THE SILICON SCRAPER: HOW GENERATIVE AI IS REWRITING INDIA’S INTELLECTUAL PROPERTY REFORMS
The Delhi High Court Confronts the Input-Output Paradox in ANI v. OpenAI
From Blanket Licensing to Output Remedies: Designing a Modern Framework for India's Creative Economy
By Legal Editor
New Delhi: July 06, 2026:
The rapid growth of generative artificial intelligence (GenAI) has triggered an unprecedented intersection between deep learning technology and traditional intellectual property regimes. Globally, creative industries are actively resisting the unauthorized scraping of text, imagery, and code by tech conglomerates. In India, this legal battleground found its focal point in the landmark case of , currently pending before the Delhi High Court under case registration CS(COMM) 1028/2024. The dispute exposes deep fissures within the statutory framework of the , revealing that centuries-old legal doctrines designed to regulate human-scale reproduction are ill-equipped to address machine learning architectures that operate on a structural scale.
Article 14 of the Indian Constitution embodies the principle of equality before the law and equal protection of the laws, ensuring that every individual receives fair and non-arbitrary treatment. The Supreme Court has consistently interpreted this fundamental right through landmark judgments addressing discrimination, arbitrariness, reasonable classification, and constitutional fairness. To understand the legal significance and evolving judicial interpretation of this fundamental right, read this detailed guide on Article 14 of the Constitution of India.
The Legal Crux: Reading vs. Plagiarizing in the AI Era
Historically, copyright jurisprudence relies heavily on a binary distinction regarding the utilization of protected works. The first category involves utilizing a work for its expressive value. This encompasses standard commercial actions like translating a text, adapting a novel into a cinematic screenplay, or distributing unauthorized sequels. Because these acts directly exploit the author’s unique market expression, they strictly require a license.
The second category encompasses utilizing a work for non-expressive or transformative purposes. This includes actions like critically reviewing a text, referencing an excerpt for classroom instruction, or analyzing structural patterns to create an entirely new, non-infringing work. Under the , such activities are legally shielded under the doctrine of "fair dealing" via Section 52(1)(a). For decades, intellectual property disputes were litigated strictly along this dividing boundary.
Large Language Models (LLMs) break this binary paradigm. When a machine learning algorithm ingests millions of news articles, it does not "consume" or display the literal expression for its aesthetic value. Instead, it processes the text as raw numerical data, extracting underlying statistical, syntactic, and semantic patterns. Under standard definitions, processing open-source or publicly available web content to deduce abstract patterns aligns closely with non-expressive, fair use.
However, the paradox emerges at the output stage: by processing an author’s entire corpus to map their unique writing patterns, the trained AI can subsequently generate fresh articles, prose, or code that directly mimics that author’s stylistic voice. It can seamlessly produce highly competitive substitute material in real-time. Consequently, the AI exploits protected creative works at the input stage as mere mathematical data to build a commercial engine that systematically competes against the original human authors at the output stage. The current law offers no seamless solution for this dual reality because the was established in a human-scale era where using a work as an analytical reference point and reproducing its commercial expression were mutually exclusive events.
Evaluating the State's Response: The "One Nation, One License" Fallacy
Recognizing that existing statutory exemptions fail to address systemic data scraping, the Department for Promotion of Industry and Internal Trade (DPIIT) issued a comprehensive policy intervention. In December 2025, a government-appointed committee published Part One of its , advancing a centralized regulatory model titled "One Nation One License One Payment."
+------------------------------------------------------------------------+
| DPIIT's Proposed Blanket Licensing Architecture |
+------------------------------------------------------------------------+
| |
| [ Copyrighted Works] ---> [ Mandatory Ingestion by GenAI Models] |
| | |
| v |
| [ Centralized Royalty Pool (CRCAT) ] |
| | |
| +------------------------------+------------------------------+
| | | |
| v v v
| [ Media Aggregators ] [ Large Publishers ] [ Individual Creators ]
| (Receives Bulk Share) (Receives Bulk Share) (Receives Negligible Pennies)
| |
+------------------------------------------------------------------------+
The DPIIT framework attempts to establish a mandatory, retroactive blanket licensing scheme. Under this initiative, AI developers receive a statutory right to ingest any copyrighted material for model training, provided they route a percentage of their commercial revenue into a centralized state fund managed by a newly proposed collecting society—the Copyright Royalties Collective for AI Training (CRCAT). While this intervention rightly acknowledges that the current legal framework is outdated, its operational execution remains fundamentally flawed.
First, a domestic blanket licensing tax cannot realistically track or capture AI training operations that occur beyond India's borders. Because machine learning datasets are aggregated and deployed across distributed cloud networks globally, local statutory mandates face immediate enforcement limitations.
Second, the structural design of the revenue distribution mechanism fails to support individual creators. When royalties are pooled centrally and divided mathematically based on total content volume, institutional media aggregators and massive publishing conglomerates inevitably capture the bulk of the financial layout. The individual freelance journalist, independent photographer, or novelist receives negligible compensation. Instead of fostering organic human creativity, this system functions as an administrative tax that validates algorithmic scraping while leaving individual human creators economically marginalized.
Reframing the Target: Shifting Liability from Model Input to Generative Output
The core limitation of the DPIIT working paper lies in its foundational premise: it assumes that copyright infringement occurs the moment a piece of data is inputted into a training model. In reality, the act of training a model on data does not cause immediate economic harm. An LLM that ingests a news database but remains contained within a private research laboratory inflicts zero commercial damage on publishers.
The economic injury occurs exclusively at the output stage, when the commercial model generates alternative reports that directly substitute for the human work it learned from. This output generation threatens journalists, novelists, and creators by undercutting the target market.
By focusing regulatory energy entirely on licensing data inputs, policymakers are attempting to regulate an abstract, borderless process where the economic harm remains highly speculative. Meanwhile, the output stage—where actual market substitution and plagiarism materialize—remains unaddressed. This regulatory approach is comparable to taxing citizens for reading library books to curb plagiarism, rather than addressing the instances of plagiarism themselves.
Path Forward: Judicial Injunctions vs. Parliamentary Statutory Reform
While individual litigations like provide temporary clarity regarding interim injunctions and the deployment of robots.txt exclusion protocols, courts face structural limitations when managing systemic technological shifts. A judicial body can only issue binary rulings: it must either declare an action an outright copyright infringement or protect it entirely under fair dealing. Because AI training exhibits qualities of both, forcing a court to apply outdated definitions will result in inconsistent legal precedents.
True resolution requires deliberate legislative intervention. Parliament must introduce updated statutory exceptions, modelled after international frameworks like the European Union's Digital Single Market Directive, which pair explicit text and data mining (TDM) exceptions with robust opt-out rights for creators. More importantly, India's legislative approach must pivot toward establishing strict output-stage liability. When an AI platform generates material that directly reproduces substantial protected expressions or misattributes fabricated information to credible creators, liability must be enforced directly. Protecting the creative economy requires regulating where the commercial impact lands, rather than implementing broad upfront data taxes that fail to reward human ingenuity.
Detailed FAQ: Navigating AI and Indian Copyright Jurisprudence
Index of Legal Questions
What is the central legal dispute in the case of ANI v. OpenAI?
Does the Indian Copyright Act, 1957 contain provisions for text and data mining (TDM)?
Can AI systems claim independent authorship or copyright ownership under current Indian law?
What is the "One Nation One License One Payment" framework proposed by the DPIIT?
Why do digital rights groups argue that mandatory blanket licensing harms individual creators?
How does the doctrine of "Fair Dealing" under Section 52(1)(a) apply to machine learning?
What is the difference between input-stage infringement and output-stage liability?
1. What is the central legal dispute in the case of ANI v. OpenAI?
The news agency in the Delhi High Court, alleging unauthorized scraping and storage of its proprietary news data to train ChatGPT. ANI argues that OpenAI directly consumes its journalistic output without an active license, occasionally reproducing content verbatim or generating false news reports inaccurately attributed to ANI. OpenAI contends that its training utilizes publicly available web data for non-expressive pattern extraction, which it argues does not constitute standard copyright infringement.
2. Does the Indian Copyright Act, 1957 contain provisions for text and data mining (TDM)?
No. Unlike the European Union or Japan, the contains no explicit statutory exception for text and data mining (TDM). Because the law contains no clear provisions for automated data analysis, courts must evaluate large-scale machine ingestion using standard fair dealing provisions, which were not originally drafted to accommodate automated web scraping.
3. Can AI systems claim independent authorship or copyright ownership under current Indian law?
No. Section 2(d) of the explicitly requires an "author" to be a natural human person. Because AI systems lack an independent legal personality, purely machine-generated outputs cannot claim copyright protection. Legal protection is extended only to collaborative works where a human creator exercises definitive creative choice, curation, or transformation over the AI-generated material.
4. What is the "One Nation One License One Payment" framework proposed by the DPIIT?
Developed by a specialized committee under the , this proposed model introduces a mandatory, retroactive blanket licensing framework for India. Under this system, GenAI developers receive statutory permission to ingest copyrighted works for model training. In return, they must pay a set percentage of their commercial revenue into a centralized collective, which redistributes those royalties to registered rights-holders.
5. Why do digital rights groups argue that mandatory blanket licensing harms individual creators?
Civil society organizations point out that by denying them the right to opt out of AI training datasets. Furthermore, because centralized royalty allocation schemes prioritize volume, financial payouts tend to favor massive media publishing groups over individual freelance journalists, writers, and artists. The model also complicates matters for Free and Open-Source Software (FOSS) developers by overriding standard copyleft licensing terms.
6. How does the doctrine of "Fair Dealing" under Section 52(1)(a) apply to machine learning?
Section 52(1)(a) permits the unauthorized use of copyrighted material for specific activities, including private research, criticism, review, or current affairs reporting. While AI firms argue that training models to extract abstract linguistic or structural data aligns with analytical research, copyright owners counter that systematic commercial data ingestion to develop a competing generative software program extends beyond the scope of traditional fair dealing exceptions.
7. What is the difference between input-stage infringement and output-stage liability?
Input-stage infringement focuses entirely on whether copying occurs the moment an AI system ingests copyrighted data into its training repository. Output-stage liability focuses on the generated product, asserting that legal and financial liabilities should apply only when the AI model produces an output that closely duplicates protected material or directly undercuts the original creator's market share.