Who Is Liable For AI-Created Content - Copyright Act And Civil Responsibility In Aotearoa New Zealand

Author: LawSphereX

Category: Intellectual Property

Published: 2026-08-23T21:43:29.145913+00:00

Updated: 2026-08-26T17:05:00.641+00:00

WHO IS LIABLE FOR AI-CREATED CONTENT?

AOTEAROA NEW ZEALAND

Abhinav Singh, M.A., B.A., LL.B., Panjab University, Chandigarh Orcid ID: https://orcid.org/0009-0003-9089-3518

ABSTRACT

The proliferation of generative artificial intelligence (AI) technologies has raised complex questions about the allocation of civil liability in copyright law. In Aotearoa New Zealand, the Copyright Act 1994 offers no statutory guidance on how to assess works produced by AI systems, nor does it define authorship or originality in contexts where human input may be minimal or absent. This article examines whether the existing legal framework is equipped to address the risks of copyright infringement arising from generative outputs, particularly where such outputs are derived from copyrighted training data or imitate protected styles.

Through doctrinal analysis and comparative evaluation of legal developments in the United Kingdom, the United States, and the European Union, the article demonstrates that Aotearoa New Zealand lags in articulating a position on derivative and transformative use in the context of AI-generated content. It further identifies critical evidentiary and procedural gaps that make enforcement difficult for rights-holders and create uncertainty for developers and users alike. In response, the article proposes targeted legislative reforms, including statutory clarification of AI-related authorship, safe harbours for unintentional reproduction, mandated transparency in training datasets, and the establishment of a specialised tribunal to resolve AI-related copyright disputes.

This article contributes to an emerging body of legal scholarship on AI and copyright by offering one of the first jurisdiction-specific analyses of liability in the Aotearoa New Zealand context. It argues that without timely and principled reform, the existing regime will struggle to maintain legal certainty and doctrinal coherence in the face of rapidly evolving generative technologies. Keywords: AI-generated Content; Copyright Liability; Authorship and Originality; New Zealand Copyright Act 1994; Generative Artificial Intelligence

Introduction The rapid integration of generative technologies within the creative industries has prompted renewed scrutiny of copyright frameworks globally. In Aotearoa New Zealand, the Copyright Act 1994 continues to operate on traditional assumptions of human authorship and intentional creation. However, as artificial intelligence (AI) systems capable of producing literary, musical, artistic, and audiovisual works become more prevalent, the foundational premises of copyright law face considerable strain. Aotearoa New Zealand’s creative industries are already beginning to confront these concerns. The Screen Production and Development Association (SPADA) has warned that generative platforms trained on publicly available scripts may produce outputs that mimic dialogue, narrative structures, or stylistic patterns from copyrighted works, often without author consent or licensing.

This has raised questions around ownership, protection, and attribution within the domestic screenwriting sector, and further underscores the need for urgent doctrinal clarity. 1 Unlike earlier tools of creativity, many AI models operate with minimal or no direct human input in the final expression of content, thereby raising intricate questions about ownership, originality, and liability. Of particular concern is the question of civil liability when generative outputs either reproduce or are substantially similar to protected works, especially in circumstances where those outputs may not be directly traceable to any one user or dataset. New Zealand law, like that of other common law jurisdictions, distinguishes between ‘original’ and ‘derivative’ works for the purposes of copyright protection and infringement.

A derivative work typically involves a modified or adapted version of an existing work, whereas an original work must be the product of independent intellectual effort. The use of AI complicates these categories. While the Act does not explicitly address machine-generated or machine-assisted outputs, it continues to apply a framework based on human authorship and conscious expression. This leaves a doctrinal gap when assessing the liability of individuals who employ generative AI tools that may intentionally or otherwise reproduce aspects of preexisting copyrighted material.

In the absence of a statutory doctrine of fair use akin to that in the United States, Aotearoa New Zealand courts rely on narrower statutory exceptions and traditional infringement tests, such as whether a ‘substantial part’ of the original work has been 1 Screen Production and Development Association (SPADA), Annual Report 2023: Aotearoa’s Screen Sector in Transition (SPADA, 2023) https://www. spada. co. nz/resources accessed 3 June 2025.

copied. 2 Although the case laws in Aotearoa New Zealand on AI-generated content remain limited, analogous developments in other common law jurisdictions provide an interpretative foundation. The United States Supreme Court decision in Andy Warhol Foundation for the Visual Arts v Goldsmith has clarified the limits of transformative use in a copyright context, concluding that commercial licensing of an AI-modified image did not amount to fair use where the new work preserved the essential elements of the original. 3 This line of reasoning may influence Aotearoa New Zealand courts, which have historically drawn upon international jurisprudence where domestic authority is lacking. Nevertheless, Aotearoa New Zealand’s reliance on a strict statutory framework rather than an open-ended fair use exception poses unique challenges when applied to the question of AI-mediated content generation.

This article examines the doctrinal and policy implications of civil liability for copyright infringement where AI systems are involved in generating or transforming existing works. It seeks to address the extent to which Aotearoa New Zealand’s copyright framework can accommodate such cases under its current structure, and whether liability should attach to the user, the developer, or the platform operator. By analysing the distinction between derivative and transformative use in domestic and comparative law, and assessing potential reforms, this research aims to contribute to a more coherent and technologically attuned understanding of copyright liability in Aotearoa New Zealand. Structure and Scope of Copyright Protection under Aotearoa New Zealand Law The Copyright Act 1994 serves as the principal legislative instrument governing copyright protection in Aotearoa New Zealand.

It provides exclusive rights to authors and rights-holders in relation to various categories of works, including literary, dramatic, musical, and artistic works, as well as sound recordings, films, and communication works. 4 The Act adopts a rights-based model of protection, where infringement is primarily framed around the unauthorised exercise of exclusive rights such as reproduction, communication, and adaptation. Importantly, the Act’s underlying assumptions rest on the premise of human authorship and the 2 Copyright Act 1994 (NZ), s 29; Hawkes & Sons (London) Ltd v Paramount Film Service Ltd [1934] 1 Ch 593 (CA).

3 Andy Warhol Foundation for the Visual Arts, Inc v Goldsmith 143 S Ct 1258 (2023).

4 Copyright Act 1994 (NZ), s 14(1).

application of intellectual skill and labour, concepts which are increasingly strained by the proliferation of machine-generated and AI-assisted outputs. The statutory definition of a “copyright work” under section 14 of the Act presupposes originality, which, under common law interpretation, does not demand novelty but instead a modicum of skill, judgment, and labour on the part of the author. 5 The decision in University of London Press Ltd v University Tutorial Press Ltd affirmed that originality does not denote innovation but rather the personal intellectual effort involved in the expression of an idea. 6 Aotearoa New Zealand courts have followed this interpretation, maintaining a relatively low threshold for originality so long as there is no direct copying and the work is not a product of mere mechanical reproduction.

This jurisprudential approach creates doctrinal ambiguity when assessing AI-mediated works, particularly where the creative input may be algorithmically determined or generated through stochastic processes without direct human involvement in the final expression. It is necessary at this point to distinguish between AI-assisted and AI-generated works. AI-assisted works involve meaningful human contribution—such as editing, prompting with intention, or exercising creative control over the output—whereas AI-generated works are autonomously created by a system without human shaping of form or content. This distinction is critical under New Zealand’s Copyright Act 1994, which presupposes a human “author” under section 5(1) and requires that copyright subsists only in works originating from human intellectual effort.

7 While this distinction is not explicitly recognised in the statute, it has significant implications for the subsistence of copyright and the assignment of liability where infringement occurs. Authorship, as a cornerstone of copyright law, is similarly tethered to human agency. Section 5 of the Act defines the “author” in relation to a literary, dramatic, musical, or artistic work as the person who creates it, without recognising artificial agents or automated systems as legal authors. 8 In contrast to the United Kingdom’s Copyright, Designs and Patents Act 1988, which under section 9(3) attributes authorship of computer-generated works to the person who undertakes the arrangements necessary for the creation, the Aotearoa New Zealand

5 Peterson v Workmen’s Compensation Board [1959] NZLR 404 (CA).

6 University of London Press Ltd v University Tutorial Press Ltd [1916] 2 Ch 601 (Ch).

7 Copyright Act 1994 (NZ), s 5(1); see also University of London Press Ltd v University Tutorial Press Ltd [1916] 2 Ch 601 (Ch) (defining originality in terms of intellectual effort).

8 Copyright Act 1994 (NZ), s 5(1).

Act remains silent on such matters. The absence of a statutory provision addressing the legal status of AI-generated or AI-assisted content poses serious questions about the subsistence of copyright in such outputs and the potential attribution of liability where infringement arises. Moreover, the concept of derivative works is treated conservatively under Aotearoa New Zealand law. Although the term ‘derivative’ is not defined within the Act itself, derivative creation is effectively governed by provisions relating to adaptations and translations, both of which require consent from the rights-holder of the original work. Section 29 prohibits the unauthorised adaptation of copyright works, with infringement established by demonstrating that the resulting work reproduces a substantial part of the original.

9 The test for substantiality, though not rigidly defined, considers both qualitative and quantitative significance of the portion copied, as reaffirmed in Hawkes & Sons (London) Ltd v Paramount Film Service Ltd. 10 In practice, this standard necessitates a contextual analysis of how the new work relates to the original, which becomes particularly complex in the context of algorithmic remixing, where outputs may amalgamate fragments of multiple works or draw stylistic influence without direct reproduction. Infringement, under section 29 and subsequent provisions, may result in civil liability where exclusive rights are exercised without a licence or exception. Civil remedies under section 120 include injunctions, damages, account of profits, and delivery up of infringing copies.

These remedies apply equally to primary and secondary infringers, the latter encompassing those who deal with infringing copies in the course of trade. However, the Act does not yet accommodate situations where the reproduction is effected indirectly through algorithmic generation, leaving a legislative lacuna in attributing liability for outputs produced without traditional human creative processes. The limitations inherent in the current legal framework become more pronounced when evaluated against emerging forms of AI-mediated expression. As the subsequent sections will demonstrate, the dichotomy between derivative and transformative use—although welltheorised in jurisdictions such as the United States—lacks statutory foundation or judicial elaboration in Aotearoa New Zealand.

This results in a significant degree of uncertainty both for users of AI systems and for original rights-holders whose works may be used, often

9 Copyright Act 1994 (NZ), s 29.

10 Hawkes & Sons (London) Ltd v Paramount Film Service Ltd [1934] 1 Ch 593 (CA).

unknowingly, in the training or prompting of generative models. Absent doctrinal reform, Aotearoa New Zealand’s copyright regime may struggle to delineate the contours of liability in this rapidly evolving technological context. Generative Systems and the Emerging Landscape of Copyright Infringement The integration of generative technologies into content production processes has significantly complicated the assessment of copyright infringement under existing legal doctrines. Such technologies, particularly those using deep learning architectures, are trained on large datasets comprising text, images, music, or other forms of creative content, many of which may be protected by copyright.

These systems, including but not limited to transformerbased models for text and diffusion models for images, generate outputs based on probabilistic patterns learned from their training material, rather than through direct copying in the traditional sense. Nevertheless, the risk that these outputs may closely resemble, replicate, or reconstruct protected works is no longer speculative. Several high-profile legal disputes and scholarly critiques have underscored the growing concern that generative systems, whether operated by developers or end users, may give rise to unauthorised reproductions within the meaning of copyright law. 11 The process by which these outputs are produced lacks transparency, complicating the evaluation of infringement.

Machine learning models are typically described as “black boxes” due to the absence of a clear, traceable path from input data to output expression. As a result, determining whether a particular image, piece of music, or paragraph generated by such systems constitutes a substantial reproduction of a copyrighted work is inherently difficult. Furthermore, because the systems are trained using vast corpora, often harvested from online repositories without granular attribution or licensing, the specific origin of particular elements in the output may be unknowable to the user or even to the model developer. 12 This evidentiary opacity presents challenges when applying traditional legal tests for infringement.

Under Aotearoa New Zealand law, establishing civil liability for copyright infringement involves demonstrating that the alleged infringing work copies a substantial part 11 Andres Guadamuz, ‘Artificial Intelligence and Copyright’ (2017) 39(1) European Intellectual Property Review 18; Sarah Silverman v OpenAI Inc (Case No. 3:23-cv-03416, N. D. Cal, filed 7 July 2023).

12 Margot E Kaminski, ‘The Right to Explain AI Decisions’ (2019) 68(1) UCLA Law Review 74, 92–94.

of the original, and that the defendant had access to the original work. 13 In the context of generative systems, access is indirect and embedded within the model’s training phase, which is typically conducted by third-party developers and not the end users. Moreover, the output may contain only fragments or stylistic elements of an original work, raising questions about whether it meets the threshold for substantial reproduction. The issue has already come before the courts in jurisdictions such as the United States. In Andersen v Stability AI, a group of visual artists alleged that Stable Diffusion’s generative image model had been trained on copyrighted artworks and produced outputs imitating their distinctive styles and compositions without permission or attribution.

14 While the claim is still pending, it illustrates the evidential and legal complexity of determining whether statistical mimicry constitutes infringement. These cases signal the kind of disputes New Zealand may soon face, particularly as similar platforms become widely accessible to local users and content creators. For example, a user prompting a model to produce a digital portrait “in the style of Ralph Hotere” may receive an output that reflects key aesthetic attributes associated with the artist’s oeuvre, without reproducing any particular painting. Whether such mimicry constitutes infringement depends on a qualitative assessment of the artistic expression being replicated and whether it forms a protectable element under the Copyright Act. 15 This issue is further complicated by the differing degrees of human involvement in the generation of such outputs.

In some cases, the user provides minimal prompts and exerts little control over the final result. In others, users may iteratively refine their input or post-process the output, thereby contributing creative input that could be characterised as authorship. The question then arises whether liability should attach to the user as the proximate cause of the infringing act, to the developer for facilitating the infringing potential through model training, or to both under a theory of joint responsibility. Consider the following scenario: A freelance designer in Wellington uses a text-toimage generator hosted by an overseas platform to create promotional artwork for a client. The user prompts the system to generate an image “in the style of a well-known local Māori artist.

” The resulting image adopts distinctive compositional and cultural motifs found in the original 13 Copyright Act 1994 (NZ), s 29; Hawkes & Sons (London) Ltd v Paramount Film Service Ltd [1934] 1 Ch 593 (CA).

14 Andersen v Stability AI Ltd (ND Cal, Case No. 3:23-cv-00201, filed 13 January 2023).

15 Copyright Act 1994 (NZ), s 14(1); Bonz Group (Pty) Ltd v Cooke [1994] 3 NZLR 216 (HC).

artist’s work. If that output is used commercially, potential liability may arise at multiple levels: the user for deploying a prompt with foreseeable risk, the platform provider for offering a model trained on copyrighted Māori art without attribution or license, and the developer for incorporating those materials into the training dataset. This example illustrates how legal uncertainty surrounding knowledge, control, and foreseeability complicates enforcement and attribution of liability in AI-driven creative workflows. In the absence of statutory direction, Aotearoa New Zealand law provides limited guidance for assessing such diffuse accountability. Additionally, there is currently no legislative requirement in Aotearoa New Zealand for AI developers or platforms to disclose the datasets on which their models are trained.

This contrasts with proposed reforms in the European Union under the draft Artificial Intelligence Act, which includes transparency obligations for high-risk systems, including documentation of training data. 16 The lack of such requirements in Aotearoa New Zealand contributes to the legal uncertainty for both content creators and users, as there is no way to ascertain whether a given model has been trained on infringing content or whether safeguards are in place to prevent the reproduction of protected works. In light of these developments, it is evident that generative technologies have introduced novel forms of legal risk that are not adequately addressed by Aotearoa New Zealand’s current copyright regime.

The traditional emphasis on direct copying, clear authorship, and observable access does not easily accommodate the probabilistic and distributed nature of machine-generated outputs. Moreover, the divergence between the technological process of generation and the legal standards for infringement creates interpretative challenges for courts tasked with adjudicating disputes in this domain. These tensions will be further explored in the subsequent section, which examines the comparative treatment of derivative and transformative works in jurisdictions that have begun to engage with the problem of AI-mediated creativity.

The lack of statutory recognition for AI platform operators within Aotearoa New Zealand’s copyright enforcement framework leaves users and developers to shoulder the entirety of legal risk, despite the increasingly central role platforms play in content generation and dissemination. 16 European Commission, ‘Proposal for a Regulation Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act)’ COM (2021) 206 final.

Derivative versus Transformative Use: Comparative Legal Perspectives on Copyright and Generative Output The legal characterisation of outputs created through machine learning and generative systems has brought renewed focus to the distinction between derivative and transformative works. While Aotearoa New Zealand’s Copyright Act 1994 does not expressly employ the term “transformative use”, this concept has gained doctrinal prominence in jurisdictions that have adopted broader fair use frameworks, such as the United States. The relevance of these comparative approaches lies not in their direct applicability to Aotearoa New Zealand, but in their conceptual frameworks for distinguishing between permissible re-use and unlawful reproduction—a distinction that becomes particularly strained in the context of AI-generated content.

The United States remains the leading jurisdiction in developing the doctrine of transformative use within copyright law, especially under the fair use exception in section 107 of the Copyright Act of 1976. A significant judicial development occurred in Andy Warhol Foundation for the Visual Arts, Inc. v Goldsmith, where the United States Supreme Court held that the unauthorised commercial licensing of a Warhol silkscreen image, based on a photograph by Lynn Goldsmith, did not constitute fair use. 17 The Court clarified that while Warhol’s image added some new expression to the original photograph, it did not sufficiently alter the “meaning or message” in a way that justified the reproduction of the original work, especially when the secondary use was for a similar commercial purpose.

18 This decision reined in prior judicial enthusiasm for finding transformative use and emphasised the need for a meaningful alteration of expressive content beyond superficial changes or restyling. The reasoning in Warhol bears relevance for the treatment of generative systems, which often produce outputs that are stylistically or compositionally similar to the input data but do not directly reproduce protected works in their entirety. For instance, when an image generator trained on a corpus of thousands of artworks produces a new work that visually resembles a particular artist’s style, the question arises whether such resemblance constitutes the creation of a derivative work or represents a transformative reimagining. Under US law, if the output fails to alter the underlying expression or purpose of the original in a meaningful way, it may

17 Andy Warhol Foundation for the Visual Arts, Inc v Goldsmith 143 S Ct 1258 (2023).

18 ibid [28]–[30].

not qualify as transformative, and its unauthorised use could be deemed infringing. 19 This threshold sets a relatively high bar for AI-generated content that draws from protected materials without clearly departing from the original’s aesthetic or conceptual essence. In contrast, the United Kingdom adopts a more restrictive approach to unauthorised reuse of protected works. The UK Copyright, Designs and Patents Act 1988 does not provide a general fair use defence; rather, it offers narrowly defined exceptions, such as fair dealing for criticism, review, or parody under sections 29 and 30. The leading case of Ashdown v Telegraph Group Ltd affirmed that even non-commercial re-use of copyrighted material may infringe if it appropriates a substantial part of the original work, regardless of the purpose.

20 UK courts have tended to focus on whether a substantial part of the work—assessed both quantitatively and qualitatively—has been copied, rather than on the broader context of whether the new work adds a new expression or message. This approach offers limited scope for claiming legality in AI-generated outputs that closely mimic or replicate existing artistic styles, particularly where the original work’s distinctive elements are reproduced or emulated. Australian law similarly follows a conservative path, with no explicit recognition of transformative use and a strict interpretation of exceptions.

In IceTV Pty Ltd v Nine Network Australia Pty Ltd, the High Court of Australia held that originality resides in the skill and labour expended in producing a work, and infringement depends upon copying a substantial part of the original, as judged by the quality and significance of the copied material. 21 This reinforces a legal environment in which AI-generated works that emulate structural or stylistic features may be scrutinised for infringing protected elements, even in the absence of direct reproduction. As in Aotearoa New Zealand, the absence of a transformative use doctrine restricts the scope for judicial discretion in determining whether AI-generated material that departs aesthetically from its source is sufficiently original or remains legally derivative.

In Aotearoa New Zealand, courts have not yet been required to adjudicate cases involving AI-generated content in the context of derivative or transformative use. However, the analytical principles applicable to derivative works, as discussed in Bonz Group (Pty) Ltd v Cooke, suggest a preference for assessing whether the new work incorporates identifiable and 19 Rebecca Tushnet, ‘Worth a Thousand Words: The Images of Copyright’ (2023) 71(1) Journal of the Copyright Society 1, 19–21.

20 Ashdown v Telegraph Group Ltd [2001] EWCA Civ 1142, [2002] Ch 149.

21 IceTV Pty Ltd v Nine Network Australia Pty Ltd [2009] HCA 14; (2009) 239 CLR 458.

protectable elements of the original. 22 Without a statutory fair use defence or judicial precedent endorsing the concept of transformation as a justification, the Aotearoa New Zealand legal framework lacks the flexibility present in other jurisdictions to accommodate nuanced cases involving recontextualisation, parody, or machine remixing. In sum, the comparative landscape reveals a divergence in how jurisdictions conceptualise and regulate derivative versus transformative uses. Jurisdictions with codified fair use or broader judicial recognition of transformation have begun to grapple with AImediated creativity, albeit within a contested and evolving doctrinal space.

Aotearoa New Zealand’s more rigid statutory architecture leaves little room for interpreting transformation as a defence or exception to infringement, thereby exposing creators, developers, and users of generative technologies to heightened legal risk. The implications of this position are increasingly salient as domestic use of such systems expands, and global litigation shapes the expectations of rights-holders and content platforms alike. Unlike jurisdictions that have begun articulating either statutory exceptions or judicial reasoning to accommodate transformative use in AI contexts, Aotearoa New Zealand remains without any legislative recognition or interpretative framework to distinguish permissible re-use from infringement.

Attribution of Civil Liability for Copyright Infringement in Generative Systems The increasing use of generative technologies in content creation has introduced considerable complexity in the attribution of civil liability for copyright infringement. Under Aotearoa New Zealand law, liability traditionally attaches to the party who exercises the exclusive rights reserved to the copyright owner without proper authorisation, as set out in section 29 of the Copyright Act 1994. 23 The primary infringer is generally the person who commits the infringing act, such as reproducing or communicating the work to the public. However, in the case of outputs created through generative systems, where the infringement may occur indirectly or unintentionally, the question of who should bear responsibility—user, developer, or platform provider—remains legally unsettled.

Where an individual uses a generative platform to produce content that closely resembles or replicates a protected work, civil liability may arise even in the absence of an

22 Bonz Group (Pty) Ltd v Cooke [1994] 3 NZLR 216 (HC).

23 Copyright Act 1994 (NZ), s 29.

intent to infringe. The law does not require a subjective element of knowledge or intention for a finding of primary infringement. 24 Thus, where the output is substantially similar to an existing work and that similarity results from the reproduction of a protected expression, the user may be liable even if unaware that the generated content infringes copyright. This is consistent with the broader principles of strict liability for primary infringement, which focus on the act rather than the mental state of the infringer. However, the attribution of liability becomes more contentious where the user lacks meaningful control over the content generation process, or where the infringing character of the output is the product of design choices embedded in the model’s training and architecture.

In such cases, the platform developer or service provider may be implicated under doctrines of secondary or joint liability. Although Aotearoa New Zealand case law in this area is limited, the courts have recognised joint tortfeasorship where two or more parties act pursuant to a common design or where one party facilitates the tortious act of another with knowledge. 25 In the context of generative systems, a plausible case for joint liability may arise if a platform knowingly provides tools capable of producing infringing material and fails to implement effective safeguards or moderation mechanisms. The requisite knowledge threshold remains uncertain, particularly in environments where model outputs are not deterministic and the provider lacks direct oversight of specific user interactions.

Further complexity arises in relation to the training of generative models using datasets that include copyright-protected material. If the developer has incorporated such material without obtaining licences or relying on a valid exception, the model itself may be regarded as a tool built on infringement. Although Aotearoa New Zealand does not currently impose data transparency requirements on AI developers, the absence of such obligations does not exempt them from potential liability if the outputs can be causally linked to infringing training data. In the United States, claims have been filed against developers whose models are alleged to have generated outputs substantially similar to protected works used during training.

26 Should similar litigation arise in Aotearoa New Zealand, courts would need to evaluate the extent to which the developer’s conduct constituted authorisation or facilitation of infringement.

24 Bauman v Fussell [1978] 1 NZLR 327 (CA); APRA v Coolgrove Pty Ltd [1995] 57 FCR 89 (Fed Ct, Aus).

25 Fish & Fish Ltd v Sea Shepherd UK [2015] UKSC 10, [2015] 2 All ER 613.

26 Doe 1 v GitHub, Inc. (Case No. 3:22-cv-06823, ND Cal, filed 3 Nov 2022); Sarah Silverman et al v OpenAI, Inc. (Case No. 3:23-cv-03416, ND Cal, filed 7 July 2023).

Platform liability also intersects with broader questions of intermediary responsibility. Under section 92B of the Copyright Act 1994, internet service providers are afforded limited safe harbours if they act as passive conduits and respond appropriately to takedown notices. 27 However, the statutory safe harbour regime does not clearly extend to operators of generative platforms who actively develop and maintain proprietary models capable of producing userdirected outputs. In the absence of a statutory classification for such services, the applicability of existing safe harbours remains ambiguous, further exposing developers to potential legal action from rights-holders. The cumulative effect of these legal uncertainties is a fragmented liability framework that fails to provide clear guidance on responsibility for infringement involving generative technologies.

Users face potential liability despite their limited understanding of the underlying model behaviour; developers operate in a legal grey area in the absence of clear standards on training data and output moderation; and platform operators may be subject to liability under joint tortfeasorship or deemed to authorise infringement by omission. As the next section will explore, this fragmentation is compounded by evidentiary difficulties and structural gaps in enforcement, which further complicate the resolution of such disputes under Aotearoa New Zealand’s current legal regime. Evidentiary Complexities and Procedural Constraints in Copyright Enforcement While the conceptual attribution of liability for AI-mediated infringement raises substantive legal questions, the practical enforcement of such claims presents equally significant procedural obstacles.

Foremost among these are the evidentiary burdens borne by claimants, who must demonstrate access to the protected work and substantial similarity between the allegedly infringing output and the original expression. These requirements, relatively straightforward in conventional copyright disputes, become exceptionally challenging in the context of generative technologies, where the process of content creation is often opaque, decentralised, and lacking in documentary traceability. In traditional copyright litigation, courts rely heavily on the factual matrix surrounding the act of copying. This typically includes evidence of access, temporal proximity, and visual or textual comparison of the original and infringing works. In Bonz Group (Pty) Ltd v Cooke,

27 Copyright Act 1994 (NZ), ss 92A–92E.

the High Court reaffirmed that the claimant must prove that a substantial part of the work was taken and that this part involved protectable original expression. 28 However, the generative process used in machine learning models—especially large-scale, unsupervised training methods—lacks the linear causality usually present in human reproduction. The models themselves do not maintain logs that can be easily queried to establish which specific training data informed a particular output. Without this link, rights-holders face substantial evidential deficiencies when seeking to prove that their works were used or that any similarities are the result of reproduction rather than mere stylistic convergence. Compounding these difficulties is the issue of access. For a finding of infringement, it is typically necessary to show that the infringing party had access to the original work.

29 In the case of generative systems, the access is mediated through training data that may have been collected by the developer and is not disclosed to users or, in many cases, to regulators. For instance, models such as Stable Diffusion and GPT variants have been trained on web-scraped datasets, which may contain copyrighted content obtained without licences. 30 Yet, in the absence of transparency requirements or public documentation of such training corpora in Aotearoa New Zealand, proving access becomes speculative. The claimant cannot easily demonstrate whether a specific image, text, or recording formed part of the model’s training data or whether its presence in an output resulted from that exposure. Further evidentiary limitations arise from the inability of current forensic methods to reverse-engineer model outputs to their constituent data.

Generative systems, by design, do not replicate inputs in a pixel-for-pixel or word-for-word manner. Rather, they operate through statistical representations, often referred to as embeddings, that abstract the structure and features of input data without preserving identifiable fragments. As a result, an output may be recognisably similar to a protected work without containing any segment that would ordinarily be construed as a “copy” under legal standards. Some legal scholars have proposed the use of technical watermarking, cryptographic provenance tracing, or model auditing protocols to improve attribution and enable courts to link outputs back to infringing inputs. 31 This structural

28 Bonz Group (Pty) Ltd v Cooke [1994] 3 NZLR 216 (HC).

29 Sawkins v Hyperion Records Ltd [2005] EWCA Civ 565, [2005] 1 WLR 3281 [86]–[87].

30 Joe Mullin, ‘AI Training Data Dispute Could Decide Future of Machine Learning’ (Electronic Frontier Foundation, 11 September 2023) https://www.eff.org/deeplinks/2023/09/ai-training-data-dispute-could-decidefuture-machine-learning accessed 3 June 2025. 31 Ramalho A, ‘Technology to the Rescue? Tracing Copyright Data in the Age of Generative AI’ (2023) Journal of Intellectual Property Law & Practice.

challenge weakens the evidentiary position of claimants, who may suspect unauthorised use of their works but lack admissible proof to support such claims in civil proceedings. Moreover, unlike direct copying, the outputs of generative systems may involve recombined elements or stylistic approximations, making it difficult to apply the traditional test of substantial similarity in a meaningful way. The difficulty of identifying responsible parties also undermines enforcement efforts. As noted in the earlier section on liability, the potential infringers in generative workflows may include users, developers, or hosting platforms. However, in practice, claimants may struggle to identify a specific party against whom to initiate proceedings, particularly when services are operated from offshore jurisdictions or under complex corporate arrangements.

The absence of local representatives or compliance mechanisms further restricts the ability to serve legal notices or obtain preliminary injunctive relief. Moreover, the Aotearoa New Zealand legal system does not currently offer a tailored process for dealing with automated or semiautomated infringement, nor does it possess a copyright small claims tribunal capable of handling low-value but procedurally complex disputes, unlike jurisdictions such as the United Kingdom. 32 Finally, rights-holders face practical disincentives to pursuing litigation given the cost, uncertainty, and public scrutiny associated with such cases. Legal action against developers of widely used generative systems may require expert testimony on machine learning architectures, evidentiary discovery from reluctant defendants, and judicial interpretation of technical evidence.

These hurdles, coupled with the present doctrinal ambiguities in Aotearoa New Zealand law regarding machine-generated content, create a climate of enforcement fatigue among creators. The net result is a disjunction between the theoretical availability of remedies under the Copyright Act 1994 and the practical ability to secure those remedies in cases involving AI-generated works. This evidentiary opacity is exacerbated by the absence of any legal obligation in Aotearoa New Zealand for AI model developers to retain logs of training datasets or outputs, in contrast to proposed European legislation that mandates data documentation in high-risk systems. 33

32 Intellectual Property Office (UK), Small Claims Track for IP Disputes: Guidance Note (2020)

https://www.gov.uk/guidance/small-claims-track-for-ip-disputes accessed 3 June 2025. 33 European Commission, Proposal for a Regulation Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) COM (2021) 206 final, art 52.

Legislative Inertia and Comparative Reform Initiatives in the Regulation of Generative Technologies Despite the legal complexities and risks posed by generative technologies in the field of copyright, Aotearoa New Zealand’s policy response has remained largely dormant. The Copyright Act 1994, last substantively amended in 2008, does not reflect the realities of automated content production, nor does it engage with the challenges posed by AI-based tools to traditional conceptions of originality, authorship, or reproduction. No formal review has been initiated to assess the compatibility of current legislation with the emergence of generative machine learning systems.

This legislative inertia stands in stark contrast to developments in comparative jurisdictions that have begun to consider, and in some instances implement, regulatory frameworks tailored to address the disruptive effects of generative AI. In the United Kingdom, the Intellectual Property Office conducted a public consultation in 2021 regarding the implications of AI for copyright and related rights, culminating in a policy decision to preserve the current scope of protection for computer-generated works while deferring the question of authorship reform. 34 Although this response was measured, it acknowledged the policy tensions between fostering innovation and ensuring fair remuneration for human creators.

Of particular note is section 9(3) of the Copyright, Designs and Patents Act 1988, which deems the person making the “arrangements necessary” for the creation of a computer-generated work to be the author. 35 This provision, absent from Aotearoa New Zealand’s legislation, at least offers a statutory foothold for assigning authorship in contexts where no individual directly crafts the expressive output. However, the UK Government ultimately declined to extend copyright protection to AI-generated works absent human input, reaffirming the centrality of human creativity in its copyright doctrine. 36 The European Union has adopted a more expansive approach through its ongoing legislative work on the Artificial Intelligence Act, which includes obligations for transparency and accountability in high-risk AI systems.

Although the proposal does not amend the EU’s copyright framework directly, it introduces documentation duties that may aid in identifying 34 UK Intellectual Property Office, Artificial Intelligence and Intellectual Property: Government Response to the Call for Views (2022) https://www. gov. uk/government/publications/artificial-intelligence-and-ip-call-for-viewsgovernment-response accessed 3 June 2025.

35 Copyright, Designs and Patents Act 1988, s 9(3).

36 UKIPO (n 1) 16–18.

whether protected content was used in training data.37 Furthermore, the European Parliament has called for the development of a legal status for AI-generated content and the clarification of liability rules under the Product Liability Directive.38 These developments suggest a regulatory trajectory that, while cautious, acknowledges the need to align IP law with the risks introduced by generative AI. In the United States, while there has been no federal legislation to date specifically addressing generative AI and copyright, the United States Copyright Office issued guidance in

2023 indicating that works created solely by AI without human authorship are not eligible for

protection. 39 This clarification followed high-profile litigation over the use of copyrighted material in the training of large language and image models. The Office’s position underscores a broader concern in American policy that expanding copyright protection to non-human entities may dilute the incentive structure that underpins the creative economy. At the same time, pending litigation in the US courts—including claims involving alleged unauthorised use of photographs, artwork, and code—may lead to judicially driven policy developments in the absence of legislative intervention. Aotearoa New Zealand’s failure to engage in any comparable policy process reveals a serious lacuna in its legislative approach.

The Ministry of Business, Innovation and Employment (MBIE), which oversees intellectual property policy, has not conducted any targeted review or consultation on the implications of AI for copyright law, despite international developments and growing domestic use of generative platforms. The Copyright Act contains no provisions requiring developers to document or disclose the sources of training data, nor does it offer any specific defence or liability regime applicable to AI-mediated production. Notably, the Ministry of Business, Innovation and Employment (MBIE), which oversees IP and copyright legislation, has omitted AI-related risks from its most recent reviews on copyright reform, including the Copyright Act Modernisation Programme.

40 The absence of a statutory position on the copyrightability of AI-generated content, as well as the liability 37 European Commission, Proposal for a Regulation Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) COM (2021) 206 final, art 52. 38 European Parliament Resolution of 20 October 2020 with recommendations to the Commission on a civil liability regime for artificial intelligence (2020/2014(INL)). 39 US Copyright Office, Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence (2023) https://www. copyright. gov/ai/ accessed 3 June 2025.

40 Ministry of Business, Innovation and Employment (MBIE), Copyright Act Review Programme Overview

(2023) https://www.mbie.govt.nz/business-and-employment/business/intellectual-property/copyright/copyrightact-review/ accessed 3 June 2025.

of those who deploy or profit from such content, creates legal ambiguity that could undermine enforcement and deter innovation simultaneously. In view of the reform momentum elsewhere, Aotearoa New Zealand’s current posture reflects a lack of regulatory foresight. Without substantive engagement on the questions of training data, authorship, originality, and liability, Aotearoa New Zealand risks becoming a jurisdictionally underdeveloped territory in an increasingly internationalised field. This regulatory passivity places both creators and AI developers in a position of uncertainty, impeding their ability to assess legal risk or secure rights with confidence. As the following section will argue, a principled reform agenda is required to bring clarity to this area and to align Aotearoa New Zealand’s copyright framework with technological, commercial, and doctrinal realities.

Recommendations for Reform To address the doctrinal, evidentiary, and policy limitations outlined above, the following legislative and regulatory reforms are proposed: In light of the doctrinal ambiguity and policy inertia surrounding the treatment of generative technologies under Aotearoa New Zealand copyright law, a structured legislative reform programme is both necessary and overdue. The Copyright Act 1994, while adequate in many traditional contexts, is ill-equipped to regulate works produced or materially altered by machine learning systems, particularly where authorship is diffuse or derivative elements are embedded in probabilistically generated outputs. Addressing these shortcomings will require both targeted statutory amendments and broader regulatory engagement to ensure legal certainty for rights-holders, technology developers, and end users.

The first and most pressing area of reform concerns the definition and scope of authorship. At present, the Act implicitly assumes that all copyrightable works originate from human creativity, as reflected in its reliance on the notion of the “author” as a natural person. 41 While this approach aligns with the foundational philosophy of copyright as a reward for intellectual labour, it does not reflect the practical realities of content generated by AI systems. Following the example of the United Kingdom’s section 9(3) of the Copyright, Designs and Patents Act 1988, Aotearoa New Zealand could introduce a statutory provision that attributes

41 Copyright Act 1994 (NZ), s 5(1).

authorship in computer-generated works to the person who undertakes the necessary arrangements for creation. 42 This would provide a legally recognised mechanism for allocating rights and responsibilities in relation to AI-generated works, while maintaining the humancentric structure of copyright law. Secondly, the Act should be amended to clarify the threshold of originality as it applies to AI-assisted works. While Aotearoa New Zealand courts have traditionally followed the qualitative approach to originality articulated in University of London Press Ltd v University Tutorial Press Ltd, technological developments necessitate a more nuanced understanding of what constitutes original expression. 43 Specifically, legislation should draw a distinction between works that are wholly machine-generated and those in which human intervention meaningfully shapes the final output.

Such a distinction would allow courts to exclude purely autonomous outputs from copyright protection, while recognising derivative works as potentially protectable where a human contributor exercises sufficient creative judgment over form or content. In tandem with these changes, the Act should include a specific statutory exception or defence relating to unintentional reproduction by generative systems. Current infringement provisions impose liability irrespective of intent, thereby exposing users and developers to potential claims for outputs that they could not reasonably foresee or control. 44 While preserving the integrity of copyright enforcement, a limited exception could be introduced for situations where the reproduction is incidental, technically unavoidable, or not commercially exploited.

This approach would mirror the rationale behind existing exceptions for transient copying and could provide a proportionate balance between technological innovation and rights protection. Moreover, Aotearoa New Zealand must address the evidentiary and transparency challenges that hinder effective enforcement. At a minimum, legislative reform should require developers of large-scale generative models to maintain records of training data and to disclose, upon request by affected rights-holders, whether a particular work or dataset formed part of their model’s training corpus. Such a requirement would align with the European Union’s approach under the proposed Artificial Intelligence Act and would enable claimants to establish

42 Copyright, Designs and Patents Act 1988, s 9(3).

43 University of London Press Ltd v University Tutorial Press Ltd [1916] 2 Ch 601 (Ch).

44 Copyright Act 1994 (NZ), s 29; see also APRA v Coolgrove Pty Ltd [1995] 57 FCR 89 (Fed Ct, Aus).

access in copyright litigation with greater confidence. 45 To ensure compliance, these obligations could be incorporated into Aotearoa New Zealand’s consumer protection or privacy regulatory frameworks, particularly where generative tools are offered as commercial services to the public. Additionally, a regulatory mechanism should be established to manage disputes involving low-value or high-volume infringement claims. The introduction of a Copyright Tribunal with jurisdiction over AI-related disputes, modelled on the Intellectual Property Enterprise Court in the United Kingdom, could provide an accessible and cost-effective forum for resolving claims involving generative content.

46 Such a tribunal could develop expertise in the technical and legal nuances of machine-generated works, thereby relieving the burden on general civil courts and improving the consistency of outcomes in this novel area of law. Finally, reform must be guided by public consultation and interdisciplinary engagement. The Ministry of Business, Innovation and Employment (MBIE), in partnership with legal, technical, and creative sector stakeholders, should initiate a comprehensive review of the Copyright Act with respect to emerging technologies. As other jurisdictions have demonstrated, proactive policymaking in this domain is not only desirable but essential to preserving the relevance and legitimacy of copyright in the digital age.

These proposed reforms are not exhaustive, but they represent a starting point for aligning Aotearoa New Zealand’s legal framework with the realities of generative content production. Absent such reform, the Copyright Act will continue to operate in a vacuum, undermining both the enforceability of rights and the viability of innovation in the emerging digital economy. Conclusion The intersection of copyright law and generative artificial intelligence presents one of the most urgent and underdeveloped areas of legal doctrine in Aotearoa New Zealand.

As AI systems increasingly produce outputs that resemble or derive from protected works, the existing statutory framework—centred on human authorship, originality, and intentional 45 European Commission, Proposal for a Regulation Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) COM (2021) 206 final, art 52.

46 Intellectual Property Office (UK), Small Claims Track for IP Disputes: Guidance Note (2020)

https://www.gov.uk/guidance/small-claims-track-for-ip-disputes accessed 3 June 2025.

reproduction—proves inadequate. The Copyright Act 1994 does not contemplate the realities of content generated by autonomous systems, nor does it provide clarity on the legal position of those who prompt, develop, or deploy such technologies. This article has shown that the absence of legislative definitions for AI-generated authorship, the lack of transparency obligations for model training, and the evidentiary opacity in infringement disputes have created a fragmented and uncertain legal environment. These shortcomings not only burden creators and developers with legal ambiguity but also inhibit enforcement, compliance, and the sustainable integration of AI tools into creative industries.

Moreover, the failure to adapt places New Zealand behind jurisdictions such as the United Kingdom, the European Union, and the United States, which have either introduced targeted reforms or initiated structured policy consultations in response to similar challenges. To address these deficiencies, this research recommends a principled package of reforms, including statutory clarification of authorship for AI-generated works, revised originality standards for human–machine collaboration, safe harbour provisions for incidental infringement, mandated disclosure of training data provenance, and the creation of a specialised adjudicative forum for AI-related copyright disputes. These proposals are grounded in comparative experience but tailored to New Zealand’s legal and institutional context.

In the absence of meaningful intervention, the law will continue to operate reactively and inconsistently, undermining both its normative legitimacy and practical enforceability. A forward-looking, evidence-based reform agenda is essential—not only to ensure that copyright remains adaptable and just in the face of technological disruption, but also to uphold New Zealand’s standing in the evolving landscape of international intellectual property governance.

References ● Andy Warhol Foundation for the Visual Arts, Inc.

v Goldsmith 143 S Ct 1258 (2023) ● APRA v Coolgrove Pty Ltd [1995] 57 FCR 89 (Fed Ct, Aus) ● Ashdown v Telegraph Group Ltd [2001] EWCA Civ 1142, [2002] Ch 149 ● Bauman v Fussell [1978] 1 NZLR 327 (CA) ● Bonz Group (Pty) Ltd v Cooke [1994] 3 NZLR 216 (HC) ● Copyright Act 1994 (NZ) ● Copyright, Designs and Patents Act 1988 (UK) ● Doe 1 v GitHub, Inc (Case No 3:22-cv-06823, ND Cal, filed 3 November 2022) ● European Commission, Proposal for a Regulation Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) COM (2021) 206 final ● European Parliament, Resolution with Recommendations to the Commission on a Civil Liability Regime for Artificial Intelligence (2020/2014(INL)) ● Fish & Fish Ltd v Sea Shepherd UK [2015] UKSC 10, [2015] 2 All ER 613 ● Guadamuz A, ‘Artificial Intelligence and Copyright’ (2017) 39(1) European Intellectual Property Review 18 ● Hawkes & Sons (London) Ltd v Paramount Film Service Ltd [1934] 1 Ch 593 (CA) ● IceTV Pty Ltd v Nine Network Australia Pty Ltd [2009] HCA 14; (2009) 239 CLR 458 ● Intellectual Property Office (UK), Small Claims Track for IP Disputes: Guidance Note (2020) https://www.

gov. uk/guidance/small-claims-track-for-ip-disputes ● Kaminski ME, ‘The Right to Explain AI Decisions’ (2019) 68(1) UCLA Law Review

● Ministry of Business, Innovation and Employment (MBIE), Copyright Act Review Programme Overview (2023) https://www.mbie.govt.nz/business-andemployment/business/intellectual-property/copyright/copyright-act-review/ ● Mullin J, ‘AI Training Data Dispute Could Decide Future of Machine Learning’ Electronic Frontier Foundation (11 September 2023) https://www.eff.org/deeplinks/2023/09/ai-training-data-dispute-could-decide-futuremachine-learning ● Ramalho A, ‘Technology to the Rescue? Tracing Copyright Data in the Age of Generative AI’ (2023) Journal of Intellectual Property Law & Practice ● Sawkins v Hyperion Records Ltd [2005] EWCA Civ 565, [2005] 1 WLR 3281 ● Screen Production and Development Association (SPADA), Annual Report 2023: Aotearoa’s Screen Sector in Transition

(SPADA,

2023) https://www.spada.co.nz/resources ● Tushnet R, ‘Worth a Thousand Words: The Images of Copyright’ (2023) 71(1) Journal of the Copyright Society 1 ● UK Intellectual Property Office, Artificial Intelligence and Intellectual Property: Government Response to the Call for Views (2022) https://www.gov.uk/government/publications/artificial-intelligence-and-ip-call-forviews-government-response ● United States Copyright Act 1976, 17 USC §107 ● United States Copyright Office, Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence (2023) https://www.copyright.gov/ai/ ● University of London Press Ltd v University Tutorial Press Ltd [1916] 2 Ch 601 (Ch)