ABSTRACT
In light of the difficulties presented by the development of generative AI, this article examines the ways in which Indian copyright law handles these issues. The concepts of authorship, originality, and responsibility as they pertain to copyright infringement caused by artificial intelligence are the primary areas of discussion. It begins by explaining how current machine-learning frameworks operate, then moves on to discuss how AIs are trained and how humans might influence their creative output. In addition, this article examines the Copyright Act of 1957 from a doctrinal standpoint, paying special attention to the statutory definition of the creator of computer-generated works included in section 2(d) and the minimal creative criterion as articulated in the case of Eastern Book Company v. D.B. Modak (vi). After that, these criteria are used for AI-assisted and AI-driven results, taking into account the fair dealing requirements of section 52 and the intermediary safe harbor provisions of section 79 of the IT Act and the Intermediary Rules of 2021. The report wraps up by suggesting normative remedies and future legal avenues to bring Indian copyright law up to speed with computer innovation. These include a significant human input test and the creation of specifically designed sui generis regimes.
INTRODUCTION
The pervasiveness of AI in modern life has made it almost difficult to distinguish between naturally occurring and artificially produced goods and services. AI has revolutionized our lives by taking over tasks that were formerly thought to be uniquely human in fields as diverse as healthcare, transportation, aviation, space, education, and the entertainment industry (including but not limited to music, art, gaming, and films). In an effort to maximize efficiency and eliminate human error, a global trend toward automating routine tasks and reducing human involvement has emerged[1]. Artificial intelligence is one of the most important things humanity is working on, Sundar Pichai said during a San Francisco town hall gathering. I don't know, it's deeper than fire or electricity.[2]
“Indian copyright law, which has been around for a long time and is responsible for protecting people's rights to intangible property, is facing serious problems as a result of this technological change. Legal concerns of authorship, ownership, originality, and infringement are complicated by the increasing global capability of AI systems to generate a range of compositions, sometimes with little or no human involvement.
This paper argues that Indian copyright law is infused with the doctrinal tools to interpret and address the AI assisted human attribution and creativity without any legislative rupture, while simultaneously rejecting the idea of autonomous AI system as authors. It examines authorship and originality under Copyright Act, 1957 and critically analyses the scope and limits of the fair dealing doctrine in the context of AI training and generative outputs, demonstrating how existing copyright principles may be judicially interpreted to retain human creative agency while responding to challenges posed by the emerging technologies.
Therefore, the rise of AI not only challenges the framework but also the values humans bring in the field of creativity. For instance, when I entered the prompt Create an apology song for my girlfriend into Chat GPT, this was the result I got-
Here’s a short and sweet apology song for your girlfriend:
Verse
I know I let you down, and I’m so sorry,
I never meant to make you cry, But I’ll do anything to make it right, I need you here, don’t say goodbye.
Chorus
Forgive me, love, I’m feeling lost,
I’ll fix the mess, no matter the cost.
I’m sorry, babe, please understand,
I’ll be your man, if you’ll take my hand.
Short, simple, and straight from the heart. I hope this helps! Let me know if you want any changes.
It’s remarkable how effortlessly the results have been provided by an AI with just a single prompt, with minimal human intervention. Furthermore, what is particularly intriguing is that, had I specified the style of a renowned artist, the AI would have customized the result accordingly. This ease of access, however brings certain concerns that we are going to address in this paper.
It is remarkable how effortless the process has become requiring just a single input to generate the content one desires. Furthermore, what is particularly intriguing is that, had I specified the style of a renowned artist, the AI would have customized the output to align with that artist’s distinctive style.
TRAINING PROCESS OF AN AI
The time and money needed to train an AI is substantial. Reasoning, problem-solving, recognizing, and a plethora of other tasks requiring human intellect are the driving forces behind its development. Advanced artificial intelligence systems can learn from massive datasets using a variety of machine learning methods. Machine learning has emerged as a crucial method for implementing AI, according to Thomas W. Malone, a professor at MIT Sloan and the head of the MIT Center for Collective Intelligence. Malone claims that this trend has occurred in the last five to ten years. Most of the current advances in AI have involved machine learning, the author writes, explaining why the words are often used interchangeably.[3]
Ai and machine learning is often confused by the people but they are nothing but subsets of each other. Machine learning is a subset of artificial intelligence that automatically enables a machine or system to learn and improve from experience. Instead of explicit programming, machine learning uses algorithms to analyze large amounts of data, learn from the insights, and then make informed decisions.
Machine learning algorithms improve performance over time as they are trained exposed to more data. Machine learning models are the output, or what the program learns from running an algorithm on training data. The more data used, the better the model will get.[4]
Along with supervised and unsupervised ML models, semi-supervised and reinforcement learning are two of the most popular types of ML models.
Observed: In supervised learning, the system learns to associate inputs with their proper outputs by studying previously tagged models. When a precise set of inputs is known to provide a desired outcome, supervised learning is used[5]. The most typical tasks in this kind of learning are classification, which entails classifying data into groups, and regression, which involves predicting numerical values. For instance, identifying the feeling or category of a piece of text, such as a tweet or product review.
Unsupervised: Unsupervised learning is able to find structures or patterns in unlabeled data without any human intervention. Anomaly detection, data grouping, and data complexity reduction are typical uses for it. Clustering, anomaly detection, dimensionality reduction, feature learning, density estimation, and establishing association rules are among the most prevalent unsupervised learning tasks[6].
By combining supervised and unsupervised learning techniques with a combination of labelled and unlabeled data, semi-supervised learning achieves better results. This makes it a happy medium between totally directed and totally unguided learning. When dealing with large amounts of unlabeled data in real-world scenarios, semi-supervised learning may be a lifesaver[7]. By combining labelled and unlabeled data, this learning seeks to improve prediction accuracy. Complex topics like machine translation, text categorization, data annotation, and fraud detection find substantial use with it.
Machines and software agents may use reinforcement learning to determine the optimal course of action in every given situation, greatly increasing their efficiency[8]. The ultimate goal of this kind of reinforcement-based learning is to utilize the knowledge obtained from environmental activists to either increase the benefit or decrease the danger9. Reinforcement learning effectively trains AI to improve complex systems like robotics and autonomous vehicles but is not ideal for simple tasks.
These training methods have allowed AI to become incredibly powerful and capable of understanding language, generating images, composing music, and even mimicking human creativity. However, much of this progress has been made possible by feeding AI systems with vast amounts of data, often sourced from the internet, where not all content is freely available or meant to be reused. This raises a number of significant legal issues, such as whether using copyrighted material to train AI is treating authors fairly and how the usage of pre-existing information is upending traditional notions of ownership and altering copyright law itself.
IMPACT OF AI AND MACHINE LEARNING ON COPYRIGHT LAW
There are currently no specific provisions in the Indian Copyright Act, 1957 that deal with the ownership or infringement of works produced or impacted by artificial intelligence (AI). As AI-generated content from music and literature to multimedia art becomes more common, this statutory lacuna has grown in importance. Indian copyright law excludes non-human creators like artificial intelligence (AI) systems because it still relies on the conventional notion of a person as the author, in contrast to human authorship. An essential condition for copyright protection is originality, which states that a work cannot be the outcome of mechanical replication or duplication but rather the product of an author's skill, judgment, and effort. With no human imagination or initiative, the question of whether AI-generated output may be really unique arises in light of this legal criterion.[9] As artificial intelligence evolves, so are the laws that govern it. Not only are artists, writers, and musicians experiencing serious issues as a result of this widening gap, but everyone whose livelihood or line of work depends on precise guidelines about accountability and ownership is as well. This in turn brings up important issues like: Who owns anything created by an AI? Is it the firm that created the AI, the individual who created it, or nobody at all? And who should be held accountable if such effort harms someone or infringes upon their rights? We'll examine these issues in more detail in the parts that follow, as well as how AI is changing our understanding of copyright and accountability.
The case of Thaler v. Perlmutter (U.S. Copyright Office, 2023)[10], is a landmark case that highlights the emerging tension between AI and conventional concepts of copyright law. In his copyright registration application, plaintiff Stephen Thaler made it clear that the AI system he created, the Creativity Machine, was the only author of the piece of visual artwork A Recent Entrance to Paradise. The plaintiff also stated that the machine did not require any human intervention in its creation. Despite the absence of traditional human authorship, Thaler argued that the work should nevertheless be protected by copyright. If an AI meets the creative criteria for copyright, he said, its developer should be considered the owner and the AI should be acknowledged as the inventor. The court affirmed the judgment of the U.S. Copyright Office to deny registration, reasoning that an AI-generated work cannot meet the copyright law's essential requirement of human authorship. Since there could be no copyright in such a work, Thaler's request for a summary decision was denied.
AUTHORSHIP AND OWNERSHIP IN AI‑GENERATED WORK
Artistic outputs such as writing, photographs, music, and videos may now be efficiently generated by AI and other cutting-edge technology in a way that is very close to human expression. But just because these algorithms are technically efficient doesn't mean they're inevitably authors or that their work deserves copyright protection. There has always been a difference between ownership and authorship in Indian copyright jurisprudence. In contrast to ownership, which pertains to the entity that receives the economic rights in a work which can be someone other than the author due to employment, commission, or assignment authorship pertains to the individual who really thinks of and expresses the work.
Section 2(d)[11] of the Copyright Act, 1957 defines author in relation to different categories of works, and in the case of any literary, dramatic, musical or artistic work which is computer‑generated, identifies as author the person who causes the work to be created. This formulation presupposes a human person who can be said to have exercised causal control over the coming into existence of the work, even where a computer is used as an instrument. At the same time, Section 17[12] clarifies that, subject to contract, the first owner of copyright may be someone other than the author for example, an employer in the course of employment or a person commissioning photographs or paintings for valuable consideration.
The Raghav Case brought the topic of authorship of generative AI outputs in India to main light as in this case the Indian Copyright Office first acknowledged the RAGHAV Artificial Intelligence Painting App as a co-author of the artwork Suryat alongside the human artist, Ankit Sahni. However, this acknowledgment was later withdrawn on the grounds that an AI system could not be regarded as a person in accordance with Section 2(d)[13] of Indian copyright law. A definitional gap in the current statutory language that accommodates non-human creators is revealed by the episode. In contrast to that, several foreign jurisdictions have taken a broader approach for example, Canada listed both Sahni and RAGHAV as co‑authors of Suryat, showing an inclination towards the idea of AI systems as collaborators for authorship purposes. Chinese jurisprudence has allowed organizations, in certain circumstances, to be recognized as authors of AI‑generated works if they demonstrate willingness to create and claim rights, as seen in the case of Dreamwriter[14] and in Beijing Feilin Law Firm v. Baidu Technology Co. Ltd.[15] However in Europe, particularly Germany and France, the authorship and originality are rooted in human personality, even as the EU continues to explore the possibility of granting some form of legal recognition for AI generated outputs.
The US has adopted a relatively stringent stance. In Thaler v. Perlmutter[16], Stephen Thaler, emphasizing that his artificial intelligence (AI) system Creativity Machine was exclusively responsible for creating the work of art A Recent Entrance to Paradise, sought registration for the item. The US Copyright Office's decision to deny registration was affirmed by the District Court for the District of Columbia. The court ruled that copyright does not protect AI output that is completely autonomous and that copyright has never stretched so far to include works that do not have human authorship. This follows the Office's long-standing policy of not protecting works created only by machines without any human creativity.
Given this contrasting context, the present state of Indian law maintains a focus on human beings. Copyright protection is not afforded to AI-generated outputs under Indian law unless a human being can be named as the author or as the person who causes the work to be created according to Section 2(d) (vi).[17] The scope of Sections 13 and 17[18], Scholars and expert groups have reexamined the concepts of first ownership and copyrightable subject matter in relation to AI-assisted creativity because of this interpretation constraint.
A big open question is the minimum level of human involvement required for AI-assisted jobs.
Indian courts have not yet decided whether a user's prompt to an AI tool like ChatGPT constitutes enough creative input to establish authorship, but the current skill and judgment standard under Indian originality doctrine suggests that some traceable human intellectual effort is necessary. Though not officially embraced in India, the academic discussion has been influenced by the underlying premise that even little but meaningful creative judgments in selection, organizing, or curation may be enough.
Indian courts have articulated a refined approach to decipher the originality that is instructive for AI‑generated outputs. In R.G. Anand v. Deluxe Films[19], the Supreme Court clarified that copyright protects the style and manner of expression of ideas and not the idea or plot itself, and that infringement occurs only where the defendant copies the fundamental or substantial aspects of that expression. In Eastern Book Company v. D.B. Modak[20], Rejecting the more expansive sweat of the brow standard, the Court ruled that an item must demonstrate a minimal level of inventiveness, competence, and discretion. When applied to AI situations, these principles recommend a balanced approach: outputs created solely by AI without human interaction should not be protected, whereas creations supported by AI that are skillfully and arbitrarily made may qualify.
LIABILITY FOR AI‑DRIVEN COPYRIGHT INFRINGEMENT
Authorship and ownership concerns, which address the right to rights in works helped by AI, are complex, but liability for infringement coming from AI training and deployment is equally so. Doctrinally significant steps in training AI models include ingesting, duplicating, and storing copyrighted works and producing and disseminating outputs that could duplicate or alter protected information.
Under Indian law, Section 51[21] defines infringement as the unauthorized doing of any act reserved to the copyright owner under Section 14[22], or by allowing a public space to be used for commercial purposes in a way that infringes upon others' rights. In theory, unless there is an explicit legislative exemption, like one of the fair-dealing clauses in Section 52, the massive copying and scraping of text, pictures, audio, and audiovisual information for training purposes might be considered reproduction under Section 14[23].
The kind of questions that Indian courts are likely to encounter may be seen in foreign rulings. According to reports, a German court ruled that training a model like ChatGPT on protected lyrics was considered reproduction under German law. The court went on to say that fixing works in the numerical parameters of the model was also a kind of reproduction, as it allowed technical means to reconstruct or perceive the works in the future. The proceedings were brought by the collecting society GEMA against OpenAI. In contrast, generative-image systems like Stable Diffusion have been the subject of litigation in the UK, which has cast doubt on the idea that the internal model weights are literal copies of the training pictures. Instead of being precise reproductions, the contention is that they contain abstract patterns.
These divergent opinions show how difficult it is to make a blanket determination on the legality of AI training. The exact language of the legislation, the scope of exceptions (such fair-dealing provisions), concerns over the amount of memorization and resemblance between training materials and outputs, and the possibility of market replacement are all important factors. Unlike US fair use, which allows for a wider range of purposes, including reporting on current events, private use (including study), and criticism or review, Indian law currently does not provide an exemption for bespoke text-and-data-mining. Additionally, its fair-dealing provision is more restricted.
Once it is recognized that some training or output methods may be infringing, the allocation of responsibility among the many parties involved becomes problematic. There are at least three separate groups involved: model developers, who design the architecture and decide on data collection tactics; deployers or platform operators, who provide user interfaces and may improve models; and end users, who input prompts and intentionally create infringing outputs, such as lyrics or passages word for word. Section 51(a) addresses main infringement, whereas Section 51(b) covers a narrower notion of vicarious liability for individuals who, knowing or reasonably believing that their premises would be used for infringing communication to the public, permit such use.
To platforms and model creators that purposefully build or operate systems in a manner that reliably produces infringing outputs when requested, the Indian Courts may apply these concepts in the context of AI. Assigning duties is further aided by Indian theology's emphasis on communication to the public. Section 2(ff) provides a detailed definition of communication to the public, and Section 14 states that only right-holders have the exclusive right to communicate works to the public, including using digital means. The Supreme Court acknowledged in International Confederation of Societies of Authors and Composers (CISAC) v. Aditya Pandey[24] that various entities may exercise overlapping communication rights and that multiple layers of copyright (such as lyricist, composer, and sound-recording producer) may coexist. It is possible to apply these ideas to generative AI services that, upon request, provide the public with potentially infringing results.
The Information Technology Act of 2000 applies to this setup. In Section 81, it is stated that the Copyright Act's rights are unaffected by this Act. However, in Section 79, intermediaries are granted conditional safe-harbor protection for third-party information that they host or transmit, provided that they fulfill due diligence obligations and remove the content promptly upon discovering its illegality. This means that AI systems that only allow users to provide inputs and results may, in certain cases, claim that they are in the middle ground. However, this will not absolve them of responsibility in cases where they violate training protocols or when they actively favor or curate specific results. Thus, the interaction between Section 79, Section 51, and Section 14 of the Copyright Act would be the central focus of any future case in India involving AI-driven infringement.
FAIR USE / FAIR DEALING AND AI
Financial sustainability, market control, attribution, protection against illegal use, and cultural and technical innovation are all facilitated by the copyright law, which allows authors exclusive rights to reproduce, distribute, perform, and modify their works. Additionally, the notion of fair use is a provision in the same statute that allows artists to make use of such exclusive rights. This theory allows for the reuse of copyrighted materials so long as they are substantially improved upon from the original and do not constitute a derivative or adaptive work. The theory was included into the Copyright Act of 1976 (US LAW) from the United States' common law. In deciding whether Fair Use may be used as a defense in a given case, the courts primarily consider four factors: the use's purpose and character, the copyrighted work's nature, the quantity and substantiality of the use, and the impact on the market value of the work. It is up to the discretion of the court to determine whether fair use may be used as a defense in any given case, taking into account such criteria. To prevent the repression of genuine public interests like study, education, criticism, and transformative innovation, the doctrine acts as a crucial check and balance within copyright law. Exercising fair use is a right, not an obligation, rightly says James S. Heller, who stresses that fair use is more than just a defense; it's a constitutionally protected entitlement that promotes cultural engagement and democratic access to information. This principle was notably reinforced in Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994)[25] where the U.S. Supreme Court held that transformative use such as parody can qualify as fair use even when done for commercial purposes. The importance of fair use in maintaining the constitutional balance between author incentives and public discourse promotion was acknowledged by the Court. So, even when people build upon the ideas of others, the theory nevertheless recognizes and protects their original contributions.
Section 52 of the Indian Copyright Act, 1957, which allows certain uses of copyrighted works without the author's consent, encompasses the idea of fair use in India. This section outlines different exceptions to infringement. This covers uses that do not infringe on others' rights, such as lawful software backup and functional copies, temporary reproductions for electronic transmission and depiction of public art, limited educational and non-commercial reproductions, and fair dealing for private study, research, criticism, review, news reporting, and judicial proceedings.
Although at first glance the ideas of fair dealing and fair usage are quite similar, there are important distinctions between the two that stem from their respective natures, areas of application, and scopes. Fair use encompasses a much broader range of activities than fair dealing. If an exemption is not explicitly stated in the legislation but it can be reasoned that such work has been used for fair use, then the doctrine of fair use will apply. In contrast, the latter may only be applied in relation to the uses or exceptions expressly established in the law. Several instances involving AI-driven copyright infringement have employed fair use as a defense; nevertheless, the results have varied according on the court's discretion in determining whether it constitutes fair use or not. For example, in the case of Bartz vs Anthropic PBC[26], Anthropic argued that digitizing and destroying purchased books to train its LLM is a transformative fair use. Judge Alsup while delivering the judgement, upheld fair use, finding the LLM training among the most transformative we will see in our lifetimes. In some cases, this defence has been outright rejected as well. For example, in the case of Thomson Reuters v. Ross Intelligence[27], Ross argued that it was using legal research materials in its AI system in a fair use manner, but the court disagreed, ruling that preserving whole works for an extended length of time does not qualify as fair use. The case is still ongoing, but OpenAI is claiming that training its LLM on news content falls under fair dealing because it is transformative. This is according to OpenAI's argument in ANI v. OpenAI Opco LLC, which is heard in the Delhi High Court.
It is clear from the aforementioned examples that the particular circumstances and legal standards used in each country determine the validity of a fair use defense. Courts reach diverse conclusions based on how several criteria, including as purpose, type, quantity, and impact on the market, match; no specific rule ensures victory. Jurisdictional theories differ as well; for example, although India's Section 52 fair dealing exemption is more limited and purpose-specific, the United States has a more liberal four-factor test. What passes muster in California may not be up to snuff in Delhi. Taken together, these instances show that determining fair use in the age of AI is still a question that requires a lot of evidence. Not only will the courts' use of recognized fairness principles determine the outcome, but so will shifting perspectives on training as transformation, worries about market replacement, and the sufficiency of compliance procedures implemented by AI developers.
The lack of uniformity in these results highlights the need for more specific copyright regulations for artificial intelligence, such as strengthened safe harbours, obligatory licensing systems, or sui generis rights, to ensure that innovators and rights holders have what they need.
NORMATIVE SOLUTIONS AND FUTURE LEGAL PATHWAYS
|
Model |
Core Idea |
Jurisdictional Examples |
Ownership/Protection outcome |
Doctrinal Strength |
Key Concerns |
|
Strict Human Authorship Model |
Copyright subsists only in works created by natural persons; authorship presupposes a human mind. AI and other non-human entities cannot be authors or rightholders. Purely AI-generated outputs without meaningful human creative input are ineligible for protection. |
• United States Thaler v. Perlmutter (D.D.C. 2023) confirming bedrock human-authorship requirement. • US Copyright Office practice excluding works produced by purely mechanical/automatic processes. • EU Court of Justice’s author’s own intellectual creation standard presupposing a human author.
|
• Purely autonomous AI outputs receive no copyright protection and fall into the public domain. • AI-assisted works may be protected, but only to the extent a human’s intellectual contribution (selection,arrangement, editing) meets the relevant originality test (e.g. modicum of creativity or skill and judgment). |
• High preserves traditional theoretical foundations of copyright as a reward for human intellectual creation and personality. • Aligned with existing statutory texts and judicial precedents in major jurisdictions (US, EU) without requiring legislative overhaul. |
• Large volumes of autonomous AI content remain unprotected, creating a default public domain that may influence investment incentives for proprietary generative models, though scholarship is divided on whether exclusivity is necessary for recouping costs. • Uncertainty in borderline mixed cases regarding the threshold of human input required (for example, prompt-only versus extensive editing and curation). • Practical evidentiary burdens in proving the extent of human creative contribution in AI-assisted workflows. |
|
Human Causation Model |
The author of a computer-generated work is the human (or legal person) who undertakes the arrangements necessary for its creation; authorship is derived from causal control over the generation process rather than direct creative selection of each expressive element. |
• United Kingdom CDPA 1988, s. 9(3): author of a computer-generated work is the person by whom the arrangements necessary for the creation of the work are undertaken. • India Copyright Act, 1957, s. 2(d)(vi): in the case of computer-generated literary,dramatic, musical or artistic works, the author is the person who causes the work to be created. • Other common-law jurisdictions such as Ireland and New Zealand have similar formulations (noted comparatively in Indian scholarship) |
• Depending on factual control, authorship (and thus initial ownership) may be attributed to developers, deployers, commissioners, or users who cause the work to be created through arrangements and operation of the system. • Purely autonomous outputs may be captured if courts adopt a generous view of arrangements necessary or causation, treating upstream design and deployment decisions as sufficient. |
• Moderate directly grounded in statutory text (UK CDPA, Indian s. 2(d)(vi)). • Offers a pragmatic allocation of rights where no natural person directly determines the final AI expression but human agents still structure and trigger the creative process. |
• Conceptual difficulty in identifying who, among programmers, data curators, platform operators and end users, has caused the work, especially in highly distributed AI development pipelines. • Risk of artificial attribution deeming someone author even when they did not exercise recognisable creative judgment in relation to the specific output, potentially diluting the intellectual-creation basis of copyright. • Doctrinal tension with personality-based understandings of authorship prevalent in civil-law traditions. |
|
Joint/ Collaborative Attribution (Proposed) |
Authorship and ownership are shared between multiple human contributors involved in AI development and use e.g. programmers, data providers, platform operators and creative end-users on an analogy with collaborative or joint works. |
• Canadian and EU academic proposals suggesting joint authorship between system developers and human users where both contribute to creative choices. • Comparative commentary on the RAGHAV/Suryat Canadian registration treating AI and human as co-authors, used in scholarship to argue for collaborative models |
• Rights in the AI-assisted work are distributed among different human contributors, potentially reflecting their respective roles (for example, programmer and human user as co-authors, with subsequent allocation via contract). • Economic exploitation and enforcement would require collective decision-making or agreed allocation of exploitation powers, often via contractual arrangements. |
• Low to Moderate aligns with joint-authorship concepts in copyright statutes, but existing provisions were not designed for the diffuse, layered contributions typical of AI systems. • Doctrinally plausible where multiple humans make identifiable creative contributions; harder where contributions are highly technical or remote from the final expression. |
• AI itself lacks legal personality; any co-authorship must still be between human or juristic persons, which undercuts proposals to treat AI as a co-author in its own right. • High administrative and transactional complexity in managing joint rights (clearance, licensing, enforcement) across numerous stakeholders. Risk of over-fragmentation of rights, potentially stifling downstream uses rather than promoting them. |
|
Sui Generis AI Right (Proposed) |
Instead of stretching copyright doctrine, a separate, limited statutory right is created for AI-generated works, granting certain economic rights (e.g. against commercial copying) for a shorter term or narrower scope than full copyright. |
• Discussed in UK policy reviews on AI and IP, often in the context of the computer-generated works regime under CDPA 1988 and possible reform options. • Raised in WIPO conversations on AI and IP as one way to reward investment in generative |
• Holders (likely developers or commissioning entities) receive a time-limited, narrower economic entitlement over certain AI-generated materials (for example, against commercial copying or unfair extraction), while moral rights and long copyright terms are not extended to AI outputs. |
• Conceptually relatively strong – avoids distorting copyright’s human-centric foundations by creating a distinct category tailored to AI outputs. • Allows calibration of term, scope and limitations specifically for AI-related investments. |
• Risk of over-protection and new enclosure of information resources if the sui generis right is drafted too broadly, especially in favour of large technology firms controlling major models and datasets. • Potential overlap and conflict with existing copyright and database-protection regimes, creating legal uncertainty. • Implementation would require fresh legislation and |
|
technologies without conferring full copyright status. |
• Human-authored works remain governed by ordinary copyright rules; AI-only works fall under the sui generis regime, preserving conceptual clarity. |
international coordination; at present, it remains largely at the level of policy discussion. |
|||
|
Public Domain Approach |
Purely autonomous AI outputs are treated as unprotected and belong to the public domain unless and until a human makes a sufficiently creative contribution to the expression; human-centred copyright remains intact, and AI materials become freely usable inputs for subsequent creation. |
• Supported in US and EU academic discourse as the logical consequence of strict human-authorship and intellectual creation requirements. • Reflected in US practice after Thaler v. Perlmutter, where purely AI-generated works are denied registration and thus left unprotected. |
• Autonomous AI outputs are free for all to use, adapt and commercialise, subject only to non-copyright constraints (defamation, personality rights, regulatory limits, etc.). • Protection attaches only when human authors make an original intellectual contribution, e.g. by editing, arranging or building upon AI material. |
• High doctrinal purity – firmly preserves copyright as a human-centred regime grounded in intellectual creation. • Simple and predictable rule: no human creativity, no copyright; AI outputs default to public domain. |
• Concern that absence of exclusive rights in AI outputs might reduce private incentives to invest in costly generative models, although developers can still rely on trade secrets, contracts and technical controls. • Potential for rapid commodification and mass reuse of AI-generated culture, with unclear long-term effects on creative labour markets. |
|
Contractual Allocation Model |
Instead of resolving authorship doctrinally, ownership and control over AI-generated outputs are allocated by contract (terms of service, licences, employment agreements). Platforms may assign any rights they have in outputs to users or retain them, irrespective of statutory authorship debates. |
• Industry practice: AI platform user agreements, such as OpenAI’s Terms of Use, under which users retain ownership of inputs and are assigned any operator rights in outputs to the extent permitted by applicable law. • Cloud-service and API agreements allocating data and model-output rights between enterprise customers and providers (noted in Indian commentary as emerging practice). |
• As between contracting parties, economic control over AI outputs is determined by licence and assignment clauses; for example, users may receive broad exploitation rights even where statutory copyright in the output is uncertain. • Third parties remain subject to whatever statutory rights actually subsist; contracts cannot create copyright where statutes deny protection, but they can structure relative entitlements among signatories. |
• Legally enforceable in private law; fits comfortably within existing assignment and licensing provisions (e.g. Indian Copyright Act ss. 18– 20) without need for new doctrine. • Operationally flexible – parties can tailor allocations to commercial preferences and risk-sharing arrangements. |
• Does not resolve the underlying statutory question of who is an author or whether the AI output is copyrightable at all; it merely allocates whatever rights may exist. • Heavily dependent on private ordering and bargaining power, potentially disadvantaging end-users or small creators in standard-form contracts. • Limited transparency and interoperability of terms across platforms, which may create fragmentation and uncertainty for downstream users and licensees. |
CONCLUSION
The study proved that Creative Personhood Theory still governs Indian Copyright Law, despite the fact that it is unaffected by technology.” Section 2(d)(vi) of the Copyright Rules of the Copyright Act deals with this same issue, stating that the person who causes the work to come into being is deemed to be the creator of that particular work. The conclusion drawn from this is that the ultimate product of creation was the result of human ingenuity and imagination. Works with a sufficient and substantial amount of human intellectual and creative contribution may be affordably protected by copyright, taking into account the skill and judgement criterion outlined in Eastern Book Company v. D.B. Modak and the minimal quantity needed for creativity. Nonetheless, according to the same concept, generative output and other works that are developed with little or no human intellectual innovation will not be granted intellectual property rights. Concerns about copyright infringement, fair dealing, and the implementation of the Copyright Act's text-and-data-mining exceptions have arisen in response to the growing popularity of generative models for this purpose. The present safe-harbour regulations governing the activities of intermediary service providers were not drafted with the intention of facilitating the registration of AI developers. While certain doctrinal shifts have helped clarify some of the aforementioned questions for example, whether there is a significant amount of human input into the creation of works and whether section 52 applies in a training context the courts will not be able to provide a sufficient legal framework through judicial interpretation alone due to the far-reaching economic consequences. To clarify the laws of training generative models in the limitations/exceptions regime created by the Copyright Act, to construct narrowly focused legislative solutions that embrace the constitutional commitments of access, education, and cultural dissemination reflected in the courts' interpretation of copyright law (with respect to section 52), and to establish clarity as to authorship/ownership of works created using AI technology, limited legislative change may be necessary.
Reference
[1] V.K. Ahuja, Contemporary Developments in Intellectual Property Rights: A Prologue in V.K. Ahuja and Archa
Vashishtha, Intellectual Property Rights: Contemporary Development 3-18s (Thomson Reuters, 2020)
[2] Leswing, Kif. Google CEO: AI Is More Important than Fire or Electricity. CNBC, 1 Feb. 2018, www.cnbc.com/2018/02/01/google-ceo-sundar-pichai-ai-is-more-important-than-fire-electricity.html.
[3] MIT Sloan School of Management. (2021, April 21). Machine learning, explained. MIT Sloan Ideas Made to Matter. https://mitsloan.mit.edu/ideas-made-to-matter/machine-learning-explained.
[4] Google Cloud, What is Machine Learning (ML)?, Google Cloud, accessed July 11, 2025, https://cloud.google.com/learn/what-is-machine-learning.
[5] Sarker IH, Kayes ASM, Badsha S, Alqahtani H, Watters P, Ng A. Cybersecurity data science: an overview from machine learning perspective. J Big Data. 2020;7(1):1–29. doi: 10.1186/s40537-020-00318-5.
[6] Sarker, I. H. (2021). Machine learning: Algorithms, real-world applications and research directions. SN Computer
Science. https://doi.org/10.1007/s42979-021-00592-x
[7] Mohammed M., Khan M. B., and Bashier Mohammed B. E., Machine Learning: Algorithms and Applications (Boca Raton, FL: CRC Press, 2016), https://doi.org/10.1201/9781315371658.
[8] Kaelbling, L. P., Littman, M. L., & Moore, A. W. (1996). Reinforcement learning: A survey. Journal of Artificial Intelligence Research, 4, 237–285. https://doi.org/10.1613/jair.301 9 Supra, 7
[9] AI and Copyright Law in India, The Legal School (May 5, 2024), available at https://thelegalschool.in/blog/ai-andcopyright-law-in-india (last visited July 15, 2025).
[10] Thaler v. Perlmutter, 2023 WL 5333236 (D.D.C. Aug. 18, 2023).
[11] Copyright Act, 1957, No. 14 of 1957, § 2(d)(vi) (India).
[12] Copyright Act, 1957, No. 14 of 1957, § 17 (India).
[13] Supra note11
[14] Shenzhen Tencent Computer Sys. Co. v. Shanghai Yingxun Tech. Co., (2019) Yue 0305 Min Chu No. 14010 (Nanshan Dist. People’s Ct., Shenzhen, China Dec. 24, 2019).
[15] Beijing Feilin Law Firm v. Beijing Baidu Netcom Science & Technology Co., Ltd., (2018) Jing 0491 Min Chu No.239
[16] Thaler v. Perlmutter, 2023 WL 5333236 (D.D.C. Aug. 18, 2023).
[17] Supra note 12
[18] Copyright Act, 1957, § 13, 17 (India).
[19] R.G. Anand v. Deluxe Films, (1978) 4 S.C.C. 118 (India).
[20] Eastern Book Co. v. D.B. Modak, (2008) 1 S.C.C. 1 (India). 23 Supra 14
[21] Copyright Act, 1957, § 51 (India).
[22] Copyright Act, 1957, § 14 (India).
[23] Copyright Act, 1957, § 52 (India).
[24] Int’l Confederation of Soc’ys of Authors & Composers (CISAC) v. Aditya Pandey, (2017) 8 S.C.C. 670 (India).
[25] Campbell v. Acuff-Rose Music, Inc. | 510 U.S. 569 (1994) | Justia U.S. Supreme Court Center
[26] ORDER ON 122 FAIR USE for Bartz et al v. Anthropic PBC :: Justia Dockets & Filings
[27] Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., No. 1:20-cv-613-SB, 2020 WL 5518423