Stealing The Style - Conflicts Of Copyright Law In Algorithmic Art

Author: LawSphereX

Category: Intellectual Property

Published: 2026-08-23T21:09:14.519026+00:00

Updated: 2026-08-26T17:06:12.161+00:00

STEALING THE STYLE: CONFLICTS OF COPYRIGHT LAW

IN ALGORITHMIC ART

Gauri Nandakumar Varma, O.P. Jindal Global Law School

ABSTRACT

Generative artificial intelligence (AI) has presented a new challenge to copyright law by revealing a doctrinal flaw in the treatment of artistic style. Traditionally, copyright did not protect the idea itself and only the expression of an, a doctrine that was formulated in Baker v Selden (1879) and affirmed in the Indian case of RG Anand v Deluxe Films (1978). An artist's style, the patterns of aesthetic choices and composition that characterise a work of art, has been excluded as an unprotectable idea. This was deemed important to avoid monopolies of methods of production and avoid locking up the creative commons. This proposition has been upset by generative AI. Models trained on billions of images can now mimic styles at scale and produce works that are indistinguishable from a particular artist's style, but which do not copy any specific protected work.

Case law in Andersen v Stability AI (2023ongoing) and Getty Images v Stability AI (UK, 2025) highlights the pressing need and existing doctrinal shortcomings. The idea-expression distinction, the doctrine of works derived from other works, the fair use defence and the doctrine of secondary infringement have failed to resolve the problem of the harms to human creators from AI-generated style mimicry. This paper shows that while the doctrine of stylistic exclusion is theoretically valid, its application to generative AI requires fresh consideration.

Drawing on case law, the fair use and transformativeness doctrines, the passing off and moral rights doctrines, and the European Union's text and data mining regime, it offers a fair balance through preserving the public domain, while creating specific legal tools, such as transparency, meaningful opt-out, and a brief sui generis protection, to prevent the unconsented commercial use of identifiable styles through generative AI.

I.

Introduction Generative AI and copyright law are one of the pressing jurisprudential issues of the 21st century. Software like Stable Diffusion and Midjourney enables anyone to type a prompt for a model to generate an image "in the style of" a living artist, mimicking their style with an accuracy that would formerly have taken years of practise for an artist to master. This presents us with an awkward reckoning with a core copyright doctrine: style is an idea and ideas are common property. The idea-expression dichotomy, established in Baker v Selden1 (1879) and reinforced in RG Anand v M/S Deluxe Films2 (1978), means that copyright covers the expression of an idea, not the idea itself.

Artistic style has long been recognised as an idea, and ideas are not protected by copyright: to do so would grant a monopoly over the manner of creation and hinder the art world's conversation with itself. This paper does not claim that the doctrine is incorrect. It does argue that generative AI is making its application problematic. When a diffusion model is trained on billions of artworks and used to sell their aesthetic expressions, the normative underpinnings of the ideaexpression rule, that ideas are free because humans can always choose how to express them, are undermined. By examining key cases, doctrines, and regulatory models, this paper offers legal reforms commensurate with that threat. II.

The Idea-Expression Dichotomy and the Merger Doctrine The idea-expression dichotomy is enshrined in section 102(b) of the US Copyright Act of 19763, wherein only expressions are protected not ideas4. In Baker v Selden, Bradley J found Selden's written description of his bookkeeping system to be copyrightable, but that

the system as a method, could not be monopolised: “[w]here the truths of a science or the methods of an art are the common property of the whole world, any author has the right to express the one, or explain and use the other, in his own way”5. The case also established the merger doctrine wherein an idea can only be expressed in one or limited ways, the expression and idea are merged together, and neither attracts copyright protection because

1 Baker v Selden 101 US 99 (1879).

2 RG Anand v Deluxe Films AIR 1978 SC 1613.

3 Copyright Act 1976 (US), 17 USC § 102(b).

4 Copyright Act 1976 (US), 17 USC §§102(b), 106(2), 107.

5 (n 1).

to do so would amount to monopolising the idea itself. These doctrines have been confirmed by Indian law in RG Anand v M/S Deluxe Films6, where the Supreme Court of India confirmed that ideas, themes and plots are part of the public domain and cannot be protected by copyright infringement. The Court articulated the doctrine of substantial similarity via its "lay observer test": infringement is only present where the viewer is left with an "unmistakable impression" that one work is a copy of the other, requiring similarities to be material and substantial, rather than incidental7. Both cases are founded on a common normative point: monopolising methods of creation slows innovation, which copyright aims to promote. III.

Derivative Works and Artistic Style Copyright owners have the exclusive right to make derivative works based on their originals under Section 106(2) of the Copyright Act of 19768. The plaintiffs in Andersen v Stability AI invoked this doctrine to argue that the "in the style of" AI outputs created by the defendants are infringing derivative works created from the protected expression within the training data. This claim faces a head-on doctrinal challenge: derivative works protect against copying of protected expression, not imitation of unprotected stylistic ideas. As Pam Samuelson has argued, generative AI is uniquely positioned to produce works that closely resemble stylistic features of human authors, but do not literally copy them, a new type of harm that the derivative works doctrine, as currently interpreted, does not address9.

Jane Ginsburg and Luke Budiardjo also point out that authorship norms developed for human creation are problematic for large-scale algorithmic creation10. IV.

The Litigation Landscape

A. Andersen v Stability AI Ltd (2023–Ongoing)

The litigation of Andersen v Stability AI 11 commenced on 13 January 2023 in the Northern District of California, and is the largest copyright case brought against AI image

6 (n 2).

7 ibid.

8 Copyright Act 1976 (US), 17 USC § 106(2).

9 Pamela Samuelson, 'Generative AI Meets Copyright' (2023) 381 Science 158.

10 Jane C Ginsburg and Luke A Budiardjo, 'Authors and Machines' (2019) 34 Berkeley Technology Law Journal 343.

11 Andersen v Stability AI Ltd No 3:23-cv-00201 (ND Cal 2023–ongoing).

generators12. The case was brought by three artists as a class action against Stability AI, Midjourney, and DeviantArt for alleged unlicensed use of their copyrighted works (from the five billion-image LAION-5B data set) to train Stable Diffusion. It included claims for direct infringement and creation of infringing derivative works via "in the style of" AIgenerated artworks13. Judge Orrick dismissed most motions to dismiss on 30 October 2023, but allowed the direct infringement claim against Stability AI to proceed, finding it reasonable that all of Andersen's registered works were in the training data14. On 12 August 2024, the court permitted direct infringement, trademark and inducement claims and dismissed DMCA copyright management information claims. The trial will take place in September 2026.

The doctrinal problem here, as with the case above, is that the plaintiffs' complaint is not that any particular image was copied; rather, their aesthetic styles and years of work were commercially exploited without permission or compensation at an industrial scale. B. Getty Images v Stability AI (UK, 2025)

In January 2023, Getty Images brought an action in the High Court of England and Wales, claiming primary and secondary copyright infringement, database right infringement, trademark infringement, and passing off15. Mrs Justice Joanna Smith DBE's judgment of 4 November 2025 is the first substantive IP determination on generative AI in the UK16. The outcome was narrow. Getty withdrew its foremost infringement claims because it did not establish training in the UK, a prerequisite for UK copyright jurisdiction. The secondary claim for copyright infringement under sections of the Copyright, Designs and Patents Act 198817 was also unsuccessful: the Court found that while Stable Diffusion may be an intangible "article", its weights did not contain copies of Getty's images and could

12 ibid.

13 Andersen v Stability AI: The Landmark Case Unpacking the Copyright Risks of AI Image Generators’ (NYU Journal of Intellectual Property and Entertainment Law, 2024) https://jipel.law.nyu.edu/andersen-v-stability-aithe-landmarkcase-unpacking-the-copyright-risks-of-ai-image-generators/ accessed 28 April 2026.

14 Andersen v Stability AI Ltd, Order on Motions to Dismiss (ND Cal, 30 October 2023).

15 Getty Images (US) Inc & Ors v Stability AI Limited [2025] EWHC 2863 (Ch).

16 ibid [87–108] (Mrs Justice Joanna Smith DBE).

17 Copyright, Designs and Patents Act 1988 (UK), ss 22, 23, 27, 77–80.

not therefore be an "infringing copy"18. Some limited trademark infringement was found in the reproduction of watermarks. As Katten Muchin Rosenman noted, the case 'leaves critical questions unanswered', particularly the question of whether training on copyrighted works at scale constitutes infringement19. V.

Fair Use, Transformativeness, and the Market Harm Problem The main defence used by AI companies for their training has been the fair use doctrine commissioned under Section 107, Copyright Act 197620. It is necessary to take into account four factors: the nature of the copyrighted work, the fraction exploited, the market effect, and the purpose and character of the usage. The key element in the AI defence is the doctrine of transformativeness, expounded in seminal work by Pierre Leval: If an AI application adds something new and gives the original work a new function or personality, it is transformational21.

In Authors Guild v Google and Authors Guild v HathiTrust22, the Second Circuit found mass digitisation of copyright works for search and analysis a transformative fair use; the resulting corpus was used for purposes, search, research, that were distinct from the expressive purposes of the originals. 23 In mid-2025, a US District Court described Anthropic's AI training as "spectacularly transformative", comparing it to a person reading books24. But this expansive understanding of transformativeness is severely tested by the commercial sale of AI art "in the style of" artists. In Andy Warhol Foundation v Goldsmith25, the Supreme Court stressed that commercial use for the same intent as the original strongly militates against fair use26. The marketing of an AI system that produces unlimited output in the distinctive style of a living artist clearly displaces the artist's commissions.

Matthew Sag suggests that fair use should apply to "fair learning" so long as 18 Berne Convention for the Protection of Literary and Artistic Works (adopted 9 September 1886, revised 28 September 1979), art 6bis. 19 Katten Muchin Rosenman LLP, 'Getty Images v Stability AI: IP Rights in the Age of Generative AI' (November

2025) https://katten.com accessed 28 April 2026.

20 Copyright Act 1976 (US), 17 USC § 107.

21 Pierre N Leval, 'Toward a Fair Use Standard' (1990) 103 Harvard Law Review 1105.

22 Authors Guild v Google Inc 804 F 3d 202 (2d Cir 2015).; Authors Guild Inc v HathiTrust 755 F 3d 87 (2d Cir 2014).

23 ibid.

24 Irwin IP LLC, 'A Black Flag for Authors: AI Training Ruled Fair Use' (July 2025) https://irwinip.com accessed

28 April 2026.

25 Andy Warhol Foundation for the Visual Arts Inc v Goldsmith 598 US 508 (2023).

26 ibid.

the products of AI systems do not reproduce individual originals,27 while Lemley and Casey conclude that "the most important variable is not the nature of the input but the nature of the output"28. Both observations point to a conclusion that the courts are yet to fully accept: if the output systematically displaces the market for the creator's output at scale, then the transformative nature of the training cannot, by itself, form a full defence. VI.

The European Framework: TDM and the AI Act The European Union has advanced furthest on legislating AI-copyright. Article 4 of the Copyright in the Digital Single Market Directive (DSM Directive, 2019/790) establishes a text and data mining (TDM) exception for commercial AI training on works lawfully made available, except where rightsholders appropriately reserve their rights, such as machine-readable means'29. Article 53 and Recital 105 of the AI Act (2024) further support this, requiring providers of generalpurpose AI to detect and respect TDM opt-out reservations, and clarifying that TDM applies to generative AI30. The regime is illustrative as a balance of recognition of copyright implications in AI training, while keeping the innovation door open with a structured opt-out (rather than a blanket ban). Yet it has considerable limitations.

Opt-out regimes are technically problematic; many AI companies do not comply; and, as Getty's experience showed, training occurs largely outside of any one jurisdiction, potentially confining the impact of any national or regional framework31. The UK's December 2024 consultation on a similar EU-style TDM exception illustrates the recognition that lawmakers must resolve these issues, as courts cannot do so alone32. VII.

Critical Analysis: A Coherent Legal Framework A. The Limits of Existing Doctrine

The paper lists five doctrines that are in play: the idea-expression dichotomy, merger 27 Matthew Sag, ‘The New Copyright Issues in AI Training’ (2024) 37 Harvard Journal of Law and Technology 1.

28 Mark A Lemley and Bryan Casey, 'Fair Learning' (2021) 104 Texas Law Review 743.

29 Directive (EU) 2019/790 on Copyright and Related Rights in the Digital Single Market [2019] OJ L130/92, Art 4. 30 Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 (AI Act) [2024] OJ L1689/1.

31 IAPP, ‘The EU AI Copyright Playbook’ (February 2026) https://iapp.org accessed 28 April 2026.

32 DLA Piper, 'Training AI Models: UK Government Proposes EU-Style Opt-Out Copyright Exception' (2025) https://dlapiper.com accessed 28 April 2026.

doctrine, substantial similarity doctrine, derivative works, and fair use, which, individually or collectively, do not meet the particular challenge of AI-mediated style exploitation. The idea-expression dichotomy assumes an era in which the reuse of an idea still involved human creativity. When Picasso painted Velázquez, he brought his genius. When a diffusion model imitates Karla Ortiz, it conducts mechanical statistical extraction at an unlimited scale, a very different process. The merger doctrine is also undermined: once a unique artist's style was literally inseparable from her expression; now it can be statistically isolated and reused. The derivative works doctrine does not apply because it regulates copying of protected expression, not imitation of unprotected style33.

And fair use, when applied to works that compete with the artist whose work provided the model, is being used beyond its limits. B. Passing Off, Moral Rights and Their Limits

Two related doctrines provide some answers. The doctrine of passing off, based on the common law tort of misrepresentation, protects traders against a defendant misrepresenting their goods as originating from, or being connected with, the claimant, resulting in goodwill loss. Where AI-generated works are sold as being "in the style of" a particular artist, passing off may occur, especially if there is consumer confusion about the commercial origins of the work. The moral rights doctrine - embodied in Article 6bis of the Berne Convention34 and domestically enshrined in the UK Copyright, Designs and Patents Act 198835 protects authors' rights of attribution and integrity, and is routinely infringed when AI commercially uses an artist's aesthetic signature without attribution or permission.

Both doctrines are, however, limited: passing off requires that there be misrepresentation and damage to goodwill in each instance; moral rights under UK law do not apply to works for hire or are often waived. Neither was designed to address systemic algorithmic exploitation36.

33 Mark A Lemley, ‘How Generative AI Turns Copyright Upside Down’ (2024)

https://ssrn.com/abstract=4517702 accessed 28 April 2026.

34 (n 18).

35 Copyright, Designs and Patents Act 1988 (UK), ss 77–80.

36 Christophe Geiger and Vincenzo Iaia, ‘The Forgotten Creator: Towards a Statutory Remuneration Right for Machine Learning of Generative AI’ (2024) 52 Computer Law and Security Review 105925.

C. A Proportionate Framework

Three changes are suggested. First, the fair use doctrine, when applied by courts, should be more rigorous in assessing the fourth factor (market harm) for AI systems that depend on substitutes for human-commissioned artwork at scale, and avoid conflating transformativeness of the training process with transformativeness of the output's market impact. The Warhol approach supports this: use for the same purpose is highly disfavoured. 37 Second, the law should provide for transparency and opt-out options: AI companies should be required to disclose datasets used for training, and artists should be able to detect and challenge the use of their works in training through standardised machine-readable reservations with legal sanctions for non-compliance. The knowledge gap between artists and AI companies is otherwise unbridgeable.

38 Third, over time, a qualified sui generis right against targeted commercial uses of a recognisable style of an identifiable living artist by AI systems should be explored. As Shan and others have shown, different artistic styles are mathematically extractable and reproducible39, showing that "style" can be defined to a legally relevant degree of specificity. This right would be limited: to living artists, to those whose style was trained into a model, and to those whose outputs are used commercially to replace the personal labour of the artist. It would not inhibit human artistic expression or further creativity: it would protect the specific harm of the commercial use of an unconsented algorithmic simulation of a human artistic identity. VIII.

Conclusion Generative AI has highlighted a disconnect between the normative goals of copyright and its doctrinal arsenal. The idea-expression distinction, the merger doctrine, the derivative works doctrine, fair use, passing off, and moral rights were all created to work in an environment in which the creative capacity of humans provided a buffer between

37 Andy Warhol Foundation for the Visual Arts Inc v Goldsmith 598 US 508 (2023).

38 ‘Consent and Compensation: Resolving Generative AI’s Copyright Crisis’ (Virginia Law Review, 2024) https://virginialawreview.org/articles/consent-and-compensation-resolving-generative-ais-copyright-crisis accessed 28 April 2026. 39 Shawn Shan and others, 'Glaze: Protecting Artists from Style Mimicry by Text-to-Image Models' (2023) arXiv 2302.04222.

inspiration and exploitation. Generative AI eliminates that buffer. The cases of Andersen v Stability AI40 and Getty Images v Stability AI41 show courts are sympathetic but helpless, and the EU's TDM regime shows that law can recognise the problem but not solve it. The precept that style is in the public domain remains true between human authors. But the idea that AI companies may algorithmically extract, mimic and monetise the entire style of a creative individual, merely because a single iteration does not copy a single protected expression, is not an idea that the values of copyright law can support. An appropriate response, robust market harm analysis, transparency and a narrow sui generis right can secure the rights of human creators without locking the creative commons. Without legal intervention, algorithmic art will not complement human creative endeavour: it will cannibalise it.

40 (n 14).

41 (n 15).