Copyright Infringement and Fair Dealing in AI Training: The Indian Statutory Framework under Section 52

I. The Inevitable Collision Between Neural Training and Intellectual Property

The modern generative artificial intelligence revolution is built upon an insatiable appetite for data. To train large language models (LLMs) and diffusion architectures capable of synthesizing coherent text, complex computer code, and photo-realistic imagery, artificial intelligence laboratories scrape vast swathes of the public internet. In doing so, these systems ingest millions of copyrighted literary, artistic, journalistic, and musical works without the express authorization or licensing consent of original creators. This massive, unauthorized data ingestion has catalyzed a profound jurisprudential clash across global courts, with India emerging as a pivotal battleground.

Under Section 14 of the Indian Copyright Act, 1957, the copyright owner possesses the exclusive statutory right to reproduce the work in any material form, including the storing of it in any medium by electronic means. When an AI developer crawls digital news archives, books, and creative blogs to assemble a multi-terabyte training corpus, transient and permanent copies of these works are created within the developer's server infrastructure. Prima facie, this reproduction constitutes actionable infringement under Section 51 of the Act unless shielded by a recognized statutory defense.

The legal controversy deepens when neural models, having internalized the statistical syntax and expressive nuances of the training data, output verbatim or substantially similar extracts upon user prompting. Indian intellectual property jurisprudence is thus confronted with two distinct legal questions: Does the unauthorized ingestion and processing of copyrighted material for AI model training constitute copyright infringement? And does the commercial generation of synthetic works that compete directly with original creators violate statutory economic rights?

II. The Narrow Contours of Fair Dealing under Section 52(1)(a)

Unlike the open-ended, four-factor 'Fair Use' doctrine codified under Section 107 of the United States Copyright Act, India adheres to an exhaustive, rule-based 'Fair Dealing' framework under Section 52 of the Copyright Act, 1957. Indian courts have consistently affirmed that an unauthorized use of copyrighted material is permissible only if it falls strictly within the specific statutory categories enumerated by Parliament. Section 52(1)(a) provides fair dealing immunity exclusively for private or personal use (including research), criticism or review of that work, and reporting of current events.

Artificial intelligence corporations frequently argue that model training constitutes 'computational research' protected under Section 52(1)(a)(i). However, Indian judicial precedent imposes severe limitations on this defense. In landmark rulings such as *The Chancellor, Masters & Scholars of the University of Oxford v. Rameshwari Photocopy Service* (the Delhi University Photocopy Case) and *Academy of General Education, Manipal v. B. Malini Mallya*, the Supreme Court of India and the Delhi High Court established that the purpose and commercial character of the exploitation are decisive.

Commercial technology corporations cannot legitimately characterize the development of proprietary, multi-billion-dollar commercial AI products as 'private research'. Research conducted with the direct objective of commercial monetization, subscription licensing, or enterprise deployment falls outside the protective umbrella of Section 52(1)(a)(i). Consequently, without an express statutory amendment introducing a broad Text and Data Mining (TDM) exception, commercial AI training on copyrighted Indian works remains exposed to severe infringement liability.

III. The Absence of a Codified Text and Data Mining (TDM) Exemption in India

While foreign jurisdictions have enacted statutory adaptations—such as Japan's Article 30-4 of the Copyright Act permitting computational analysis without commercial restriction, and the European Union's Articles 3 and 4 of the Digital Single Market Directive establishing opt-out regimes for commercial TDM—the Indian Parliament has deliberately refrained from codifying a blanket TDM exemption. The Department for Promotion of Industry and Internal Trade (DPIIT) reiterated in 2024 that the existing framework of the Copyright Act 1957 is adequate and that unauthorized commercial ingestion requires licensing.

The technical mechanics of AI training render the absence of a TDM exception critical. Training does not merely involve reading text; it requires tokenization, embedding generation, vector transformation, and mathematical weight updates during backpropagation. Each technical stage involves transient electronic copying under Section 14(a)(i). While Section 52(1)(b) exempts transient and incidental electronic storage purely for transmission, this safe harbor is strictly conditioned on the reproduction being non-infringing at inception and having no independent commercial purpose.

Furthermore, where content creators deploy technical protection measures—such as `robots.txt` exclusion headers, CAPTCHAs, or paywall architectures—the circumvention of these barriers to extract training data triggers civil and criminal liabilities under Section 65A of the Copyright Act (protection of technological measures) and Section 43 of the Information Technology Act 2000 (unauthorized data extraction and extraction of computer databases).

IV. Judicial Precedents: The Landmark ANI v. OpenAI Litigation

The legal debate transitioned from academic scholarship to high-stakes judicial adjudication with the filing of *Asian News International (ANI) v. OpenAI Inc. & Ors.* before the Delhi High Court in late 2024. ANI, one of India's premier multimedia news agencies, instituted a civil suit alleging that OpenAI systematically scraped ANI's copyrighted news reports to train its GPT models without license or commercial remuneration, and that ChatGPT reproduced ANI's original reporting verbatim or with synthetic attributions.

The Delhi High Court's preliminary hearings in *ANI v. OpenAI* mark a watershed moment for Indian technology jurisprudence. The plaintiff contended that OpenAI's actions constitute direct copyright infringement under Section 51, unfair competition under common law, and trademark infringement by falsely attributing synthetic or hallucinated statements to ANI's brand. OpenAI argued that training occurred on public web servers outside India, that LLM training does not copy expression but learns statistical language patterns, and that outputs are novel algorithmic syntheses.

The jurisdictional and substantive findings in this litigation will define the future of generative AI in India. If the High Court rejects the fair dealing defense for commercial training, AI developers operating in India or utilizing Indian creative works will face mandatory collective licensing negotiations, statutory damages, and permanent injunctions restraining the commercial exploitation of models trained on unauthorized datasets.

V. Substantial Similarity and Output Infringement Analysis

Beyond the ingestion phase, infringement claims frequently center on the outputs generated by generative models. Under the doctrine of substantial similarity established in *R.G. Anand v. Delux Films* (1978 AIR 1613), the Supreme Court held that copyright does not protect abstract ideas, themes, or historical facts, but rather the unique expression of those ideas. An infringement occurs where the defendant's work is a substantial copy of the plaintiff's work, such that an ordinary observer would unmistakably identify the original expression.

When applied to Large Language Models, if a user issues an adversarial prompt and the model reconstructs copyrighted paragraphs verbatim or with trivial semantic substitutions (memorization or output leakage), both the user and the platform face joint tortfeasor liability. For diffusion models synthesizing visual art, where a model generates an image bearing the idiosyncratic artistic signature, composition, and visual expression of a living artist, the output constitutes an unauthorized derivative work under Section 14(a)(iv).

To insulate themselves from output infringement, AI developers must implement rigorous algorithmic filtering, retrieval-augmented verification (RAG), and post-generation similarity screening. Relying solely on disclaimer terms of service shifting all liability onto the end-user is increasingly rejected by courts where the underlying model architecture was intentionally trained to emulate proprietary expressive styles.

VI. Chambers Advisory: Structuring Legally Defensible AI Training Pipelines

Advocate Ritik Khurana Chambers advises leading artificial intelligence startups, media enterprises, and academic institutions on intellectual property compliance, training data licensing, and copyright litigation defense. For AI developers, our team conducts Comprehensive Data Provenance Audits, verifying whether training sets contain scraped copyrighted material, structuring bilateral content licensing agreements with Indian publishers, and implementing robust automated guardrails against output memorization.

For copyright owners, authors, and news corporations whose creative assets have been misappropriated by AI platforms, our litigation practice initiates high-court copyright infringement suits, files formal Cease-and-Desist notices under Section 55 of the Copyright Act, and petitions the court for dynamic blocking injunctions and account-of-profits decrees. As algorithmic synthesis challenges conventional notions of authorship, proactive copyright governance is indispensable to protecting corporate valuation.

Through these multi-layered legal interventions, our chambers ensures that innovation proceeds upon legally secure foundations, balancing technological acceleration with absolute respect for intellectual property rights.