Copyrightability of AI Model Weights, Latent Embeddings, and Training Datasets in India: Authorship and Fair Dealing Frontiers

I. I. The Technological Upheaval of Generative AI in Intellectual Property

The exponential emergence of Generative Artificial Intelligence (GenAI), Large Language Models (LLMs), and diffusion neural networks has precipitated the most profound legal crisis in the history of intellectual property law. AI models do not operate in a vacuum; training a modern frontier foundation model requires ingesting billions of web pages, digitized books, artistic illustrations, audio recordings, and proprietary software repositories.

This computational paradigm raises two existential legal dilemmas under Indian copyright jurisprudence: First, can the automated outputs, mathematical model weights, and latent vector embeddings of an AI system be copyrighted? Second, does the unauthorized scraping and ingestion of copyrighted human works to train an AI model constitute actionable copyright infringement under Section 51 of the Copyright Act, 1957, or is it shielded by statutory fair dealing?

With Indian technology enterprises and global AI developers establishing massive engineering clusters across Bengaluru, Hyderabad, and Delhi NCR, mastering the statutory contours of AI copyrightability is essential for corporate risk governance.

II. II. The Human Authorship Rule under Section 2(d) of the Copyright Act

The foundational pillar of Indian copyright law is that copyright can subsist exclusively in original works emanating from a **human author**. Section 2(d) of the Copyright Act, 1957 defines 'author' across categories of works: for literary, dramatic, musical, or artistic works, the author is the natural person who created the work.

While Section 2(d)(vi) introduces a specific definition for computer-generated works—stating that in relation to any literary, dramatic, musical or artistic work which is computer-generated, the author is 'the person who causes the work to be created'—Indian jurisprudence has consistently interpreted this provision to require active human intellectual engagement.

In 2020, the Indian Copyright Office briefly registered an AI painting app ('RAGHAV Artificial Intelligence Painting App') as a co-author of an artwork, before issuing a formal withdrawal notice recognizing that statutory law permits only legal persons (natural or juristic) to hold copyright. An autonomous AI model cannot be recognized as an author, nor can copyright subsist in raw, unedited AI generations.

III. III. Legal Status of AI Model Weights and Latent Embeddings

A foundational technical inquiry is the legal categorization of **Model Weights**—the trillions of floating-point numerical parameters derived through iterative gradient descent during the training of a neural network.

Under Indian intellectual property law, model weights occupy an ambiguous, contested frontier: (a) They cannot be protected as 'literary works' because raw floating-point numbers lack human expressive authorship; (b) They cannot be patented per se under Section 3(k) of the Patents Act as they represent abstract mathematical values; and (c) They do not qualify as 'databases' under Section 2(o) if they lack human creative selection and arrangement.

Consequently, leading enterprise AI labs must protect frontier model weights primarily through **Trade Secret Law and Technological Protection Measures (TPM)** under Section 65A of the Copyright Act. Circumventing digital locks or exfiltrating model weight tensors from secure cloud servers constitutes actionable cyber theft under Section 43/66 of the Information Technology Act and criminal breach of trust under Section 316 BNS.

IV. IV. Copyrightability of Training Datasets: The Eastern Book Company Doctrine

While raw data points cannot be copyrighted, curated training datasets represent immense commercial capital. Whether an AI training dataset is protected under Indian copyright law depends on the landmark Supreme Court ruling in *Eastern Book Company v. D.B. Modak* (2008 1 SCC 1).

In *EBC*, the Supreme Court rejected the outdated English common law doctrine of 'sweat of the brow' (which granted copyright based on mere labor and capital expenditure) and adopted the modern **Modicum of Creativity** standard formulated by the U.S. Supreme Court in *Feist Publications*.

The Supreme Court held that to claim copyright in a compilation or database under Section 2(o), the author must demonstrate that the selection, coordination, and arrangement of data involved a minimum degree of intellectual effort, judgment, and skill. A raw, uncurated scrape of web pages lacks creativity; conversely, a meticulously curated, annotated, and deduplicated instruction-tuning dataset satisfies *EBC* and is fully protected as a copyrighted literary work.

V. V. Web Scraping, AI Training, and the Section 52(1)(a) Fair Dealing Defense

When copyright holders initiate infringement suits against AI developers for ingesting their copyrighted books, news articles, and code repositories without a license, AI developers routinely raise the statutory defense of **Fair Dealing** under Section 52(1) of the Indian Copyright Act.

However, Indian fair dealing jurisprudence is significantly narrower than the open-ended 'Fair Use' doctrine of Section 107 of the U.S. Copyright Act. Section 52(1)(a) of the Indian Act enumerates an exhaustive, closed list of permitted fair dealing purposes: (i) private or personal use, including research; (ii) criticism or review; and (iii) the reporting of current events.

Commercial AI corporations that scrape copyrighted content to train commercial, revenue-generating generative AI models encounter immense legal difficulty fitting commercial training within 'private research'. In the absence of an explicit statutory Text and Data Mining (TDM) exception in Indian copyright law, unauthorized commercial scraping constitutes substantial prima facie copyright infringement under Section 51, exposing AI developers to severe damages and injunction liabilities.

VI. VI. Chambers Artificial Intelligence and Emerging IP Practice

Advocate Ritik Khurana Chambers provides specialized intellectual property advisory, dataset licensing negotiation, and copyright defense representation to artificial intelligence laboratories, foundation model developers, and digital publishers across India.

Our practice drafts AI Training Data Acquisition Agreements, structures proprietary model weight security charters, and audits algorithmic training pipelines to mitigate copyright infringement exposure. In high-stakes intellectual property litigation before the Delhi High Court IPD, our team delivers pioneering legal advocacy at the frontiers of artificial intelligence jurisprudence.