AI-Generated Content And Copyright Ownership - Who Holds The Right
Author: LawSphereX
Category: Intellectual Property
Published:
Updated:
AI-GENERATED CONTENT AND COPYRIGHT
OWNERSHIP: WHO HOLDS THE RIGHT?
Aman Verma, Amity Law School, Amity University, Lucknow, Uttar Pradesh Dr. Axita Shrivastava, Associate Professor, Amity Law School, Amity University, Lucknow, Uttar Pradesh
ABSTRACT
Generative artificial intelligence has changed the basic ideas that support copyright law. The Indian Copyright Act, 1957, like similar laws in other major regions, assumes a human author. However, AI systems now create literary, artistic, and musical works with little to no human creativity involved. This paper explores who should hold copyright over AIgenerated content. It looks at the claims from the AI system, the user or prompter, the developer, and the public domain. It also surveys the legal situations in India, the United Kingdom, the United States, the European Union, and China. The paper critically reviews important cases, including Thaler v Perlmutter, the Zarya of the Dawn ruling, and the ongoing ANI v OpenAI case in the Delhi High Court.
It discusses the issues with training data and the debated application of Section 52 of the Copyright Act, 1957, regarding AI use of copyrighted works. The paper concludes that current frameworks are not sufficient and suggests a tiered authorship model along with a short-term sui generis right as the fairest and most practical solution for India Keywords: Artificial Intelligence, Copyright, Authorship, Indian Copyright Act 1957, Generative AI, Training Data, Sui Generis Right, Fair Dealing, Section 52, WIPO I. Introduction Intellectual property law has always been pushed forward by technology. The printing press produced the Statute of Anne. Photography forced courts to reconsider what 'authorship' really meant. Software prompted major legislative overhauls across the world. Each wave of disruption unsettled the law, and each time the law eventually found a new footing.
Generative artificial intelligence — systems capable of producing novels, artworks, legal drafts, and music with minimal human involvement — may be the most disruptive wave yet, because it does not merely change what is created; it challenges who creates. The legal difficulty this creates is deceptively clean: if an AI system produces a poem or a painting without meaningful human creative input, who owns the copyright? The question is not academic. It determines the rights of AI developers who invest heavily in building these systems, of users who deploy them commercially, of human creators whose works train them, and of courts that must ultimately resolve the disputes between these parties. India's position is particularly vulnerable. The Copyright Act, 1957 defines 'author' under Section 2(d) with reference to natural persons in every limb of the definition.
1 Section 13 requires works to be 'original' — a word the Supreme Court has interpreted to demand a minimum degree of human creativity. 2 Section 9(3) of the United Kingdom's Copyright, Designs and Patents Act 1988, which vests authorship in the person who 'makes the necessary arrangements' for a computer-generated work,3 has no express equivalent in the Indian Act, though some scholars have argued that Section 17B could serve an analogous function by implication. 4 That argument has never been tested before any Indian court. The Delhi High Court is currently seized of the ANI v OpenAI proceedings, which raise questions about the reproduction of copyrighted news content in AI training,5 but output ownership — the question of who holds copyright in what the AI produces — remains entirely unaddressed by Indian jurisprudence. This paper proceeds in eight parts.
Part II examines the philosophical foundations of copyright and why they resist application to AI. Part III maps the competing ownership claims. Part IV compares the legal responses of five jurisdictions. Part V analyses landmark cases. Part VI addresses the training data problem and Section 52. Part VII examines the Indian legal landscape in depth. Part VIII proposes a reform framework and concludes. II. The Philosophy of Authorship and Its Limits Copyright law rests on three philosophical pillars, each of which has something to say about AI — and none of which produces a comfortable answer. The first is Locke's labour theory: a person acquires property rights by mixing their labour with the natural world, and the fruit of that labour belongs to them. 6 Applied to copyright, the creative effort of the human author justifies exclusive rights in the resulting work.
The difficulty with AI is immediate: the morally relevant labour belongs to the engineers who built the model and to the human creators whose works trained it — not to the AI itself, which expends no effort in any sense the law has traditionally recognised, and arguably not to the casual user either, whose contribution may amount to little more than a sentence-long prompt. The second is Hegel's personality theory, which grounds intellectual property in the expression of the author's individual will and personality. 7 On this view, a work of authorship is the externalisation of selfhood, and it deserves protection because it embodies something irreducibly personal. An AI has neither selfhood nor will. Its outputs are sophisticated statistical patterns derived from training data. They express nothing about the machine's inner life because the machine has none.
The personality theory, the most philosophically rich justification for copyright, is simply inapplicable to AI-generated works.
The third is the utilitarian incentive model, which underpins much of American copyright doctrine: society grants temporary monopolies to incentivise human creators to produce socially valuable works. AI systems do not respond to economic incentives — they produce regardless of whether copyright protection exists. The utilitarian case for protecting AI-generated works would have to rest on incentivising the humans and companies that develop and deploy AI tools, which is an instrumental argument that relocates the justification for protection in ways that carry significant distributional consequences. Ginsburg has argued that authorship in the Berne Convention framework has always been a fundamentally human concept.
8 The Convention's silence on the question9 has sometimes been read as openness to non-human authorship, but the weight of scholarly opinion and consistent state practice points the other way. The real question for national legislatures, therefore, is not whether AI can philosophically qualify as an author — it clearly cannot — but rather how to allocate, as a matter of policy, the rights in works that AI generates. III. Competing Claims: Who Could Own the Right?
Four parties could conceivably hold copyright in an AI-generated work. Each claim has a logic, and each has a flaw.
3.1 The AI System
The most instinctively appealing answer — that the AI should own what it creates — is also the legally weakest. An AI is not a legal person. It cannot hold property, enforce rights, or enter contracts. Courts in the United States10 have rejected AI authorship outright, and no jurisdiction currently grants an AI system copyright ownership. The philosophical objections are equally decisive: copyright exists to reward human creativity and incentivise human effort, neither of which the AI performs.
3.2 The User or Prompter
The person who prompts the AI — who describes what they want and chooses among the outputs — has the strongest claim to human involvement. But the strength of this claim varies directly with the nature and extent of the contribution. A user who types 'write a poem about rain' and accepts the first output contributes almost nothing that copyright law would recognise as creativity. A user who iterates extensively, selects, arranges, and substantially modifies the AI's outputs may cross the threshold into protectable human authorship — as the US Copyright Office recognised in the partial registration of Zarya of the Dawn.11 The difficulty lies in drawing and enforcing the line between these two poles.
3.3 The Developer or Company
AI developers invest substantial resources in building, training, and maintaining the models that generate creative outputs. Under a work-for-hire or entrepreneurial copyright analysis — of the kind applied to films, sound recordings, and databases in various jurisdictions — there is a reasonable argument for extending copyright to the entity that made the creative output possible. Abbott's concept of 'AI neutrality'12 supports this reasoning: law should not disadvantage AIgenerated outputs relative to human-created ones merely because of the nature of their production process. Guadamuz similarly observed that existing frameworks were not designed to accommodate machine-generated creativity and that some form of legal recognition may be necessary to prevent a regulatory vacuum.
13 The risk is one of concentration: extending copyright in this way would vest enormous amounts of intellectual property in the hands of a small number of technology corporations, potentially at the expense of human creators and the public domain.
3.4 The Public Domain
The cleanest answer under current law is that purely AI-generated works — those lacking substantial human creative contribution — attract no copyright and enter the public domain. This is broadly the consequence of the US position following Thaler v Perlmutter and the logical implication of civil law systems that require 'own intellectual creation' as a threshold for protection. The public domain default has genuine appeal: it preserves the philosophical integrity of copyright, enriches the creative commons, and prevents AI outputs from undercutting the market for protected human-created works. The disadvantage is that developers receive no protection for AI outputs, which may reduce the incentive to invest in generative AI tools. IV.
Comparative Jurisdictional Analysis The table below summarises how five major jurisdictions currently approach AI-generated works, followed by analytical observations. Jurisdiction Statutory Position Judicial / Administrative Key Gap Stance USA
17 USC — human
USCO denies copyright No framework for authorship required; for purely AI works compensating creators no express (Thaler v Perlmutter, whose works trained provision for AI- 2023); partial copyright the AI generated works where human selection and arrangement are present (Zarya, 2023) UK CDPA 1988, s 9(3) No appellate ruling on 'Necessary — author of a generative AI specifically; arrangements' test is computer-generated s 9(3) predates LLMs and unclear for modern work is the person its application to text-toneural networks making necessary image and LLM outputs is arrangements uncertain Jurisdiction Statutory Position Judicial / Administrative Key Gap Stance provides sui generis obligations but is silent on right output ownership China Copyright Law Li Yunkai v Liu (2023) — No judicial ruling on
2020 — protects
copyright granted where works generated works with human meaningfully without human 'originality'; no AImodified AI output; not a modification specific provision ruling on purely autonomous AI generation India Copyright Act No precedent on AI- Entire framework 1957, ss 2(d), 13 — generated works; ANI v untested for human author OpenAI (Delhi HC) generative AI; implied throughout; ongoing — addresses legislative reform s 17B possibly training data, not output overdue extends to 'arranger' ownership by analogy (untested) The United Kingdom's Section 9(3) CDPA 1988 remains the most practically workable of the existing statutory frameworks. By locating authorship in the person who makes the 'necessary arrangements' for a computer-generated work, it provides a mechanism for protecting AI outputs without granting rights to the machine or abandoning the work to the public domain.
The problem for generative AI is that 'necessary arrangements' was conceived in the 1980s for rule-based programs, not for neural networks operating over billions of parameters. Whether a text prompt to an LLM constitutes 'necessary arrangements' is far from obvious, and no UK appellate court has addressed the question. China's 2023 ruling is instructive but must be read carefully. 14 The Beijing court did not hold that AI can be an author. It held that a human who selects, arranges, and meaningfully modifies AI-generated content can acquire copyright in the result. This is substantively close to the US Copyright Office's position in Zarya of the Dawn.
The convergence of these otherwise divergent legal systems on this single point — human creative contribution is required, but need not be exhaustive — suggests an emerging international consensus on which Indian reform could credibly build. V. Key Cases and Their Implications
5.1 Thaler v Vidal (Fed Cir, 2022)
Dr Stephen Thaler sought US patent protection for inventions he attributed to his AI system, DABUS. The Federal Circuit held unanimously that the Patent Act requires inventors to be natural persons.15 Although the decision concerned patents, its reasoning — that intellectual property rights presuppose human agency — has been widely cited in the copyright context and lent weight to the Copyright Office's subsequent position in Thaler v Perlmutter.
5.2 Thaler v Perlmutter (DDC, 2023)
Thaler sought copyright registration for an AI-generated artwork attributed to DABUS. Judge Howell upheld the Copyright Office's refusal, holding that human authorship is a bedrock constitutional and statutory requirement of copyright law. The decision is the clearest judicial statement of the current US position. It does not resolve what degree of human contribution would suffice for protection — a question that remains actively litigated — but it firmly closes the door on purely AI-originated copyright claims.
5.3 Zarya of the Dawn (USCO, 2023)
The Copyright Office's registration decision in this graphic novel case is the most nuanced statement of current US administrative practice. The Office permitted copyright in the textual narrative and the human author's selection and arrangement of AI-generated images, but declined to extend protection to the AI-generated images themselves. The case confirms that partial human authorship can coexist with AI generation — a pragmatic position that sidesteps the all-or- nothing problem — but creates difficult line-drawing challenges that lower courts and the Office have not yet resolved.
5.4 Getty Images v Stability AI (2023)
Getty Images sued Stability AI in England, alleging that training the Stable Diffusion model on millions of Getty's copyrighted photographs without a licence infringed copyright and trade mark rights. The English High Court permitted the claim to proceed, finding it reasonably arguable.16 The case is significant because it shifts the legal battleground from output ownership to input legality — a distinction that has major implications for the training data debate addressed in Part VI.
5.5 The Authors Guild v OpenAI (SDNY, ongoing)
A coalition of prominent authors sued OpenAI in the Southern District of New York, alleging that GPT models were trained on their copyrighted books without authorisation, constituting direct and contributory infringement.17 The outcome of this case will significantly shape the legal and commercial landscape for AI training across common law jurisdictions, including India, whose courts will inevitably look to it for guidance in the absence of domestic precedent.
5.6 ANI v OpenAI (Delhi HC, ongoing)
India's first significant AI-copyright dispute involves a claim by the news agency ANI that OpenAI reproduced and stored its copyrighted news content in the course of AI training, in violation of Sections 14 and 51 of the Copyright Act, 1957. The matter remains pending, with hearings listed through 2025, and no final judgment has been delivered. The case raises important questions about the jurisdiction of Indian courts over foreign AI companies and the scope of the reproduction right in the context of AI training. Crucially, it does not directly address output ownership — but the reasoning Indian courts apply to the training data question will lay the doctrinal groundwork for the output ownership question that must eventually follow.
VI. The Training Data Problem and Section 52
A conceptually distinct but practically connected issue concerns the copyright status of data used to train AI systems. Generative AI models are typically trained on datasets containing billions of text samples and images sourced from the internet, a significant proportion of which are protected by copyright. Two questions arise: whether the ingestion of copyrighted works for AI training constitutes infringement, and what licensing frameworks could legitimately govern such use. Under Indian law, the primary provision is Section 52 of the Copyright Act, 1957, which provides a fair dealing exception for purposes including research and private study. 18 Whether large-scale commercial AI training qualifies as 'research' within this exception is analytically contentious.
The provision was designed with individual scholarly use in mind — a student or academic working privately with copyrighted material — not systematic machine ingestion of millions of works to build commercial products that compete directly with the human creators of those works. A purposive reading of the exception would struggle to extend it to this use, and no Indian court has yet addressed the question. The EU AI Act 2024 takes a transparency-based approach: it does not create a new copyright exception for AI training but requires providers of general-purpose AI models to publish detailed summaries of the content used in training. 19 This obligation, while not resolving the infringement question, creates a mechanism for rights holders to identify and challenge unauthorised use of their works — a step that Indian law does not yet contemplate.
The DPIIT Working Paper of December 2025 canvasses three regulatory models for India: extended collective licensing, under which AI developers pay into a central fund administered by collecting societies; compulsory licensing with remuneration, under which a statutory licence permits AI training against a government-set royalty; and an opt-out registry, under which works are presumed available for AI training unless the rights holder actively registers an objection. 20 The PSA White Paper of March 2026 recommends a hybrid model combining a statutory licence with a dedicated compensation fund. 21 Neither document is yet binding, but together they signal serious governmental engagement with the problem that is likely to inform forthcoming legislative activity. The training data problem is not merely legal — it is deeply ethical.
Human creators receive nothing when their works train AI systems, while developers profit from the resulting products. This structural asymmetry has mobilised global creative industry opposition, evidenced by the Authors Guild litigation and the Getty Images proceedings. Any sustainable Indian framework for AI and copyright must address this asymmetry directly, whether through licensing, remuneration, or mandatory attribution. VII. The Indian Legal Landscape
7.1 Statutory Framework
The Copyright Act, 1957 was last substantially amended in 2012, a decade before the generative AI revolution. Its core provisions on authorship and subsistence of copyright were drafted for a world in which all creative production was human. Section 2(d) defines 'author' for each category of protected work — the author of a literary work, the composer of a musical work, the artist for an artistic work — in terms that presuppose a natural person throughout. Section 13(1) conditions copyright on the work being 'original', a requirement the Supreme Court has held demands a minimum degree of creativity beyond mere mechanical reproduction. There is no definition of 'computer-generated work' in the Act, and no provision that addresses works generated by an AI system.
Section 17B has attracted scholarly attention as a possible route to protecting AI-generated works within the current statutory framework. By analogy with Section 9(3) CDPA 1988 — under which the person who makes the 'necessary arrangements' for a computer-generated work is treated as its author — it has been argued that the human or entity that 'causes' an AI to generate a work should be recognised as its author under Indian law. This interpretive argument has real merit, and if accepted by an Indian court, it would provide a workable mechanism for protecting AI outputs within the existing framework. But it remains entirely untested, and the analogy with the UK position is imperfect because Section 9(3) CDPA was enacted expressly and specifically, whereas the Section 17B argument rests on judicial implication. Sections 14 and 51 are directly in issue in the ANI v OpenAI proceedings.
Section 14 confers on the copyright owner the exclusive right to reproduce a work in any material form, including storing it in electronic or digital form. Section 51 provides that copyright is infringed where any such act is done without the owner's licence. The central question — whether ingesting news articles into an AI training corpus constitutes 'reproduction' within the meaning of Section 14 — remains unanswered.
7.2 Policy and Institutional Developments
The National IPR Policy of 2016 set out India's strategic vision for intellectual property but predates generative AI entirely and contains no provision relevant to AI authorship or training data. The DPIIT Working Paper of December 2025 and the PSA White Paper of March 2026 represent the first substantial governmental engagement with the intersection of AI and copyright in India, signalling that legislative reform is under active consideration. India's engagement at WIPO is equally significant. The WIPO Standing Committee on Copyright and Related Rights has conducted ongoing conversations on IP and AI since 2019,22 exploring disclosure requirements, authorship standards, and liability frameworks. India's domestic legislative choices will inevitably reflect and influence its positions in these multilateral fora.
A coherent and principled Indian framework would significantly strengthen the country's credibility and leverage at WIPO. VIII. Reform Proposals and the Way Forward Four principal models have been advanced in the literature and policy debate. This Part evaluates each before proposing a preferred approach for India.
8.1 Public Domain Default
The simplest approach is to accept that purely AI-generated works attract no copyright and fall into the public domain. This is the practical consequence of the current US position and the logical implication of any framework that requires human creativity as a threshold for protection. The advantages are real: the public domain is enriched, the philosophical integrity of copyright is preserved, and human creators are not undercut by AI competition. The disadvantage is that developers receive no protection for their AI outputs, potentially reducing the incentive to invest in generative AI tools — a concern that is more than trivial for a country with India's ambitions in the AI sector.
8.2 Work-for-Hire Extension
A second model extends the work-for-hire doctrine to AI developers, treating AI-generated works as belonging to the company that built and maintains the AI system — analogous to how films are owned by producers rather than individual crew members. This has intuitive appeal. The concern is concentration: extending copyright in this way would vest enormous amounts of intellectual property in the hands of a handful of technology corporations, potentially for the full copyright term, while doing nothing to compensate the human creators whose works trained the AI.
8.3 Sui Generis Right
A third model creates a purpose-built 'AI-generated works right' — a sui generis protection analogous to the EU database right23 that provides a shorter term of protection (perhaps fifteen years), is vested in the developer or deployer of the AI system, and is subject to exceptions for human creative use. This is architecturally elegant because it addresses the gap without distorting the existing copyright framework. Its drawback is legislative complexity and the risk that a narrowly drawn right fails to keep pace with rapid technological change.
8.4 Tiered Authorship Model (Recommended)
The fourth model — which this paper commends as the most workable for India — calibrates protection to the degree of human creative contribution. At one end of the spectrum, purely AIgenerated works with no meaningful human input receive no copyright and enter the public domain. In the middle, works involving substantial human selection, arrangement, modification, or direction attract full copyright protection in the human contributor. At the other end, a shortterm sui generis right is granted to AI developers for works generated with minimal human input, coupled with a mandatory remuneration obligation to compensate the rights holders whose works trained the AI. This model draws on converging jurisprudence in the United States (Zarya of the Dawn), China (Li Yunkai v Liu), and the underlying logic of the UK position under Section 9(3) CDPA.
It is consistent with the DPIIT's exploration of hybrid licensing models and the PSA's recommendation of a statutory licence with a compensation fund. Most importantly, it preserves space for human creativity at the top tier, incentivises AI investment at the bottom, and addresses the training data equity problem directly through a mandatory remuneration mechanism.
8.5 International Harmonisation
No domestic reform can succeed in isolation when AI systems, training data, and creative outputs routinely cross borders. The WIPO Conversation on IP and AI, ongoing through 2026, provides the natural multilateral forum for developing a harmonised framework. The existing instruments — the Berne Convention and TRIPS24 — are insufficient because they were not designed with AI in mind. India should participate actively in WIPO deliberations and advocate for a framework that balances AI investment against the rights of human creators and the public interest of developing countries.
The minimum elements of any such international framework should include: a common definition of 'AI-generated work'; a harmonised threshold for the degree of human contribution necessary to attract copyright; mandatory transparency obligations on AI developers regarding training data; and an international mechanism for compensating rights holders whose works contribute to AI training. None of these elements is technically beyond reach. The primary obstacle is political will among states whose AI industries have divergent interests. IX. Conclusion The question of who holds copyright in AI-generated content is not a technical puzzle in statutory interpretation alone. It sits at the intersection of philosophy, economics, and ethics, and the choices India makes in answering it will shape its AI industry, its creative economy, and its standing in global IP governance for a generation.
This paper has argued that existing Indian copyright law is structurally inadequate to address generative AI. Section 2(d) implies human authors throughout. Section 13 requires originality that presupposes human creativity. Section 17B offers a possible interpretive bridge, but it is untested and uncertain. The fair dealing exception under Section 52 is unlikely on a principled reading to extend to commercial AI training. And the ANI v OpenAI proceedings, while important as a first engagement, address training data liability rather than output ownership — leaving the core question entirely open. Comparatively, a detectable convergence is emerging across the US, UK, China, and EU: purely AI-generated works receive little or no protection; partial human creative contribution can attract copyright; and the training data question remains sharply contested.
India would be well-served by building on this emerging consensus rather than developing an idiosyncratic position that isolates it in international negotiations. The reform this paper recommends is a tiered authorship model: full copyright for works with substantial human creative contribution; no protection for purely autonomous AI output; and a short-term sui generis right for AI developers coupled with a statutory compensation mechanism for rights holders whose works trained the AI. This approach is philosophically defensible, practically workable, constitutionally sound, and aligned with the direction of travel in international IP governance. It is the architecture on which a just and sustainable Indian AI copyright regime can be constructed.