimilar “as a whole.”
5. The “adversarial prompt” issue
The court attached weight to ANI asking ChatGPT to state “exactly” what Neeraj Chopra’s mother had said. It characterised the request as adversarial prompting designed to extract protected material.
The nature of the prompt is relevant, but its legal significance should be limited.
An adversarial prompt can affect:
whether reproduction is representative of ordinary use;
whether the provider could reasonably foresee the output;
responsibility between the user and provider;
the adequacy of output safeguards;
injunctive proportionality; and
damages.
It should not, however, alter whether the resulting output objectively reproduces a substantial part of a protected work. A photocopy does not cease to be a copy because an investigator deliberately pressed the photocopier’s button. Similarly, a model provider cannot necessarily avoid output liability merely because a user asked directly for protected text—particularly where the system is designed to respond and the provider controls safeguards.
The foreign cases make this distinction important. The Munich Regional Court’s decision in GEMA v OpenAI involved repeated reproduction of protected lyrics and treated memorisation and output reproduction as legally relevant. The Delhi court legitimately distinguished GEMA because its own record lacked repeated verbatim lyrics and involved post-cut-off material. But the adversarial character of a prompt should be evidence about causation and controls, not a freestanding defence.
6. The interview and translation: a potentially serious mistake
The court held that Neeraj Chopra’s mother, rather than ANI, was prima facie the copyright owner of the quoted words under Section 17(cc), which concerns an address or speech delivered in public. It further reasoned that the translation right belonged to her and that ANI had not shown an assignment.
Several issues were left unresolved.
First, an interview answer is not automatically an “address or speech delivered in public.” The court said it was applying the “underlying principle” of Section 17(cc), but copyright ownership should not ordinarily be transferred through analogy where the statutory category may not fit.
Second, the author of the underlying spoken words and the author of a translation are not necessarily the same person. A sufficiently original translation may itself be a literary work created by the translator, without prejudice to the rights in the underlying work. Whether ANI owned the translated version could depend on:
consent to record and translate the interview;
the originality of the translation;
the employment status of the translator;
contractual terms;
implied licences; and
whether the underlying speaker retained enforceable rights in the particular circumstances.
Third, ANI’s complaint was not necessarily that it owned the underlying Hindi speech. It may have claimed rights in the particular English rendering, editorial selection and surrounding article.
The court therefore conflated the exclusive right to authorise a translation of an underlying work with ownership of the original expression contributed by the translator. This point should have been left for evidence rather than used affirmatively against ANI at the interim stage.
7. Training copies: the court correctly recognised reproduction
OpenAI and supporting intervenors argued that model training used only non-expressive information—linguistic patterns, grammar, correlations or statistical relationships—and that model weights were not copies of ANI’s works.
The court wisely did not use this theory to deny that copying occurred. It held that electronically storing a literary work, whether temporarily or permanently, is reproduction within Section 14(a)(i). The purpose or intention behind that storage is irrelevant at the initial infringement stage.
That is one of the judgment’s strongest holdings. Training normally requires the acquisition, parsing and temporary or persistent storage of expressive works. Even if the final weights are not themselves readable copies, intermediate copies may still engage the reproduction right.
The “non-expressive use” concept is better considered under a statutory exception or fair-dealing analysis than used to pretend that no reproduction took place.
The court also correctly treated Sections 14, 51 and 52 as an integrated statutory scheme. An act within Section 52 is not infringing merely because it would otherwise fall within the rights enumerated in Section 14.
8. Commerciality: the court is broadly right, but only up to a point
ANI argued that a profit-making corporation could not rely on fair dealing for private use or research. The court rejected an automatic commercial-use bar. It observed that Section 52 expressly requires non-commercial use in certain other provisions but does not do so in Section 52(1)(a). Criticism, book reviews and news reporting can also be conducted commercially without automatically becoming unfair.
This reasoning is persuasive. Commerciality should not be a categorical exclusion. A newspaper’s commercial review, a lawyer’s paid research or a commercial documentary’s criticism may qualify as fair dealing.
The court also reasonably distinguished Jiwan Publishing House and TIPS v Wynk. In those cases, defendants commercially supplied the protected materials to users and tried to rely on the users’ private or research purposes. OpenAI asserted that its training corpus was internal and not delivered to users.
But “commerciality is not automatically fatal” is not equivalent to “commerciality is irrelevant.” It remains central to fairness, motive, market effect and proportionality. The court’s later analysis gives insufficient weight to the scale of copying, OpenAI’s commercial objectives and the direct contribution of high-quality content to the value of its products.
9. Is commercial LLM training “private use”?
The court interpreted “private” through dictionary definitions as including something restricted to a person, group, class or company. Because the training corpus was held internally and was allegedly accessible only to the model, the court concluded that OpenAI’s use was “purely private.”
This is the judgment’s most vulnerable statutory conclusion.
“Private” describes more than whether a process is conducted behind closed doors. The objective use of ANI’s material was to improve a product supplied to millions of members of the public, enterprise customers and API developers. The training process may be confidential, but its commercial and public-facing purpose is not private in any ordinary copyright sense.
Under the court’s interpretation, almost any internal copying by a corporation could qualify as private—even copying undertaken to manufacture and launch a mass-market substitute—provided the internal dataset was not itself publicly downloadable. That would transform “private use” into a broad industrial copying exception.
B. Malini Mallya, on which the court relied, involved performance within an educational institution and contained brief observations about the availability of several Section 52 provisions. It is weak authority for the proposition that a corporation’s internal product-development pipeline constitutes private use.
The stronger interpretation is that secrecy or internal access is relevant to the character and possible fairness of the dealing, but does not by itself establish the statutory purpose of private use.
10. Is model training “research”?
The conclusion that LLM development can involve research is plausible. Training involves systematic experimentation, analysis, optimisation and the development of computational capabilities. The term “research” should not be limited to traditional academic institutions or human beings manually reading texts.
The court invoked updating construction, reasoning that an ongoing statute should be applied to technological developments unforeseen when it was enacted. It even suggested that replacing a human teacher with an AI teacher should not defeat an educational exception. It consequently held that machine learning could qualify as research undertaken at human direction and for human benefit.
The difficulty is not whether machines may assist research. It is whether the phrase “private or personal use, including research” extends to wholesale copying for the development of a general-purpose commercial product.
The legislative history is important. The original 1957 formulation covered “research or private study.” In 1994 it became “private use, including research,” and in 2012 “private or personal use, including research.” Grammatically, “including research” may be read as a species of private or personal use rather than as an entirely independent, unlimited category.
The court’s reliance on Canada is therefore incomplete. Canadian Section 29 separately identifies “research” and “private study.” In CCH, commercial legal research was capable of qualifying, but the Canadian statute did not place research inside the phrase “private or personal use.” The provisions are similar but not truly pari materia on this critical point.
Canadian jurisprudence also performs much heavier analysis at the fairness stage. It examines the purpose, character and amount of dealing, alternatives, the nature of the work and market effect. Commercial research may be less fair, and reasonable safeguards matter.
Updating construction permits a statutory term such as research to operate in a technological environment. It does not permit a court to remove adjacent limiting words or construct a general text-and-data-mining exception that Parliament did not enact.
The court’s AI-teacher analogy is similarly weak. Whether an AI system can act “in the course of instruction” under a different statutory clause does not demonstrate that every act used to build the AI system is private research. Teaching with an AI tool and copying billions of works to manufacture that tool are different uses.
11. The “non-infringing copy” question
The Explanation to Section 52(1)(a) refers to electronic storage and the incidental storage of a computer program “which is not itself an infringing copy.” ANI argued that fair dealing required the stored source to be a lawful, non-infringing copy.
The court concluded, primarily from punctuation and statutory structure, that the non-infringing-copy qualification applied only to the incidental computer program, not to the literary work being stored. That grammatical interpretation is plausible.
The court then went further. It reasoned that OpenAI obtained ANI’s content from freely available web pages, did not circumvent a paywall and therefore could not be said to have obtained an infringing copy.
That step is problematic. Lawful access to a webpage and permission to make a new reproduction are distinct. A person can lawfully read an online article while lacking authority to download, aggregate and retain it for another purpose. Public availability does not place a work in the public domain.
The court used the analogy of a researcher who may photocopy a page in a library and should not lose fair-dealing protection merely because the copy is electronic rather than physical. That supports technological neutrality, but it does not answer whether systematic acquisition of entire works for commercial model development is fair.
The court cited Bartz v Anthropic, but Bartz sharply distinguished lawfully purchased and scanned books from millions of books downloaded from pirate libraries. Training copies were held fair use, while the creation of a permanent central library from pirated sources was not. The source and acquisition pathway were therefore legally material.
OpenAI’s precise acquisition sources, crawler practices, contractual restrictions, syndicated copies and any rights reservations were not fully tested in the ANI proceedings. The court should not have treated free web access as affirmative proof of lawful copying.
12. Delhi High Court precedent and the improvised fairness test
The court candidly recognised that Indian decisions have not adopted one uniform fair-dealing test. It surveyed:
Narendera Publishing, which emphasised transformative character in the context of review and educational guidebooks;
B.D. Bhandari, which considered whether guidebooks possessed a materially different or derivative character;
Rameshwari Photocopy, which concerned the specific educational exception and held that the US four-factor cases had little persuasive force in that context;
Blackwood, which examined the quantity and quality taken;
ESPN Star Sports, which considered commercial competition, prior publication and the amount and importance taken;
Super Cassettes v Chintamani Rao, which considered proportionality and intellectual contribution; and
India TV v Yashraj Films, which referred to four US factors but ultimately applied de minimis principles.
The judgment accurately concluded that no single Indian formula has been applied consistently.
It then created three factors of its own:
whether OpenAI’s use was limited to model training;
whether the use economically competed with or prejudiced ANI’s legitimate commercial interests; and
whether ChatGPT served the public interest.
These are relevant considerations, but they are not sufficient.
The new test omits or marginalises several matters repeatedly recognised in Indian, Canadian, UK and US fair-dealing or fair-use jurisprudence:
the quantity and qualitative importance of what was copied;
whether entire works were systematically reproduced;
the duration and character of storage;
whether copies were deleted or retained;
the availability of less intrusive alternatives;
whether the copying was reasonably necessary;
the nature and creativity of each work;
the lawfulness of acquisition;
compliance with technological or machine-readable reservations;
safeguards against memorisation and regurgitation;
the availability of attribution or referral;
whether licences were practically available;
whether a licensing market existed or was emerging; and
whether the defendant’s conduct was proportionate to its claimed purpose.
The test is therefore structurally weighted towards AI developers: demonstrate internal training, deny direct substitution and invoke broad technological benefits.
13. The Berne three-step test was only partially applied
The court said its factors were broadly aligned with Article 9(2) of the Berne Convention, which permits reproduction exceptions only in “certain special cases” that do not conflict with normal exploitation and do not unreasonably prejudice an author’s legitimate interests. The Convention’s formulation is confirmed by WIPO’s official text.
But the court did not meaningfully examine the first limb: whether its interpretation remained a “certain special case.” A rule covering internal copying by any commercial corporation for the development of any AI system could become extremely broad.
The court also interpreted “normal exploitation” primarily through present substitution and ANI’s existing subscriber revenue. Normal exploitation can include reasonably developing licensing markets, especially where commercial AI companies are already entering content-licensing agreements.
Finally, “legitimate interests” extend beyond demonstrable short-term subscriber loss. They may include control over wholesale reproduction, licensing opportunities, attribution, the integrity of translated or curated content, referral traffic and the ability to develop competing AI products.
The judgment’s invocation of Berne therefore gives the test less constraining effect than its language requires.
14. Transformative use and the selective borrowing of US fair use
The court stated that the US statutory four-factor test was not directly applicable, but then relied extensively on the American idea of transformative use and on Authors Guild v Google, Bartz and Kadrey.
That creates an internal tension. Indian fair dealing is purpose-limited. US fair use is open-ended and expressly balances four statutory factors. Transformative use may help determine fairness once an Indian statutory purpose is established, but it cannot itself manufacture a missing statutory purpose.
The US authorities are also more divided and fact-sensitive than the Delhi judgment suggests.
Bartz v Anthropic
Judge William Alsup held that the use of books to train Anthropic’s LLMs was highly transformative. But he separately held that building a permanent central library from pirated shadow-library copies was not fair use. The subsequent litigation and settlement concerned that piracy-related copying.
The case therefore supports a possible fair-use defence for training, but also strongly supports source provenance and lawful acquisition—issues the Delhi court downplayed.
Kadrey v Meta
Judge Vince Chhabria granted Meta summary judgment on the particular evidentiary record but strongly rejected the idea that market harm from generative AI is analogous to children learning to write. He explained that market dilution from automated production could become decisive in future cases. Meta prevailed because the plaintiffs failed to produce sufficient evidence connecting Llama to harm to their own markets, not because all training was categorically fair.
The Delhi court adopted the pro-transformative passage while giving less attention to this warning about market dilution.
There also appears to be at least one attribution error in the Delhi judgment: a passage discussed as part of Bartz refers expressly to Meta and Llama and corresponds to the reasoning in Kadrey. That does not determine the result, but it illustrates the danger of treating the two US decisions as a single consensus.
Thomson Reuters v ROSS Intelligence
The Delaware federal court reached the opposite result in a dispute involving an AI-assisted legal research competitor. ROSS used Westlaw headnotes as data to develop a competing legal research system. The court held the use was commercial and not transformative because it shared substantially the same purpose, and it recognised both the legal-research market and the market for AI-training data.
ROSS was not a generative LLM, so the case is distinguishable. Nevertheless, it disproves any proposition that using material “as AI data” necessarily creates a different purpose.
Warhol v Goldsmith
The US Supreme Court has warned against treating any new expression or technology as transformative. The inquiry concerns the specific challenged use, its objective purpose and the degree of difference from the original. The same copying may be fair for one use and unfair for another. Commercial uses sharing substantially the same purpose ordinarily weigh against fair use unless independently justified.
That methodology should have led the Delhi court to separate:
acquisition and dataset creation;
temporary training copies;
persistent archival copies;
model-weight use;
RAG indexing;
live retrieval;
output generation;
summaries;
translations; and
licensing or commercial deployment.
Instead, the judgment sometimes evaluates “ChatGPT” as one undifferentiated socially valuable system.
15. Why Google Books is an imperfect analogy
The court relied on Authors Guild v Google, where digitisation, search and limited snippets were held fair use. That case supports intermediate copying for a non-substitutive computational purpose.
But important safeguards distinguished Google Books:
books were acquired through participating libraries;
public access was limited to search and controlled snippets;
security measures restricted reconstruction;
the system did not generate unlimited substitute prose;
it facilitated discovery of the original works; and
its market effect was assessed in light of those architectural constraints.
A general-purpose LLM is capable of summarisation, rewriting, translation and generating content that may compete with the economic functions of the source material. Google Books is therefore relevant but cannot supply a blanket rule.
16. Canada does not support the judgment as strongly as suggested
The court relied on CCH for a large and liberal interpretation of research and the proposition that commercial legal research can qualify. That is correct.
But Canadian law also insists upon an objective assessment of the “real purpose or motive” and a detailed fairness analysis. In SOCAN v Bell, commercial music services could facilitate consumers’ research because previews were short, streamed, automatically deleted, lower quality, non-substitutive and accompanied by safeguards. The Supreme Court examined amount, character, alternatives, necessity, nature and market effect.
OpenAI’s use differs materially:
it copies entire works rather than short previews;
the copies are assembled systematically at immense scale;
the purpose is product development rather than helping users select the original work;
deletion and retention practices were not fully established;
referral and remuneration are not necessarily increased;
safeguards against output substitution vary by model and service; and
training can strengthen a product that competes for information, attention and future licensing.
The Delhi court borrowed Canada’s generous gateway to “research” without importing the full analytical discipline applied to fairness.
17. European and UK legislation points in a different direction
Comparative legislation shows that text and data mining is not normally resolved simply by relabelling all machine processing as research.
The EU Digital Single Market Directive created:
a mandatory scientific-research TDM exception for research organisations and cultural-heritage institutions with lawful access; and
a broader TDM exception applying to lawfully accessible content, but allowing rights owners to reserve their rights, including through machine-readable means for online works.
UK Section 29A is more limited and is directed to computational analysis for non-commercial research where the researcher has lawful access.
These systems reflect legislative choices about lawful access, commerciality, eligible institutions, rights reservation and retention. They demonstrate why updating construction has limits. Courts can apply old words to new technologies, but the creation of a comprehensive commercial TDM regime—with or without opt-outs or remuneration—is fundamentally a legislative policy choice.
18. Market harm: the judgment’s greatest economic weakness
The court found that ANI had produced no evidence of lost subscriptions, reduced market share or declining revenue. It emphasised that ANI syndicates news, while ChatGPT performs many functions and generally supplies summaries or snippets rather than ANI’s exact wording. It concluded that ChatGPT was not a substitute for ANI’s service.
At the interim stage, ANI’s lack of concrete evidence was a genuine weakness. But the court’s legal conclusion went beyond the evidence.
First, absence of proven present loss is not absence of potential market harm. The court’s own test included actual or potential damage.
Second, the relevant markets are broader than ANI’s present subscriber base:
licences for AI training;
licences for RAG and search;
premium feeds for AI assistants;
summary and alerting services;
advertising and referral traffic;
future ANI-developed AI products;
authenticated news and provenance services; and
derivative translation or analytical markets.
Third, ANI’s offer to license its content for USD 7.5 million was treated mainly as proof that damages were quantifiable. It was also evidence—though not conclusive evidence—of an asserted licensing market. OpenAI’s agreements with other news organisations further indicate that content licensing is commercially possible.
Fourth, “OpenAI does many things” does not answer whether it competes with ANI in the particular use at issue. A multifunction platform can substitute for a specialist service in one segment. When a user asks for the latest ANI news and receives a sufficient summary without visiting ANI or a subscriber, the relevant function may overlap directly.
Fifth, the court states that requiring multiple licences would make LLM development economically unviable. Even if empirically correct, transaction cost is not itself a statutory defence. Copyright often makes some unlicensed business models more expensive. The appropriate policy responses may include collective licensing, statutory licences, remuneration rights or carefully designed exceptions—not a judicially created entitlement to free inputs.
19. Public interest was applied asymmetrically
The court identified substantial benefits from AI in education, accessibility, research, translation, software development and knowledge dissemination. Those benefits are real and relevant, particularly to the balance of convenience.
But the public-interest analysis was one-sided. It did not give comparable attention to:
sustainable professional journalism;
incentives to investigate and verify information;
provenance and attribution;
the correction of inaccurate outputs;
concentration of informational power in large platforms;
barriers faced by smaller publishers;
the distribution of economic value between AI firms and creators;
the use of authoritative sources to improve commercial models; or
the possibility that unremunerated extraction undermines the information ecosystem on which AI depends.
Public interest can justify caution before disabling a widely used service. It should not automatically convert a private company’s input acquisition into fair dealing.
The court could have denied a broad injunction on proportionality and public-interest grounds without declaring the underlying copying fair.
20. The opt-out problem
In assessing the balance of convenience, the court placed substantial weight on ANI’s ability to block web crawlers and its failure to do so before the dispute. It noted that OpenAI had subsequently blocked ANI from training and RAG, and suggested that ANI’s subscribers could also block their sites.
This risks reversing the structure of copyright law.
Copyright ordinarily grants the owner an exclusive right. Permission is opt-in, not opt-out. A failure to deploy robots.txt, crawler blocks or technological restrictions does not normally constitute consent to reproduction, waiver or estoppel.
Opt-out mechanisms remain practically important. Rights owners who make clear machine-readable reservations strengthen their evidential and equitable position. But the absence of a technical block should not by itself legalise copying unless the statute expressly creates that consequence, as the EU has done for parts of its TDM regime.
Requiring ANI to ensure that every licensee and subscriber blocks AI crawlers is especially unrealistic. It transfers compliance costs to the rights owner and may conflict with the publisher’s legitimate need to remain accessible to ordinary search engines, customers and public audiences.
21. Evidential asymmetry and the burden placed on ANI
The court repeatedly relied on ANI’s failure to demonstrate:
which older ANI works entered training;
memorisation;
regurgitation;
use beyond training;
subscriber loss; and
precise market damage.
Those deficiencies matter. But virtually all decisive evidence about training datasets, model versions, filtering, retention, memorisation and safeguards is controlled by OpenAI.
The court accepted OpenAI documents as demonstrating that models were not designed to reproduce training material and concluded that use was limited to training because ANI had shown no other use. Design intent is not proof of system behaviour, and absence of plaintiff evidence is particularly difficult where the defendant controls the relevant records.
At trial, proportionate confidential discovery should address:
dataset manifests and source categories;
dates of collection;
model-specific cut-offs;
retention and deletion;
copies used for evaluation or post-training;
crawler identities and locations;
RAG indexes and cached copies;
memorisation testing;
output-suppression systems;
copyright opt-out processes; and
use of ANI material through licensees or third parties.
Without such evidence, broad findings about fair dealing are premature.
22. Remedies the court overlooked
The court was entitled to reject ANI’s requested blanket relief if it would effectively disable training, require uncertain deletion from model weights or grant final relief before trial.
But the choice was not necessarily between a global injunction and no protection. Possible narrower interim measures included:
requiring OpenAI to maintain its block on future training from ANI domains;
maintaining the stated RAG/search exclusion pending trial;
prohibiting outputs reproducing substantial protected passages;
preserving dataset and model-version evidence;
confidential disclosure of whether identified ANI works were included;
maintaining anti-regurgitation safeguards;
preventing false attribution to ANI;
providing correction and complaint channels; and
preserving referral links where ANI information is retrieved.
The court did not seriously evaluate these alternatives. That is particularly notable because OpenAI had already voluntarily blocked ANI’s site. Converting that voluntary measure into a temporary undertaking might have protected both parties without disrupting Indian AI development.
23. The most important mistakes and omissions
Taken together, the judgment’s principal vulnerabilities are these:
First, it conflates the court’s jurisdiction to hear a connected dispute with the territorial reach of Indian copyright law over copies made abroad.
Second, it treats an internal or secret technical process as “private use,” even where the objective purpose is a public commercial product.
Third, it interprets Canadian “research” jurisprudence without sufficiently accounting for the different wording of the Indian provision.
Fourth, it uses updating construction not merely to adapt a statutory word to new technology, but effectively to create a commercial AI-training exception.
Fifth, it overstates the requirement to compare works only “as a whole,” underprotecting qualitatively substantial extracts.
Sixth, it gives too much exculpatory significance to adversarial prompting.
Seventh, it inadequately separates model training, model memory, RAG, search and output generation.
Eighth, it makes questionable assumptions about ownership of interview quotations and translations.
Ninth, it treats public web accessibility as closely connected to lawful copying.
Tenth, its bespoke fairness test omits amount, alternatives, necessity, source lawfulness, safeguards and systematic scale.
Eleventh, it considers direct substitution but insufficiently examines licensing markets, referral traffic, future competition and market dilution.
Twelfth, it invokes Berne without rigorously testing whether the proposed exception is a “certain special case.”
Thirteenth, it treats the economic difficulty of licensing as an argument for fair dealing rather than a policy question for Parliament.
Fourteenth, it applies public interest predominantly from the perspective of AI users and developers while underweighting the sustainability of the knowledge and news ecosystem.
Fifteenth, it treats crawler opt-outs as legally significant in a manner that risks reversing copyright’s permission structure.
24. Consequences for AI makers in India
The decision gives AI developers a significant but unstable advantage.
Commercial entities can now argue that internal storage of publicly accessible literary material for model training is prima facie private research and fair dealing. Indian startups may rely on the decision when building models from public web content, especially where:
source works are factual;
no paywalls or access controls are circumvented;
the corpus is used only internally for training;
outputs do not reproduce substantial protected expression;
memorisation is controlled;
the model performs functions different from the source works; and
there is no strong evidence of market substitution.
But the judgment is not a blanket safe harbour. It is a single-judge interim ruling and remains open to appeal, trial evidence and later contrary authority. It does not conclusively authorise:
pirate-library acquisition;
circumvention of paywalls;
copying in breach of licences or confidentiality;
source-code or computer-program copying;
persistent commercial content libraries;
RAG systems that reproduce protected material;
verbatim or substantially similar outputs;
translations reproducing protected expression;
false attribution;
use of unpublished or highly confidential materials; or
training governed by foreign copyright law.
Responsible AI companies should therefore continue to maintain source provenance, model-specific dataset records, opt-out compliance, deletion policies, memorisation testing, output filters, complaint mechanisms and licensing strategies. Treating the judgment as permission to scrape everything publicly visible would be legally reckless.
25. Consequences for Indian rights owners
Rights owners will need more technically and economically sophisticated evidence.
A successful future claimant should identify:
specific protected works;
authorship, employment and assignment documents;
the model version allegedly trained on each work;
publication dates preceding the relevant training cut-off;
reproducible outputs;
non-adversarial as well as adversarial testing;
qualitative similarity in protectable expression;
evidence distinguishing model memory from RAG;
crawler logs and access evidence;
rights reservations;
licensing negotiations;
lost referral traffic;
subscriber or revenue effects;
existing AI-licensing comparators; and
realistic future markets.
Publishers should use machine-readable rights reservations, crawler policies, contractual restrictions and authentication where commercially feasible. Those mechanisms should not be legally mandatory, but after this judgment their absence may affect equitable arguments.
Collective solutions may become increasingly important. Individual licensing across millions of works is difficult, but that can support collective management, standard licences, registries and statutory remuneration—not necessarily free use.
Rights owners may also pursue claims outside the narrow training-copy theory, including RAG reproduction, communication to the public, translation rights, false attribution, passing off, trademark misuse, contract, database rights where applicable, confidentiality, privacy and competition law.
26. Consequences outside India
The judgment has no binding authority outside India. Its persuasive force will depend on the receiving jurisdiction’s statutory text.
In the United States, courts will continue to apply Section 107’s four factors, with source lawfulness, transformative purpose, market dilution and the evidential record playing major roles. Bartz, Kadrey and ROSS already show that “AI training” is not one legally homogeneous category.
In the European Union, lawful access and machine-readable rights reservations remain central to the commercial TDM framework. In the UK, the statutory non-commercial research limitation remains substantially narrower than the Delhi court’s interpretation. Germany’s GEMA judgment demonstrates a more rights-protective approach where memorised works are reproduced.
A company training in the United States or Europe cannot rely on an Indian fair-dealing ruling merely because its product is also offered in India. Conversely, the Delhi court’s jurisdiction reasoning means that foreign providers serving Indian users may face Indian claims for outputs, retrieval and other acts occurring in India even when training took place elsewhere.
The result may encourage some AI development activity in India, but it may also produce conflict-of-laws complexity. Global providers will need to map separately:
where source copies are acquired;
where datasets are stored;
where model training occurs;
where retrieval indexes operate;
where outputs are generated and received; and
which national rights reservations and exceptions apply.
The long-term pressure will be towards provenance standards, auditable reservations, collective licensing, differentiated access to trusted content and jurisdiction-specific model governance.
Conclusion
The Delhi High Court was probably correct to refuse ANI’s requested interim injunction. ANI’s principal output examples postdated the relevant training cut-offs, did not establish memorisation, and apparently did not reproduce substantial portions of the pleaded articles. A broad injunction against a globally used service would have been difficult to justify on that record.
The court could—and arguably should—have stopped there.
Its broader declaration that commercial model training constitutes “private or personal use, including research” and satisfies fair dealing rests on shakier foundations. The analysis stretches “private use,” understates the limiting language and history of Section 52, selectively imports US transformativeness, applies Canadian authority incompletely, overlooks central fairness factors, treats public accessibility and opt-out as too legally significant, and discounts licensing and market-development interests.
The most sustainable legal position is not that all AI training requires permission, nor that all publicly accessible content is free training material. A better framework would distinguish:
lawful from unlawful acquisition;
scientific from general commercial development;
temporary analytical copies from permanent content libraries;
training from retrieval;
model learning from regurgitation;
non-substitutive outputs from competing summaries;
factual information from original expression; and
responsible, auditable use from opaque extraction.
Ultimately, Parliament may need to create a clear Indian text-and-data-mining regime defining lawful access, rights reservation, transparency, retention, safeguards, commercial use and possible remuneration. Until then, ANI v OpenAIgives AI developers meaningful breathing room, but it does not provide the doctrinal certainty—or the unconditional licence—that its broadest language might initially suggest.
