Generative text-to-image systems such as DALL-E 2 produce images that can satisfy the criteria of rival theories of art at least in some cases, yet the philosophical status of these images remains unsettled: there is no principled, non-case-by-case account of when an AI-generated image is an artwork, and no agreed conditions under which the generative model itself should count as an author or artist of its outputs. The source paper (Vlaad 2024) establishes a case of art-membership under three account families and proposes conjunctive authorship (developers plus prompt-giver), while explicitly remaining agnostic about the model's own artistic status via the analogy with a twelve-sided die used in aleatoric music. Later literature (2025-2026) actively debates the authorship question through institutional, moral-agency, and relational theories, reaching directly opposed verdicts. The problem is to construct the missing account of authorship: determinate conditions for distributing authorship among the developers, the prompt-giver, and the model itself, answering the model-as-author question on which the source paper explicitly remains agnostic, and giving a determinate art-status classification and attribution for the field's fixed test case.

Background

A generative text-to-image model (paradigmatically DALL-E 2, a diffusion model) produces images from a textual prompt: it is trained on large corpora of captioned images and generates new images by iteratively denoising random noise under the guidance of the prompt, so the output is shaped by the developers' design and training data and by the prompt-giver's instructions, but not under any single agent's direct control. Boris Eldagsen's 'The Electrician' (2022), a photorealistic portrait made with DALL-E 2 that won a category of the 2023 Sony World Photography Awards before Eldagsen declined the prize on the ground that it is not a photograph, is the field's fixed test case. The philosophy of art offers three main families of accounts of what makes an object an artwork: exhibited-features accounts, which define art by properties the work displays (e.g. Ziff's 1953 criteria, including that it 'was made by an artist deliberately with skill and care' and is 'good of its kind'); genetic-features accounts, which make art-status depend on the object's origin or production history (e.g. historical or institutional definitions, where art-status is conferred by the artworld institutions, as in Dickie's institutional theory deriving from Danto's 'artworld'); and the buck-passing account (Lopes), an anti-essentialist view on which being art is not a substantive property but membership in one of the established arts, so something is art if it correctly falls under an existing art kind. Vlaad (2024) shows that The Electrician qualifies under exemplars of all three families, and diagnoses our hesitancy to call AI images art as stemming from the lack of a clear author figure. On authorship she uses 'author' in the sense of creator of a given work — she uses the terms 'author' and 'artist' interchangeably — and draws on Mag Uidhir's three-place, conjunctive account (agents A and B jointly author a work under a description, though neither does so alone) to propose that some AI images are conjunctively authored by the developers and the prompt-giver. The authorship question is pursued in the later literature (2025-2026) under two further accounts, which this task treats as among the main account families: the institutional account of authorship (Schmidt 2025), which asks whether the relevant institutional practice confers the authorial role on the candidate, and the moral-agency account (Charles 2025), which asks whether the candidate can bear moral responsibility for the work. She also notes the analogy between the model and a twelve-sided die (D12) used to compose aleatoric (chance) music: since we would not call the die a contributing artist, calling the model one may commit us to calling the die one. She explicitly remains agnostic on the model's status as artist. Separately and complementarily, she argues that DALL-E 2 itself is a work of computer art in Lopes's sense, authored by the developers, and that 'If DALL-E 2 is an artwork, then it follows that the developers are the artists' — a proposal about the system as an artwork, distinct from the model-as-author question on which she remains agnostic. Coeckelbergh (2017) had earlier shown that the criteria distinctions we would need for a general verdict (process versus outcome criteria, subjective versus objective criteria of creativity) are unstable, and framed machine art-status relationally rather than settling it. What the source paper leaves open is the authorship question: the conditions that settle whether, and when, the model itself has authorial or artistic status. The source also acknowledges that the accounts of art it examines do not solve the demarcation problem — 'there do not seem to be any straightforward ways of adjudicating which AI-generated images—if any—we ought to understand as art' — but the paper deliberately proceeds case-by-case, as 'a small case study rather than a comprehensive survey', and does not itself pose a general demarcation test as an explicit open problem.

Problem Statement

Construct a principled account that answers, for images generated by text-to-image generative models such as DALL-E 2: (i) how authorship of such images is correctly distributed among the developers of the system, the prompt-giver, and the system itself — in particular, precise conditions (necessary and sufficient, or an explicitly defended alternative such as cluster criteria) under which the generative model itself should be ascribed the status of an author or artist of the images it generates, or an argument that no such ascription is ever warranted; and (ii) how the resulting attribution, and the account's art-status classification, stand with respect to each of the main account families — exhibited-features accounts, genetic-features accounts (institutional and historical varieties), the buck-passing account, and, on the authorship side, the institutional and moral-agency accounts — stating, for each named family, whether the attribution and classification hold under the criteria that family actually applies (art-membership criteria for the art-status families; authorship and agency criteria for the authorship-side accounts) and, where they do not hold under a family's criteria, which family the account adopts and why that departure is justified. The 'authorial or artistic status' at issue is the status of a creator of the images: the source uses 'author' and 'artist' interchangeably in this sense, and the question is distinct from whether the system itself is an artwork — the source's complementary proposal that DALL-E 2 is a work of computer art authored by the developers does not answer it. The account must give a determinate art-status classification and authorship attribution for the field's fixed test case, Eldagsen's The Electrician — that is, for each family named in (ii), its determinate verdict on that case. A general demarcation criterion for AI-generated images beyond this case is not required. The account must also address the D12 analogy: it must say what principled difference, if any, separates the model's contribution to the image from a die's contribution to a chance-composed piece of music. The source paper's adjacent future-direction questions — which art competitions AI-generated images ought to be allowed into and how they should be attributed there (including the possibility of separate competitions for AI-generated images), and the aesthetics and compositional strategies unique to AI-generated images as a new artform — are related but separable programs and are not part of this problem's completion boundary.

References

  1. Vlaad, Sofie (2024). 'A portrait of the artist as a young algorithm'. Ethics and Information Technology 26(3), article 58. DOI: 10.1007/s10676-024-09796-0. https://doi.org/10.1007/s10676-024-09796-0
  2. Schmidt, João Vitor (2025). 'Can Artificial Agents be Authors?'. Philosophy & Technology 38(1), article 13. DOI: 10.1007/s13347-025-00839-y. https://doi.org/10.1007/s13347-025-00839-y
  3. Charles, David R. (2025). 'AI as artist: agency and the moral rights of creative works'. AI and Ethics 5(4), 3833-3844. DOI: 10.1007/s43681-025-00679-8. https://doi.org/10.1007/s43681-025-00679-8
  4. Ogundipe, Anne (2025). 'Generative AI, agency and aesthetic disruption in photographic art'. Journal of Aesthetics & Culture 17(1). DOI: 10.1080/20004214.2025.2557675. https://doi.org/10.1080/20004214.2025.2557675
  5. Berryman, Jim (2026). 'On Style and Mere Stylishness in AI-Generated Art'. The Journal of Aesthetics and Art Criticism (advance article). DOI: 10.1093/jaac/kpaf042. https://doi.org/10.1093/jaac/kpaf042
  6. Coeckelbergh, Mark (2017). 'Can Machines Create Art?'. Philosophy & Technology 30(3), 285-303. DOI: 10.1007/s13347-016-0231-5. https://doi.org/10.1007/s13347-016-0231-5
  7. Misselhorn, Catrin (2025). 'Artificial Intelligence, Authorship and Aesthetic Responsibility in Art'. In: Ästhetik, Digitalisierung und Künstliche Intelligenz, pp. 17-36. DOI: 10.30965/9783969753477_003. https://doi.org/10.30965/9783969753477_003