The Authorship Dilemma: Navigating Copyright Ownership in the Age of Generative AI

The problem of authorship without an author

Copyright law has developed on the premise that creative works originate with identifiable human authors whose intellectual labour issues in original expression. Authorship is not merely a symbolic term in copyright: it is the legal mechanism by which ownership, economic rights and control over creative works are established. In most systems copyright vests in the first instance in the author, and the whole scheme for the distribution of rights depends upon the possibility of identifying a human author. That author-centred model reflects the historical development of copyright in a world in which creative production could not be separated from human intellectual labour and personal creative judgment.¹ Authorship on that view performs both a moral and a proprietary function: it recognises creative personality while fixing ownership.

The philosophical foundation of copyright is likewise bound up with human authorship. Labour theory holds that a person acquires rights in the products of their intellectual work; personality theory treats creative works as an extension of the author’s personality; and utilitarian theory justifies protection as an economic incentive to human creativity. Those justifications differ in their premises but share an assumption, namely that copyright exists to protect and to encourage human creativity.² The requirement of human authorship is therefore not a technical rule but a structural foundation of the subject.

Generative artificial intelligence unsettles that foundation by introducing a mode of creative production in which expression is generated algorithmically rather than by direct human intellectual input. Models trained on large datasets can generate text, images, music and software code. The human contribution is in most cases confined to supplying a prompt, selecting among outputs or refining what is produced, rather than composing the expressive elements of the work. That raises a conceptual difficulty: if a human does not produce the expression directly, can a human be the author? The classical conception presupposes a causal connection between human creativity and the finished product, whereas generative systems interpose a detached computational process between human input and creative output.³

That interposition creates a legal grey area, because several parties contribute to a generated work. The user supplies the prompt, the developer supplies the algorithm, a company owns the system, and the training data shapes the output. Each participates in the production of the finished work and none is obviously an author in the traditional sense. Authorship is therefore distributed and difficult to attribute to any individual, which produces uncertainty of ownership, since copyright allocates ownership through authorship.⁴

The central legal question is accordingly who owns works generated by artificial intelligence. The candidates are the user, the developer, the company operating the system, or no one at all, in which case the work falls into the public domain. Each possibility exposes a structural weakness in copyright as applied to machine-generated creativity.⁵ Generative artificial intelligence therefore presents not merely a new legal question but a conceptual one: copyright rests on human authorship, while generative systems produce works without a determinate human author, leaving a gap in the determination of ownership.⁶

The collapse of traditional copyright logic

Generative artificial intelligence does not merely raise new questions in copyright; it exposes a structural weakness in existing doctrine. Copyright operates through a set of interrelated doctrines, among them originality, authorship, fixation, ownership and derivative works, which together determine whether a work is protected and to whom it belongs. Those doctrines were developed in a legal setting in which creative works were understood to be made directly by human beings. Applied to machine-generated works they yield inconsistent, contradictory or indeterminate results, which indicates a problem of incompatibility rather than of interpretation.

The first doctrine placed under strain is originality. Copyright protects only works that are original, requiring independent creation and a minimal degree of creativity, and originality has traditionally been understood by reference to human intellectual judgment. A generated work may be original in the sense that it copies no particular prior work, and yet may have no author. If originality requires human intellectual effort, then fully autonomous generation cannot yield a copyrightable work. If originality means no more than novel expression, generated works would qualify, but originality would then cease to be a proxy for human creativity, which would alter the foundation of the subject. The doctrine is therefore left uncertain as between creativity and mere novelty.⁷

Closely related is the requirement of human authorship, which is central to most systems. Courts and copyright offices have consistently proceeded on the footing that copyright arises only in works made by human beings. Generated works do not satisfy that requirement, because the expressive elements are produced by a machine. Where the user supplies only a prompt, it is unclear whether prompting constitutes authorship; and what the developer did in building the algorithm is technological rather than expressive. Many generated works may therefore fail the test of human authorship altogether and be incapable of copyright.⁸

The dichotomy between idea and expression is likewise difficult to apply. Copyright protects expression and not ideas. Where a user supplies prompts to a generative system, those prompts may resemble ideas, instructions or conceptual input, while the system produces the final expression. That raises the question whether the user contributed expression or only an idea. If prompts are ideas, the expressive element must be attributed to a system that cannot be an author. If prompts are expression, the user may be the author. The doctrine becomes unstable in prompt-based creative processes, because the line between idea and expression is difficult to locate.⁹

Ownership presents a further difficulty. Copyright ordinarily vests ownership in the author, and every subsequent dealing proceeds from that initial allocation. Where authorship cannot be determined, ownership cannot be determined either. That leaves a gap in which the user, the developer, the platform owner or the creators of the training data may each assert a claim, while none of them clearly qualifies as author under traditional doctrine.¹⁰

Other doctrines fit no better. Work made for hire cannot readily apply, because an artificial system cannot be an employee. Joint authorship requires an intention to contribute to a unitary whole, which is difficult to establish as between a human and a machine. Where systems have been trained on copyrighted works it is unclear whether the outputs are derivative works, and substantial similarity is hard to establish because the outputs are not direct copies. In the absence of a human creator, generated content may fall into the public domain by default, which may discourage investment in creative industries built upon these technologies.¹¹

Taken together, these inconsistencies show that the problem of authorship is not a recent difficulty capable of correction by modest doctrinal adjustment. It reflects a deeper incompatibility between copyright and machine-generated creativity. Copyright assumes human creativity, identifiable authorship and direct creative control; generative systems operate through distributed input, probabilistic generation and indirect human involvement. The conflict between the two produces not the adaptation of doctrine but its collapse, which suggests the need for a new framework rather than the reinterpretation of existing rules.¹²

Mapping the possible claimants

A. Reframing the inquiry from doctrines to stakeholders

The preceding part showed that traditional principles fail when applied to generated works. The question of authorship is better approached by identifying the competing claimants who might own such works. The difficulty is not only doctrinal incompatibility but the presence of several participants, each playing a different part in the creative process, none of whom can readily be described as the author. This part therefore shifts the analytical lens to the practical question whether any of the following has a legitimate claim to authorship under existing principles: the user, the developer, the owner of the system or platform, or the contributors of training data.

B. The user: creative control or creative input

The user is the most intuitive claimant, since it is the user who engages with the system by supplying prompts and selecting outputs. The argument for user authorship rests on the proposition that prompt engineering, iterative refinement and the selection of outputs amount to creative control and so satisfy the requirement of originality.

The claim is nonetheless doctrinally weak. Copyright requires direct control over the expression and not merely the giving of directions or ideas, and the courts have consistently held that ideas are not protected but only their expression.¹³ In this setting prompts are frequently characterised as conceptual direction rather than expressive contribution, so that it is questionable whether the user creates the work at all.

Administrative guidance points the same way. The United States Copyright Office has made clear that a work generated by artificial intelligence cannot be registered where it does not contain sufficient human authorship, and that prompts alone do not constitute authorship.¹⁴ Most users will therefore not qualify as authors unless they exercise substantial creative control over the output. The most intuitive claimant thus has a claim that depends on a degree of control which the law leaves undefined.

C. The developer: toolmaker or creator

A second claimant is the developer who builds the system, its architecture and its generative capacities. Developers are essential to the production of output, but copyright distinguishes between making a tool and making the expression the tool produces.

That distinction is embedded in copyright jurisprudence: the maker of a technology does not thereby acquire rights in the works created through its use.¹⁵ Developers create generative frameworks; they neither determine nor predict the output produced in response to a particular prompt.

Authorship moreover requires a causal connection between the intellectual work of the author and the resulting expression. In generative systems that connection is indirect and probabilistic rather than intentional and expressive. To attribute authorship to developers would extend copyright beyond its doctrinal limits and risk the monopolisation of machine-generated expression. Developers are therefore better described as facilitators of creativity than as authors.

D. The platform owner: economic control rather than creation

The third claimant is the platform owner or the entity that deploys and operates the system, whose claim ordinarily rests on economic investment, control of infrastructure and contractual arrangements.

Copyright does not recognise authorship on the basis of investment or ownership of infrastructure. Originality requires independent creation and a minimal degree of creativity, not effort or expenditure.¹⁶ To vest ownership in the platform owner would convert copyright into a regime of economic control rather than of creative protection. Platforms may exercise de facto control through licences and terms of service, but that control derives from contract and not from copyright. Platform owners therefore have no legitimate claim to authorship under existing doctrine.

E. Contributors of training data: the invisible participants

A more complex group of claimants comprises the creators of the works on which systems are trained. These contributors supply the data from which models learn patterns and produce outputs, which raises the question whether generated works are derivative of the training data. A derivative work must be based upon one or more pre-existing works and must display substantial similarity to them,¹⁷ whereas generated outputs are typically non-deterministic and do not reproduce recognisable portions of any single work.

Although the creators of training data may have legitimate concerns that their works were used without permission, they cannot readily establish authorship of the outputs. Their claims are better characterised as claims in infringement or in licensing than as claims to authorship.

F. The public domain by default

Since none of the identified participants satisfies the conventional requirements of authorship, the conclusion that generated works have no author at all is a tenable legal position.

That position accords with judicial authority denying non-human authorship. In Naruto v. Slater the Court of Appeals for the Ninth Circuit held that human authorship is a precondition of copyright and that non-human entities have no standing under the statute,¹⁸ and administrative authorities have reiterated that copyright subsists only in works of human origin.¹⁹

Generated works might therefore pass automatically into the public domain in the absence of a human creator. That result preserves doctrinal consistency, but it creates serious policy difficulties, among them the absence of incentives to innovate and the prospect of uncontrolled proliferation of machine-generated content.

G. The failure of the single-author model

Mapping the claimants shows that the difficulty lies not in identifying the right author but in the inadequacy of the single-author model itself. Generated works result from distributed contributions, layered inputs and probabilistic processes, none of which conforms to the classical paradigm of a work made by one person.

The question of authorship therefore cannot be resolved within that framework. What is required instead is a reconception of authorship capable of accommodating varying degrees of human intervention and technological mediation, which is the framework developed below.

Comparative approaches across jurisdictions

The problem has attracted inconsistent treatment across jurisdictions, revealing differences not only of doctrine but of legal philosophy. Some systems adhere to human authorship while others experiment with partial recognition of machine-generated content, producing a fragmented international landscape.

Copyright in the United States is firmly anchored in human authorship. Judicial and administrative decisions have consistently refused protection to works lacking a human element of creation, and the Copyright Office has made clear that registration is not available for works generated by artificial intelligence rather than by a person.²⁰ That approach preserves the philosophical basis of copyright as a reward for human intellectual labour, but leaves most generated products unprotected and effectively in the public domain.

The United Kingdom takes a more pragmatic course by recognising computer-generated works by statute. The Copyright, Designs and Patents Act 1988 defines the author of such a work as the person by whom the arrangements necessary for its creation are undertaken,²¹ which applies even where the final expression was not made by that person, and so vests authorship in whoever arranged the generative process. The model has been criticised for uncertainty, particularly in complex systems involving many participants.

The European Union is more doctrinal, proceeding from the principle of the author’s own intellectual creation. The Court of Justice has held consistently that originality requires the expression of the author’s own intellectual creation,²² which implicitly excludes generated works from protection. That preserves philosophical consistency but gives little guidance on works produced through varying degrees of human and machine collaboration.

In India the statute defines the author of a computer-generated work as the person who causes the work to be created, which might in principle extend to generated works. The point is not expressly addressed and judicial interpretation remains undeveloped,²³ leaving an area of uncertainty in which generated works may be protected in theory but not in practice.

China has taken a more permissive approach, with courts in some cases recognising a measure of protection for generated works where a human contribution to the creative process can be shown.²⁴ That reflects a policy-driven approach seeking to balance technological innovation against the protection of intellectual property.

Australia, like the United States, has emphasised human authorship and has not treated generated works as copyrightable, although policy consideration of reform continues.²⁵

The comparison shows that there is no international consensus. Some jurisdictions preserve doctrinal consistency by insisting upon human authorship, while others are pragmatic or experimental in accommodating technological change. That divergence creates legal uncertainty for works crossing borders and underlines the absence of an effective international framework. The resulting fragmentation reinforces the central difficulty, namely that a copyright system founded on human authorship is ill-adapted to machine-generated creativity.

Can a machine be an author

The problem is not merely legal but philosophical, because copyright has deep roots in human conceptions of creativity, personality and intellectual labour. To ask whether a machine can be an author requires an examination of authorship as a jurisprudential and philosophical concept. Authorship traditionally connotes originality, purpose and creative judgment. An author is not merely a producer of content but a mind that deliberately seeks to create expression. Generative systems complicate that definition by producing expressive works without awareness, will or personality, which raises the question whether authorship must necessarily be connected to a human mind.

Among the principal philosophical arguments is the distinction between creativity and computation. Human creativity is ordinarily associated with imagination, intention and subjective choice, whereas generative systems operate through algorithmic calculation and probabilistic pattern generation. Although the output may appear creative, the process by which it is produced is computational rather than intentional. The question is therefore whether copyright protects creative outcomes or creative processes. If the originality of the finished expression is what is protected, generated works may qualify. If copyright protects human creative intention and intellectual labour, there is no rationale for extending it to them.²⁶

The difficulty deepens when the major philosophical theories are applied. Personality theory, associated with Hegel, treats creative works as an extension of the author’s personality and expression, and so presupposes a personal relationship between author and work that a machine cannot have, lacking personality, consciousness and identity. Labour theory, associated with Locke, justifies property rights by the mixing of a person’s labour with resources; generative systems do not labour in any moral or philosophical sense but execute operations upon data according to their programming. On either theory a machine cannot be an author.²⁷

Utilitarian and incentive theories point in another direction. They justify copyright not because authors are entitled to protection but because society benefits from encouraging the production of works. On that view the identity of the author matters less than the economic and social incentives that protection creates. If conferring copyright in generated works would increase innovation, investment and technological development, protection might be justified notwithstanding the absence of human creativity. The emphasis shifts from moral authorship to economic policy.²⁸

A further difficulty concerns moral rights, which most systems associate with authorship. The rights of attribution and of integrity both presuppose a close relationship between author and work. They make no sense in relation to machine-generated works, since a generative system has no reputation, dignity or artistic integrity to protect. That again supports the view that the traditional conception of authorship is anthropocentric.²⁹

The philosophical analysis therefore shows that the dilemma is not one of legal classification but a conceptual challenge to the premises of copyright itself. Copyright is founded on human creativity, human personality and human labour, and generative systems unsettle each of those assumptions. Whether ownership of generated works remains the central question, or whether authorship must remain human-centred, is open to argument. Generative artificial intelligence compels copyright to revisit, and perhaps to redefine, what authorship means.³⁰

The policy risks of leaving the question unregulated

The question is often treated as hypothetical or futuristic, but leaving it unregulated carries immediate legal, economic and policy risks. Without rules governing authorship and ownership of generated works, disputes and litigation are likely to multiply. As generated content becomes common in publishing, advertising, software development and entertainment, several stakeholders may claim ownership of the same work. Without a clear framework the courts will decide case by case, producing inconsistency and uncertainty.³¹

A further risk is that large technology companies operating generative systems may come to monopolise generated content. Even if such works attract no copyright, companies may secure de facto protection through licensing agreements, terms of service and technical restrictions on access. That could allow a small number of corporations to dominate the creative industries built on these technologies, suppressing competition and concentrating economic power.³²

Conversely, where generated works pass automatically into the public domain for want of human authorship, the result may be a flooding of the public domain, in which large quantities of machine-generated content saturate markets and depress the economic value of human work. That may in turn discourage human creativity, particularly in illustration, writing, music and design, where generated content can be produced in volume at low cost.³³

A further problem concerns the use of copyrighted material in training datasets. Most generative systems are trained on protected content without express permission, which raises contested questions of data scraping, fair dealing and remuneration. Where outputs are exploited commercially and the creators of the training data receive nothing, a serious economic and ethical difficulty arises in the creative sector.³⁴

Finally, the absence of regulation creates a policy dilemma between innovation and protection. Excessive protection may stifle technological advance, while insufficient protection may harm human creators and creative enterprises. The problem is therefore not only legal but economic and social. Regulation is needed not merely to settle ownership but to hold a balance between technological innovation, market competition and the preservation of human creativity in the market.³⁵

Towards a new copyright framework

Existing doctrine is not sufficiently coherent to resolve the authorship of generated content, and attempts to force such content into the traditional framework produce doctrinal conflict, uncertainty of ownership and policy risk. What is required is not the enlargement of existing doctrine but a framework designed for machine-generated works, one that separates authorship, creative contribution and ownership and recognises degrees of human involvement rather than a binary between authorship and its absence.

Such a framework should begin by redefining authorship not as the creation of expression but as the exercise of creative control over the process of generation, with the degree of human involvement determining whether authorship arises. Where a person supplies detailed prompts, iterative instruction, selection, editing and creative direction, that person should be regarded as the author, the system operating as a tool assisting human creativity. Where the system produces work with minimal human intervention, traditional authorship should not arise. Human involvement thus becomes the qualifying criterion.³⁶

On that basis a tiered scheme of ownership can be constructed. At the first tier, where the human role is substantial and creative control is evident, the human user should be recognised as author and full owner. At the second, where human intervention exists but is slight, ownership may be allocated without full authorship, through a limited copyright or a neighbouring rights regime. At the third, where works are generated autonomously without significant human contribution, they should fall into the public domain. A tiered scheme of this kind recognises the continuum of human and machine collaboration rather than forcing generated works into a binary classification.³⁷

The framework must also address the allocation of rights. Subject to contract, ownership should ordinarily vest in the person exercising creative control or in the entity commissioning the work. Developers should not ordinarily control outputs unless they exercise direct control over the creative process. Economic rights of reproduction, distribution and commercial exploitation may be allocated to the identified owner, while moral rights should be confined to human creators, since they rest on personality and reputation.³⁸

The term of protection should also differ from the conventional term. Because generated works can be produced at scale and at low cost, full-term protection would risk over-monopolisation. A shorter term would provide an incentive to innovate while preserving public access. An optional register of generated works might also be introduced, so as to make the extent of human involvement and the basis of any ownership claim transparent.³⁹

Liability for infringement must likewise be addressed. Where a generated work infringes an existing copyright, liability should follow control over the process of generation. A user who knowingly produces infringing material should be liable. Where the infringement originates in the training data or in the design of the system, developers or operators may bear a share of responsibility. That distributed approach to liability corresponds to the distributed character of generative creativity.⁴⁰

The framework therefore indicates that the solution lies not in deciding whether a machine can be an author but in reforming copyright to recognise varying degrees of human input, technological mediation and economic control. A contribution-based and tiered scheme offers a more coherent and policy-driven answer than the attempt to apply traditional doctrines of authorship to machine-generated creativity.

Resolving the dilemma

The copyright difficulty presented by generative artificial intelligence is a dilemma of authorship arising from a structural conflict between traditional copyright and technologically mediated creativity. Copyright proceeds on the footing that a protected work issues from a human author exercising creative control over the expression, and that ownership follows authorship. Generative systems disturb that scheme by producing works in which the contributions are dispersed among users, developers, datasets and algorithms, so that no single human author can readily be identified. The difficulty lies not only in determining ownership but in the wider incompatibility between a humanist copyright law and machine-assisted creative production.⁴¹

Traditional doctrines are ineffective in this setting because they rest on originality, human authorship and direct creative expression, none of which corresponds to the operation of generative systems. Comparative approaches are inconsistent, some jurisdictions refusing protection and others extending authorship through statutory interpretation or policy, which reflects the absence of a settled theoretical position.⁴² Generated works are meanwhile difficult to accommodate within the philosophical theories of copyright, particularly personality and labour theory, since machines have neither personality, nor intention, nor moral agency. Copyright is therefore confronted with a fundamental question: whether authorship must remain tied to human creativity, or whether ownership may be justified on other grounds such as control, investment or economic incentive.⁴³

This article has argued that the answer lies in a new framework separating authorship, creative contribution and ownership, and applying a tiered approach to generated works calibrated to the degree of human intervention. Such a framework recognises the reality of distributed creativity and offers a more workable basis for allocating ownership, distributing rights and assigning liability than models founded on traditional authorship. By shifting the emphasis from authorship to creative control and contribution, copyright may adapt to technological change without abandoning its object of encouraging creativity and facilitating access to knowledge.⁴⁴

The resolution of the dilemma therefore lies not in determining whether artificial intelligence can be an author but in reformulating the treatment of creativity, authorship and ownership in copyright law. Whether machines can create will matter less to the future of the subject than whether the law can redefine what it means to create.⁴⁵

*****

Footnotes

1. Jane C. Ginsburg, The Concept of Authorship in Comparative Copyright Law, 52 DePaul L. Rev. 1063 (2003).

2. Justin Hughes, The Philosophy of Intellectual Property, 77 Geo. L.J. 287 (1988).

3. Annemarie Bridy, Coding Creativity: Copyright and the Artificially Intelligent Author, 5 Stan. Tech. L. Rev. 1 (2012).

4. Pamela Samuelson, Allocating Ownership Rights in Computer-Generated Works, 47 U. Pitt. L. Rev. 1185 (1986).

5. U.S. Copyright Office, Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence (2023).

6. World Intellectual Property Organization, Revised Issues Paper on Intellectual Property Policy and Artificial Intelligence (2020).

7. Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991).

8. U.S. Copyright Office, supra note 5.

9. Bridy, supra note 3.

10. Samuelson, supra note 4.

11. World Intellectual Property Organization, supra note 6.

12. Jane C. Ginsburg & Luke Ali Budiardjo, Authors and Machines, 34 Berkeley Tech. L.J. 343 (2019).

13. Baker v. Selden, 101 U.S. 99, 103 (1879).

14. U.S. Copyright Office, Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 88 Fed. Reg. 16,190 (Mar. 16, 2023).

15. See Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53, 58 (1884).

16. Feist Publications, Inc., supra note 7, at 345.

17. 17 U.S.C. § 101 (defining a derivative work).

18. Naruto v. Slater, 888 F.3d 418, 426 (9th Cir. 2018).

19. U.S. Copyright Office, Compendium of U.S. Copyright Office Practices § 306 (3d ed. 2021).

20. U.S. Copyright Office, supra note 5.

21. Copyright, Designs and Patents Act 1988, c. 48, § 9(3) (U.K.).

22. Case C-5/08, Infopaq International A/S v. Danske Dagblades Forening, 2009 E.C.R. I-6569.

23. The Copyright Act, 1957, § 2(d), No. 14, Acts of Parliament, 1957 (India).

24. Feilin v. Baidu, Beijing Internet Court (2019) (China).

25. Australian Government, Department of Industry, Science and Resources, Artificial Intelligence and Intellectual Property Policy Review (2021).

26. Bridy, supra note 3.

27. Hughes, supra note 2.

28. William M. Landes & Richard A. Posner, The Economic Structure of Intellectual Property Law (2003).

29. Ginsburg, supra note 1.

30. Ginsburg & Budiardjo, supra note 12.

31. Ginsburg & Budiardjo, supra note 12.

32. Ryan Abbott, Artificial Intelligence and Copyright, 21 Stan. Tech. L. Rev. 1 (2018).

33. World Intellectual Property Organization, supra note 6.

34. Samuelson, supra note 4.

35. U.S. Copyright Office, supra note 5.

36. Ginsburg & Budiardjo, supra note 12.

37. Abbott, supra note 32.

38. Bridy, supra note 3.

39. World Intellectual Property Organization, supra note 6.

40. Samuelson, supra note 4.

41. Ginsburg & Budiardjo, supra note 12.

42. World Intellectual Property Organization, supra note 6.

43. Hughes, supra note 2.

44. Abbott, supra note 32.

45. Bridy, supra note 3.