Designing Attribution: Technical Disclosure in AI-Mediated Art

Traditional attribution formats like “oil on canvas” or bronze sculpture assume stable materials and singular authorship. Generative AI (GenAI) disrupts these assumptions. In AI-based artistic practice, agency is often distributed across datasets, models, prompts, and curatorial decisions, leaving the human-AI relationship often ambiguous and raising questions about who or what contributed to the creation of an artwork.

At the same time, the use of Generative Artificial Intelligence in art has become a polarising medium, provoking both enthusiasm and resistance [5]. Empirical studies show persistent preference for human-generated art, with AI-assisted artworks often judged as less authentic or less valuable [21, 22]. Public discourse increasingly associates generative AI with low-quality, exploitation, and cultural harm [13]. The term "AI slop" has gained attention in this context, describing mass-produced, low quality AI generated content [1, 19, 24, 26, 27]. Such framings contribute to a broader climate of growing skepticism, in which AI-generated outputs are framed as careless, exploitative or environmentally harmful, further creating a risk of a public backlash towards AI [6]. These dynamics raise the stakes for how artists working with AI communicate their role, intentions, and responsibilities.

Recent research has extensively classified AI Art practices according to artistic motivations, technical control, or conceptual orientation [4, 14, 25]. These frameworks provide an understanding of how artists engage with AI, but they offer limited insights of how such engagements are publicly communicated. In practice, attribution is often treated as a neutral afterthought (reduced to a brief label description), even though it may play a central role in shaping how artworks are interpreted, valued and trusted. Studies show that audiences tend to judge artworks that include AI as less-authentic or less valuable than artworks created without the use of AI [21]. While debates in AI scholarship and regulation address whether and how AI use should be disclosed [6, 10], less attention has been paid to attribution as a public interface in artistic settings.

In this paper, we argue that attribution in AI-mediated art should be understood as a designed interface rather than a fixed descriptor. We examine how technical aspects of AI systems, such as model architecture, data provenance, interaction modalities, and energy use, are publicly framed, emphasized, or omitted in attribution materials of AI-mediated artworks. Based on a qualitative analysis of 23 artworks, with a focused discussion of 13 cases, we identify recurring patterns and asymmetries in technical disclosure practices by contemporary artists.

Our analysis suggests that artists adopt different disclosure strategies: some integrate technical detail into the conceptual framing of the work; others reference models symbolically without elaboration; and some maintain opacity as part of a narrative or critical gesture. However, our goal is not to propose a normative standard for art transparency, but to strengthen the discussion of how a disclosure and attribution can become a design act. This short paper contributes to discussions on attribution in AI-mediated art and museum labeling by framing attribution as a design space in a climate of growing skepticism towards AI-generated outputs.

To explore how artists publicly disclose their use of AI, we conducted a qualitative content analysis of publicly available attribution materials: museum labels, exhibition descriptions, artist statements, official project websites, and where necessary, interviews and essays authored by the artist.

Sampling. We assembled a purposive sample of 23 AI-mediated artworks, seeking variation across artistic strategy, institutional context (museums, galleries, online platforms), and technical approach (GANs, diffusion models, custom systems). We included widely exhibited and frequently cited works, as well as projects explicitly engaging questions of authorship and agency. The goal was not representativeness but coverage of diverse disclosure practices. We selected 13 cases for focused discussion here as they most clearly illustrate the range of disclosure strategies identified across the full sample. The remaining 10 works will be incorporated in an expanded version of this study. Where available, we recorded the year of first public exhibition for each artwork, as production dates bear on interpretation given the shift from GAN-based to diffusion-based systems around 2022–2023.

Coding. Three co-authors coded the full sample of 23 artworks independently. Differences were discussed across multiple sessions and categories refined iteratively until consensus was reached. Drawing on prior work on AI disclosure frameworks [8], taxonomies of AI art practice [4, 14], debates over training data and authorship [10], and growing evidence of AI's environmental costs [17, 29, 30], we coded each artwork along four dimensions: (1) model architecture: whether and how the AI system is named or described; (2) data provenance: the origin, scale, and licensing of training data; (3) input–output modalities: how the system receives and produces information; and (4) energy consumption: whether training, inference, or exhibition power usage are mentioned. For each dimension we documented presence, absence, and framing of technical information rather than verifying technical correctness.

Our analysis reveals uneven patterns in how technical systems are publicly disclosed (see Table 1). The different categories are not treated equally, but selectively framed depending on the artist.

Our analysis reveals clear asymmetries in how technical systems are disclosed across AI-mediated artworks. These asymmetries are not arbitrary: they reflect specific pressures on artists working in exhibition contexts, where attribution materials must simultaneously address curators, critics, general visitors, and increasingly skeptical public.

Across cases, three disclosure strategies emerge:

1. Extended technical framing. Works such as "Gender Shades", "Mosaic Virus", and "Holly+" integrate technical detail into their conceptual logic. In those artworks the data is the argument, and provenance becomes a statement about critique, consent, or ownership. Extended disclosure here may be a strategic move that strengthen legitimacy, and aligns with Messer's [21] finding that disclosing the artist's role can mitigate audience scepticism.
   
2. Symbolic technical naming. The majority of works name the AI system (GAN, VAE, StyleGAN, etc.) without elaboration. These names signal technological literacy and locate the work within a recognisable genealogy of AI art, while remaining opaque to general visitors. This partial disclosure may not satisfy the demands for transparency and accountability put forward by some scholars in other contexts [23], but such concerns may need to be weighed against an artist's aesthetic choices in only disclosing partial information.
   
3. Strategic opacity. Several works withhold technical detail as a deliberate choice. Flynn discloses almost nothing; and this strategic opacity is a choice, not an accident [7]. However, the same gesture that reads as critical in a media art context may read as misleading to a public already sceptical of AI.
   

Across all 13 cases, energy consumption is entirely absent from attribution materials, even in large-scale works like "Unsupervised". This is the sharpest asymmetry in our analysis. It may reflect absent norms, unavailable data from commercial providers, or simply a lack of established vocabulary for communicating energy use in artistic contexts. As evidence of AI's substantial environmental footprint continues to grow [17, 29, 30], this absence becomes increasingly difficult to ignore.

To sum up, the three strategies show that disclosure is shaped by artistic intent, institutional context, and strategic positioning, not by any standard of transparency. Different choices carry different implications for how authorship is constructed and communicated. Progressive disclosure formats [28] could serve both general and specialist audiences through layered attribution.

In this short paper we identified three disclosure strategies: (1) extended technical framing, (2) symbolic technical naming, and (3) strategic opacity, and found energy consumption to be the most consistently absent dimension across all cases. Our findings are grounded in a GAN-era sample and should be interpreted accordingly. The shift toward diffusion models, with their larger training corpora makes the question of technical attribution more urgent. Future work should extend this analysis to diffusion-based practice and examine how artists, curators, and institutions are navigating disclosure in this newer and more legally complicated context.

This work also opens several design directions. Layered attribution formats could serve audiences with different levels of technical literacy simultaneously. Curatorial guidelines developed with artists could establish a shared vocabulary for energy disclosure without imposing a single standard. And HCI researchers could study empirically how different attribution formats shape audience interpretation, trust, and perceived artistic value in gallery and online settings. We suggest that attribution deserves to be treated as both a design problem and an open negotiation over what it means to make and account for art with AI.