Commercial LLM Training as “Private Use, Including Research”: Section 52(1)(a)(i) Fair Dealing Shields AI Training (Prima Facie)

Case: ANI MEDIA PVT. LTD. v. OPEN AI OPCO LLC (2026 DHC 5900) — Delhi High Court
Date: 24-07-2026 — Coram: Amit Bansal, J.
Posture: Decision on interim injunction (Order XXXIX Rules 1 & 2 CPC), I.A. 45300/2024 in CS(COMM) 1028/2024

1) Introduction

The suit pits a major Indian news agency, ANI Media Pvt. Ltd. (“ANI”), against Open AI OpCo LLC (“Open AI”), alleging unauthorised use of ANI’s copyrighted news reports and interviews in two ways:

• Training claim: copying and storage of ANI’s content to train Open AI’s large language model(s) (“LLMs”) underlying ChatGPT; and
• Output/Reproduction claim: ChatGPT generating responses that allegedly reproduce ANI’s works (including by memorisation/regurgitation).

The interlocutory application required the Court to decide (at a prima facie level) whether interim injunctive relief should restrain training and/or outputs. The Court framed four issues, and decided them in an order that foregrounded jurisdiction and the mechanics of LLMs. The judgment is notable for:

• treating AI training as potentially falling within Section 52(1)(a)(i) fair dealing (“private or personal use, including research”);
• rejecting a per se bar on “commercial use” for Section 52(1)(a);
• requiring concrete proof of substantial reproduction for “output” infringement; and
• emphasising public interest and the feasibility of AI development when assessing balance of convenience.

Multiple intervenors participated, reflecting sector-wide stakes: publishers and news/music industry bodies largely supported ANI, while AI/start-up and policy groups broadly supported Open AI. Two amici curiae assisted the Court.

2) Summary of the Judgment

• Jurisdiction (Issue 4): The Court held it had territorial jurisdiction under Section 62(2) of the Copyright Act (ANI’s principal place of business in Delhi) and Section 20 CPC (services targeted and used in Delhi; alleged infringing outputs generated in Delhi). It rejected the server-location argument at the prima facie stage and declined to sever “training” from “output” in a way that would allow evasion via offshore servers.
• Output infringement (Issue 2): ANI’s exemplars largely post-dated the asserted training cut-off (April 2022 / April 2024), undermining a “memorisation” theory. The Court found no prima facie “substantial reproduction” applying R.G. Anand v. Deluxe Films and allied principles. It also treated “adversarial prompting” as relevant to the evidentiary weight of examples.
• Training/storage infringement and fair dealing (Issues 1 & 3): The Court interpreted Section 14(a)(i) broadly (digital “storing” is “reproduction”), but held the statutory structure makes Section 14 “subject to” Section 52. On a two-stage inquiry (purpose + fairness), it held (prima facie) that storage for LLM training fits “private use, including research” and is “fair dealing”. Commerciality was not treated as disqualifying for Section 52(1)(a).
• Interim injunction: Denied. Balance of convenience favoured Open AI; ANI’s alleged harm was compensable and it had opt-out technical tools; public interest in AI development weighed against restraint.

The Court repeatedly cautioned that findings are prima facie and will not bind final adjudication.

3) Analysis

A) Precedents Cited and Their Role

i) Jurisdiction and online/offshore infrastructure

• Intex Technologies v. Telefonaktiebolaget LM Ericsson: cited in the preface to underline the Indian approach of learning from foreign jurisprudence while harmonising with domestic statutes.
• Neetu Singh v. Telegram: central to rejecting the “server location” shield. The Court adopted its reasoning that cloud computing diminishes strict territoriality and that offshore server placement should not leave right-holders remediless.
• Blueberry Books v. Google India: relied on by Open AI to argue territoriality limits; the Court distinguished its relevance at the prima facie stage and used it to resist giving conclusive weight to one party’s unilateral assertions about where infringement occurs.
• Exphar SA v. Eupharma Laboratories and Dabur India Ltd. v. K.R. Industries: used (via intervenors/ANI) to reinforce that Section 62 creates an additional jurisdictional basis beyond CPC.
• Banyan Tree Holding v. A. Murali Krishna Reddy: invoked for “purposeful availment” in online jurisdiction; the Court accepted targeting/availability and transactions as supporting jurisdiction but refused to let “purposeful availment” become an instrument to immunise offshore conduct.
• Indian Performing Rights Society v. Sanjay Dalia: relied on by intervenors to support Section 62 jurisdictional principles.
• Getty Images v. Stability AI: mentioned by Prof. Scaria to note that plaintiffs abroad sometimes did not press training claims when training occurred offshore; the Court did not treat this as determinative for Indian statutory design and the chain-of-events approach.

ii) Copyrightability, originality, and “news as facts”

• Eastern Book Company v. D.B. Modak: pivotal for India’s rejection of “sweat of the brow” in favour of a “skill and judgment” standard; supported the Court’s emphasis that facts are not protected and that copyright in news is “thin” (expression only).
• Feist Publications v. Rural Telephone Service: used to reinforce the idea-expression dichotomy and that facts remain in the public domain; applied to news reporting to heighten the threshold for substantial similarity in expression.
• Walter v. Lane: relied upon by ANI to claim protection in reported speech; treated cautiously because it is “sweat of the brow” era reasoning, and was already moderated by Eastern Book Company v. D.B. Modak.
• Akuate Internet Services v. Star India: used to reject monopolisation of factual match information; reinforced “no copyright in facts/information” in public.
• Syndicate of the Press of the University of Cambridge v. B.D. Bhandari: used for “merger doctrine” and the trajectory from “sweat of the brow” to “modicum of creativity”; influenced the Court’s approach to thin protection where expression options are limited.

iii) Substantial reproduction and how to compare works

• R.G. Anand v. Deluxe Films: the leading Indian test on infringement; used to require substantial/material copying of expression and to compare works holistically, not by selective extraction.
• Star India v. Leo Burnett and Bikramjeet Singh Bhullar v. Yash Raj Films: used to reject dissection/compartmentalisation and insist on “work as a whole” comparison.

iv) AI/LLM litigation and foreign fair use/fair dealing analogies

• Andrea Bartz v. Anthropic: relied on (by Open AI/intervenors and by the Court) to illustrate that meaningful regurgitation is not readily achievable; also used to differentiate lawful acquisition vs “shadow libraries” and to discuss transformative character of training.
• GEMA v. Open AI: considered but distinguished: it involved memorised German lyrics reproduced via non-adversarial prompts; here, ANI’s examples post-dated training cut-offs and were elicited via “exactly”-type prompting.
• The Associated Press v. Meltwater U.S. Holdings, Inc., Infopaq International v. Danke Dagblades Forening, and Newspaper Licensing Agency v. Meltwater Holding: treated as limited help because those cases featured verbatim copying/excerpting; the Court found ANI had not shown comparable substantial reproduction in outputs.
• The Advance Local Media LLC et al. v. Cohere Inc.: noted as a motion-to-dismiss context with extensive alleged verbatim copying; not treated as analogous on facts at the interim stage.
• Positive Black Talk v. Cash Money Records: discounted as song-lyric similarity analysis is not readily transferable to news.
• Richard Kadrey v. Meta Platforms and Thomson Reuters v. Ross Intelligence: invoked by rights-holder-side intervenors to argue against “non-expressive use” framing; the Court nonetheless grounded its decision in Indian statutory architecture and Section 52.
• Authors Guild v. Google: used to support that transformative/limited display uses can remain fair despite commerciality, and that market substitution is a key lens.
• Andy Warhol v. Lynn Goldsmith: raised by intervenors to contest an overbroad “transformative” claim; the Court ultimately did not import a US four-factor test as binding, instead formulating an India-facing fairness inquiry aligned with Berne’s three-step logic.

v) Section 52 methodology and statutory interpretation

• Chancellor Masters & Scholars of The University of Oxford v. Narendera Publishing and Syndicate of the Press of the University of Cambridge v. B.D. Bhandari: relied on to argue liberal interpretation of Section 52 and the “transformative/derivative” character concept within Indian fair dealing analysis.
• Chancellor, Masters & Scholars of the University of Oxford v. Rameshwari Photocopy Services and The Chancellor, Masters & Scholars of University of Oxford v. Rameshwari Photocopy Services (DB): used for the proposition that Section 52 acts are not infringement; and to caution that US fair use tests are not automatically transferable to India.
• Rupendra Kashyap v. Jiwan Publishing House and TIPS Industries v. Wynk Music: relied upon by ANI/intervenors to argue commercial exploitation defeats Section 52(1)(a)(i); the Court distinguished them as cases where the defendant was commercially disseminating the copyrighted work itself, unlike training where the work is not made available to the public.
• Super Cassettes Industries Ltd. v. Hamar Television Network Pvt. Ltd. Network: used to show commerciality does not automatically negate fairness.
• Academy of General Education, Manipal v. B. Malini Mallya: used to interpret “private use” broadly, allowing institutions to invoke Section 52.
• Blackwood v. A.N. Parasuraman and ESPN Star Sports v. Global Broadcast News: canvassed for “fair dealing” factors like quantum, competition, publication status, and importance of what is taken.
• State (Through CBI/ New Delhi) v. S.J. Choudhary: foundational for the “doctrine of updating construction” applied to interpret “research” in light of AI-era realities.
• Shailesh Dhairyawan v. Mohan Balkrishna Lulla and State of Maharashtra v. Praful B. Desai: cited by Open AI for purposive interpretation (though the Court’s key interpretive move was updating construction and statutory structure).
• Society of Composers, Authors and Music Publishers of Canada v. Bell Canada and CCH Canadian v. Law Society of Upper Canada: used to treat fair dealing as a “user’s right” and to interpret “research” liberally, including in commercial contexts.

vi) Interim injunction and public interest

• Zydus Lifesciences v. E.R. Squibb, F-Hoffman-La Roche AG v. NATCO Pharma, and Astrazeneca AB and Ors. v. Intas Pharmaceuticals Limited: cited to emphasise public interest as a material consideration in interim relief for IP disputes.

B) Legal Reasoning

i) Territorial jurisdiction despite overseas servers

The Court treated Section 62(2) of the Copyright Act as decisive at the threshold: if the plaintiff carries on business within Delhi, suit can be instituted there. It supplemented this with Section 20 CPC considerations (targeting and use within Delhi).

The more innovative move lay in addressing the “training takes place abroad” objection. Without asserting extraterritorial application as a general rule, the Court treated offshore storage/training as a “terminal step” of a chain that begins with access/scraping from India and culminates in outputs in India. This “chain of events” framing pre-empted a loophole: relocating the last link to foreign servers to escape Indian copyright claims.

ii) Output claim: evidentiary discipline, substantial similarity, and timing

The Court effectively imposed three filters before allowing an “output infringement” theory to proceed at the interim stage:

• Training cut-off mismatch: ANI’s exemplars were published after GPT-4/GPT-4o training cut-offs, weakening the inference of memorisation.
• RAG possibility: Where outputs referenced fresh information, the Court inferred these could be “live links” consistent with retrieval-augmented generation, not stored memorised text. Importantly, it noted RAG-based infringement was not pleaded, signalling pleading discipline in AI cases.
• Substantial reproduction test: Applying R.G. Anand v. Deluxe Films, the Court looked for copying of “form, manner, arrangement and expression”, not facts. It rejected “part-extraction” comparisons via Star India v. Leo Burnett and Bikramjeet Singh Bhullar v. Yash Raj Films.

On the Neeraj Chopra mother interview example, the Court additionally invoked Section 17(cc) to suggest quotes/speech copyright may vest in the speaker, and absent assignment, ANI could not prima facie claim exclusive rights over the quoted speech or translation rights.

iii) Training claim: storage is “reproduction”, but Section 52 shapes the boundary of infringement

The Court read Section 14(a)(i) literally and expansively: “reproduction” includes “storing” in any electronic medium. It rejected arguments that only “human-perceivable” copies count, and held neither “temporary” nor “permanent” storage is textually excluded from Section 14(a)(i).

However, it treated Section 52 as integral to defining the infringement boundary (not merely a narrow exception), drawing support from Syndicate of the Press of the University of Cambridge v. B.D. Bhandari and Chancellor, Masters & Scholars of the University of Oxford v. Rameshwari Photocopy Services. This structural move is key: Section 14 rights are “subject to” the Act, and Section 52 declares specified acts “shall not constitute infringement”.

iv) Section 52(1)(a)(i): “commercial” is not a statutory disqualifier; “private use” and “research” are updated to AI training

The Court’s Section 52 analysis is two-step: (1) Purpose test and (2) Fairness test.

On the Explanation to Section 52(1)(a), the Court held the “not itself being an infringing copy” qualifier applies to “incidental storage of any computer programme” (punctuation/commas), not to all stored works. Factually, it also noted ANI’s content was freely accessible; there was no allegation of paywall circumvention.

Most significantly, the Court used the doctrine of updating construction (via State (Through CBI/ New Delhi) v. S.J. Choudhary) to interpret “research” and “private use” for the AI era:

• “Private” includes corporate/private entities, not merely individuals, and should not be collapsed into “personal”.
• “Research” includes machine learning: training is a form of study/investigation that produces knowledge; although the “learner” is a machine, it is ultimately for human benefit.
• The training dataset is not disclosed to the public; the process is internal, thus “private”.

v) Fairness test tailored to AI training and Berne-style balancing

Rather than importing a rigid four-factor fair use test, the Court formulated fairness factors suited to AI training and aligned them with Berne’s three-step logic (no conflict with normal exploitation; no unreasonable prejudice to legitimate interests), namely:

• Is the use limited to training (not distribution of the works)?
• Does it cause economic competition/actual or likely market harm?
• Does the technology serve public interest?

Applying these, the Court found: Open AI’s use was limited to training; ANI did not show market substitution or loss of subscribers; and public interest in AI innovation and access to information strongly supported non-restraint. It also emphasised opt-out/crawler-blocking possibilities and Open AI’s asserted blocking of ANI’s site (including for “ChatGPT search”).

C) Impact

• First major Indian interim precedent on LLM training and Section 52(1)(a)(i): The judgment signals that training-stage copying/storage can be fair dealing as “private use, including research”, even for commercial AI developers, subject to evidence and final trial findings.
• Commerciality is not per se fatal under Section 52(1)(a): By contrasting Section 52(1)(a) with provisions that explicitly exclude commercial use, the Court introduces a text-and-structure method likely to influence future “fair dealing” disputes beyond AI (e.g., data analytics, internal enterprise knowledge extraction).
• Pleading and proof discipline for “output” infringement: Plaintiffs may need (a) examples within the model’s training horizon; (b) clear substantial similarity in expression; and (c) careful pleading where RAG/search-like features are implicated.
• Quotes/speeches and translation rights: The Section 17(cc) discussion flags that news agencies may not automatically own copyright in the underlying speech they report; assignments and contractual chains may become central in AI-news disputes.
• Jurisdiction in cloud architecture: The “chain of events” approach and reliance on Neetu Singh v. Telegram strengthens Indian courts’ willingness to hear disputes despite foreign servers, a principle likely to recur across online infringement claims.
• Policy and industry effect: The Court’s emphasis on opt-out tools (crawler blockers/paywalls/robots.txt) and public interest may incentivise technical self-help and licensing negotiations rather than immediate injunctive litigation—at least at the interim stage.

4) Complex Concepts Simplified

• LLM “training” vs “output”: Training is the internal process of learning statistical patterns from vast text; output is what users see as answers. The Court treated them as linked but evaluated infringement differently for each.
• Tokenisation / vectors: Text is broken into tokens and represented numerically; the Court accepted that storage of the work (even digitally) can be “reproduction” under Section 14(a)(i), but asked whether Section 52 permits that reproduction for certain purposes.
• Memorisation / regurgitation: The claim that the model stores passages verbatim and reproduces them later. The Court found ANI’s examples did not establish memorisation because they post-dated training cut-offs and were extracted via “exactly”-style prompts.
• Adversarial prompting: Prompts engineered to force the model into reproducing specific text. The Court treated such prompting as weakening an inference of ordinary regurgitation.
• RAG (Retrieval-Augmented Generation): The model fetches fresh information from the web and then drafts an answer. The Court suspected some examples reflected RAG/search-like behaviour, and noted RAG-based infringement was not pleaded.
• Idea-expression dichotomy & “news as facts”: Facts are free for all; only the author’s expression is protected. For news, protection is “thin” because many reports inevitably describe the same event.
• Merger doctrine: Where an idea can be expressed only in very limited ways, the expression “merges” with the idea and protection narrows.
• Fair dealing (India) vs fair use (US): India’s Section 52 is purpose-linked (e.g., private use/research), whereas US fair use is open-textured with a statutory four-factor framework. The Court did not treat the US test as binding, but adopted a fairness analysis compatible with Indian text and Berne balancing.
• Updating construction: Courts interpret old statutory words to apply to new technologies while remaining faithful to legislative purpose—here, “research” was extended to machine learning.

5) Conclusion

ANI MEDIA PVT. LTD. v. OPEN AI OPCO LLC is a landmark interim ruling situating AI training within Indian copyright doctrine without legislative amendment. It makes three doctrinal moves with lasting significance:

• Territorial reach: offshore servers do not automatically defeat Indian jurisdiction where the plaintiff sues under Section 62(2) and the service targets/operates in India.
• Outputs need proof: infringement by ChatGPT outputs requires prima facie substantial reproduction of protected expression, not merely factual overlap or selectively extracted similarities.
• Training can be fair dealing: notwithstanding that storage is reproduction under Section 14(a)(i), LLM training is, prima facie, “private use, including research” under Section 52(1)(a)(i), and commerciality does not per se bar the defence.

While expressly non-final, the judgment sets a strong interim baseline: plaintiffs must plead and prove model-specific reproduction risks, whereas AI developers may—on evidence—frame training as research-oriented private dealing, especially where no market substitution is shown and public interest weighs against injunctive restraint.