Introduction
Artificial intelligence has created a copyright issue that is not about simple copying. Usually, when you copy a book, photograph or article, you can tell. The copyright holder can identify the original creation and contrast it with the alleged infringing copy. The process gets more complicated with generative AI, because the controversial action can happen a long time before the final product is created. In the context of Indian copyright jurisprudence, this judicial analysis reviews the statutory thresholds and legal principles governing Ai Copyright Ruling.
Artificial intelligence systems are trained on massive datasets that may include copyrighted literary, artistic, musical, and audiovisual works. But the creator of an original work often has no way of knowing whether that work was part of a specific training dataset or how it was processed during the model training procedure.
This raises a fundamental question that sits at the core of the debate over AI and intellectual property: If the owner of the copyright cannot see what happens with their work in the AI system, how can one establish infringement?
The question isn’t whether training AI infringes copyright. This issue relates to traceability of copyright.
Copyright Questions in AI Training
The first challenge is to understand the technical processes that occur before an AI system generates a response, image, or synthetic output. Training may require collecting, storing, processing and examining huge quantities of data. These different phases can raise separate questions of reproduction and the legal exceptions that are available from a copyright law perspective.
Under Indian copyright law, the Copyright Act, 1957 protects original literary, dramatic, musical, and artistic works, cinematograph films, and sound recordings. Section 14 enumerates the exclusive statutory rights of copyright owners, Section 51 defines infringement, and Section 52 lays down specific acts that do not constitute infringement.
The problem is that these rules weren’t explicitly written for training language models or generative AI.
The legal inquiry must therefore assess what material was ingested, how it was processed, what its intended commercial or transformative purpose was, and whether the utilization falls within a recognized statutory exception.
ANI vs OpenAI: The Indian Copyright Dilemma
This matter is now in the hands of the Indian judiciary.
In ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, ANI claimed that its copyrighted news articles were used without authorization to train OpenAI’s language models. ANI further alleged that OpenAI’s chatbot, ChatGPT, could reproduce its proprietary journalistic reporting.
On July 24, 2026, the Delhi High Court examined whether OpenAI’s ingestion of ANI’s content for AI model training constituted copyright infringement, or whether such use was protected under the fair dealing provisions of Section 52(1)(a) of the Copyright Act, 1957.
The Court considered the purpose and fairness of the use. At the preliminary stage, the Court held that the use of ANI’s literary works to train the LLMs underlying ChatGPT fell within the scope of Section 52(1)(a). Accordingly, it decided, on a preliminary basis, that the storage connected to training did not constitute an infringement. The Court found that ANI failed to adequately demonstrate that ChatGPT’s responses memorized and substantially reproduced its works.
The case, therefore, carries profound significance beyond the immediate question of fair dealing. It highlights a critical evidentiary barrier: proving copyright ownership does not automatically bridge the gap to proving that a specific work was ingested into an AI training set.
Proprietary Material, Software, and Artificial Intelligence Training
You can’t treat an AI system as one, single copyright subject.
An AI system encompasses diverse intellectual property layers: backend code, algorithmic architectures, curated training databases, and compilation pipelines. Its training inputs may span literary texts, photographs, musical compositions, audiovisual footage, and source code—each carrying independent copyright interests. Finally, the downstream generated output introduces another distinct layer of copyright analysis.
Under Section 2(o) of the Copyright Act, 1957, “literary works” explicitly include computer programs, tables, and compilations, governing both software code and database protection in India. Crucially, copyright protects original expression, not ideas, mathematical procedures, or functional methods.
This foundational distinction was articulated in R.G. Anand v. Deluxe Films (1978), where the Supreme Court held that copyright protects the expressive form of an idea and not the idea itself.
That distinction is critical in the context of machine learning. A model can extract semantic patterns, statistical correlations, and linguistic structures from a work without necessarily reproducing the underlying protected expression in its original form. A separate question of infringement arises if the deployed model subsequently generates outputs that reproduce a substantial portion of the original expression.
Therefore, the learning from data and the reproduction of protected expression cannot be considered as the same legal actions.
Public Access Is No Guarantee of Copyright Exemption
Another challenge is the availability of resources online.
Just because a work is available free of charge does not mean that it does not have copyright. An image posted on a publicly accessible site does not automatically become part of the public domain. An online newspaper article remains fully protected by copyright regardless of whether it is freely accessible without a paywall.
The importance of this distinction was shown in ANI v. OpenAI. The Court noted that ANI’s content was obtained by OpenAI from publicly available information on ANI’s website and took this into account when assessing the Section 52 defence.
However, public accessibility and copyright authorization are not synonymous. The question is whether the specific acts of collection, storage and training are protected by an applicable legal exemption.
The Evidentiary Dilemma: Who Holds the Proof?
Consider an illustrative scenario: a photographer believes her copyrighted portfolio was used to train a commercial AI image generator. She can easily demonstrate legal ownership over her photographs. She can verify that they were published on the web. She may even point to stylistic similarities between her portfolio and AI-generated outputs.
But can she prove her images were in the training set?
The AI developer could manage the relevant datasets, records, logs and technical documentation. This results in an unusual imbalance: The copyright belongs to the creator, but only the technology company possesses the internal datasets and logs needed to prove whether ingestion actually occurred.
We term this the AI Copyright Proof Deficiency. That is not to say that there is no substantial copyright protection. There is a fundamental distinction between holding a legal entitlement and possessing the factual records necessary to establish a violation of it.
This dynamic mirrors the classic problem of evidentiary asymmetry recognized in Indian procedural law. Under Section 106 of the Indian Evidence Act, 1872 (now Section 109 of the Bharatiya Sakshya Adhiniyam, 2023), when any fact is especially within the knowledge of a party, the burden of proving that fact shifts to that person. While Section 106 cannot relieve a copyright plaintiff of the initial burden of establishing a prima facie case, applying it in generative AI litigation suggests that once a creator demonstrates a plausible factual nexus, such as unique stylistic mimicry, prompt leakage, or verified crawling of their domain, the AI developer should bear an evidentiary burden of production to show whether that specific material was ingested or excluded from its training corpus.
Moving from Copyright Ownership to Copyright Provenance
One possibility to consider is to develop the idea of copyright provenance. Provenance is the origin and historical setting of a material. Simply put, within the framework of AI, it means maintaining sufficient knowledge about the provenance of important training data and how it was incorporated in the training process.
This does not require AI companies to publicly disclose every file, URL or data element used in training. This approach can uncover proprietary information and create significant compliance costs.
Developers are able to maintain adequate internal documentation of important datasets, licensing agreements and known copyright limitations. Such documents might be held in confidence until a real dispute occurs that requires their disclosure. The objective is straightforward: to create an auditable documentary record where none currently exists.
Should Copyright Owners Possess a “Right to Know”?
This suggests a possible future rule: a limited right to determine whether a copyrighted work has been used in an AI training system. This must not be interpreted as a currently established broad statutory entitlement under Indian copyright legislation. That could be a step forward in regulation.
If a copyright owner demonstrates a valid connection between a protected work and a particular AI system, the developer may be required to participate in a regulated disclosure process. It would not be about granting creators unfettered access to proprietary datasets. The aim is to prevent the near-impossibility of tracing genuine copyright claims resulting from technological secrecy.
This concept of provenance aligns closely with international regulatory developments, notably Article 53 of the European Union’s Artificial Intelligence Act (Regulation (EU) 2024/1689). Under the EU framework, general-purpose AI model providers are required to implement copyright compliance policies and publish a “sufficiently detailed summary” about the training content ingested by their systems. The EU approach demonstrates that mandating training data provenance does not require revealing proprietary model weights or compromising trade secrets. Instead, it establishes an auditable documentary trail that enables creators to identify commercial exploitation while providing AI developers with clear, standardized compliance benchmarks.
A Suggested Framework: AI Copyright Provenance Obligation
India should consider the introduction of an AI Copyright Provenance Obligation for large commercial AI developers. This obligation could rest upon three balanced operational pillars:
- Internal Record-Keeping: Commercial AI developers must maintain structured, confidential logs detailing the major datasets, URLs, and scraping repositories utilized during model pre-training.
- Regulated Inquiry Mechanism: Copyright owners who demonstrate a credible prima facie nexus can submit standardized, verified inquiries regarding specific catalogued works.
- Adverse Evidentiary Presumptions: If a claimant establishes a plausible factual connection and the developer failed to preserve records that should reasonably have been maintained, courts may draw appropriate adverse inferences regarding data ingestion.
A failure to maintain records would not constitute an independent statutory infringement in and of itself. Rather, it recognizes that when one party holds exclusive custody over the relevant operational evidence, an unexplained absence of documentation should not work to the prejudice of the creator. While the legal burden of proof cannot be eliminated entirely, procedural fairness requires balanced evidentiary presumptions.
In addition, such a system must protect real technological progress.
It should not be possible for copyright owners to claim infringement simply because a work was publicly available, or because an AI-generated output resembles a work. Similarity alone does not in itself prove duplication. Similarly, mandating the disclosure of all training data could expose sensitive information and be cost-prohibitive.
A more effective strategy would therefore have a limited evidentiary framework. First the claimant must prove ownership, and a plausible factual basis for claiming the work was used. Only then the developer’s documentation on the origin is relevant.
Licensing Is Only Part of the Picture: The DPIIT Model
India has begun talking about wider policy actions.
The Department for Promotion of Industry and Internal Trade (DPIIT), in its December 2025 Working Paper on Generative AI and Copyright, explored a centralized statutory licensing framework under the proposed “One Nation One License One Payment” model.
Licensing can be a way of authorizing and compensating, but it does not entirely resolve the question of provenance.
In a licensing regime, it’s necessary to know whose work is being used and which creators are entitled to compensation.
Provenance and licensing therefore need to go hand in hand.
Provenance tells you where the item came from. Licensing is about the legal and financial relationship.
Conclusion
Generative AI has not rendered copyright irrelevant. This has presented a problem that traditional copyright enforcement did not face to the same extent. The relevant activity may occur within a technological process that the copyright holder cannot observe.
A case in point is ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, where such questions arose under Sections 14, 51, and 52 of the Copyright Act, 1957. The Delhi High Court’s interim order does not conclude that all AI training is inherently lawful or exempt from copyright scrutiny; rather, it represents a preliminary determination grounded in the specific facts, pleadings, and procedural standards before the court.
Beyond the substantive legality of AI training, the next wave of AI copyright jurisprudence must address the procedural ability of copyright holders to determine the fate of their creations.
A framework based on provenance, proper transparency and regulated disclosure may protect progress and effective copyright enforcement at the same time.
The main question around AI copyright may therefore shift from: Who is the copyright owner?
To: Once a work has been processed by the machine, can the legal system trace its trajectory?
If the answer is no, the problem is not simply weak copyright protection. The problem is that copyright has entered a black box.
Also Read: Database Protection in India: Copyright, Originality Thresholds, and the Sui Generis Debate
References
- The Copyright Act, 1957 (Act No. 14 of 1957), India Code, Official Statute.
- Delhi High Court, ANI Media Pvt. Ltd. v. OpenAI OpCo LLC, CS(COMM) 1028/2024 (Interim Order dated July 24, 2026), Judgment on Indian Kanoon.
- Department for Promotion of Industry and Internal Trade (DPIIT), Ministry of Commerce and Industry, Government of India, Working Paper on Generative AI and Copyright, Part I: One Nation One License One Payment—Balancing AI Innovation and Copyright (Dec. 2025).
- Supreme Court of India, R.G. Anand v. Deluxe Films, (1978) 4 SCC 118, Judgment on Indian Kanoon.
- Supreme Court of India, Eastern Book Company v. D.B. Modak, (2008) 1 SCC 1, Judgment on Indian Kanoon.
- World Intellectual Property Organization (WIPO), Artificial Intelligence and Intellectual Property Policy, WIPO Knowledge Repository.
- European Parliament and Council of the European Union, Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 53 (Transparency Obligations for General-Purpose AI Models).
