Anders Sundnes Løvlie

Generative AI is rapidly transforming creative practice, yet its limitations (bias, censorship, opacity, instability, and ethical uncertainty) are often treated as flaws to be fixed. Artists, however, have historically worked with the malfunctions of emerging technologies as creative material. Drawing on interviews with 13 contemporary artists, we identify five recurring forms of constraint in GenAI systems and five design orientations artists use to rework them: Exposure (making seams visible), Redirection (working around constraints), Systematisation (making failure repeatable), Practiced Governance (ethical self-regulation), and Scaffolding (designing interpretive contexts for audiences). Instead of hiding seams, artists treat them as resources for reflection, critique, and creative expansion. We frame these practices as forms of seamful and constraint-based design, arguing that artists working at the edge of GenAI systems offer HCI a model for designing with limitations rather than against them.

Generative AI challenges established notions of authorship in contemporary art, yet little attention has been paid to how technical aspects of AI systems are disclosed in exhibition contexts. This paper presents a qualitative analysis of 23 AI-mediated artworks, with a focused discussion of 13 cases selected to illustrate distinct disclosure practices. We examine how model architectures, data provenance, interaction modalities, and energy consumption are framed or omitted in attribution materials. Our analysis reveals recurring asymmetries: model architectures are frequently named but rarely explained, data are disclosed selectively, and energy use is systematically absent. Instead of framing technical disclosure as an obligation, we approach it as a designed practice shaped by the artist. Three disclosure strategies emerge: extended technical framing, symbolic technical naming, and strategic opacity. In a climate of growing skepticism toward AI-generated outputs, technical attribution is not neutral reporting, but an active site where legitimacy and authorship are constructed.

We present a work in progress that explores using a Large Language Model (LLM) as a design material for an interactive museum installation. LLMs offer the possibility of creating chatbots that can facilitate dynamic and human-like conversation, engaging in a form of role play to bring historical persons to life for visitors. However, LLMs are prone to producing misinformation, which runs counter to museums' core mission to educate the public. We use Research-through-Design to explore some approaches to navigating this dilemma through rapid prototyping and evaluation and propose some directions for further research. We suggest that designers may shape interactions with the chatbot to emphasize personal narratives and role play rather than historical facts or to intentionally highlight the unreliability of the chatbot outputs to provoke critical reflection.

GenFrame is an interactive picture frame that mimics traditional paintings while being equipped with generative AI capabilities. Painting “portraits of a girl” has been a trope in art history for centuries. However, in the era of generative AI models, the permanence of traditional art is challenged. When an AI model is able to mimic any painting style, it can also give agency to the viewer to repaint the painting per individual desires. In this installation, we modify the role of the museum placard to provide a minimal tangible interface to change the style of the image and the mood of the depicted girl. In this way, the generative AI model is interfaced in tangible ways, instead of the regular prompt paradigm. Demo video available here: https://youtu.be/ga7pFgAOPiY

Deep Generative Models (DGM) have had a great impact both on visual art and broader visual culture. In this research-through-design project we investigate the use of a DGM for helping museum visitors explore the aesthetics of Edvard Munch’s art. We designed and built an interactive drawing table that allows a user to explore a StyleGAN model trained on sketches by Edvard Munch. The paper makes two novel contributions: 1. It presents a system that allows users to interact with a DGM by drawing on paper (rather than the typical text prompts used by most current systems). 2. We demonstrate how this mode and quality of interaction establish a unique perspective on Munch’s drawings as a practice. Through qualitative evaluation, we discuss how this setup led users towards a specific hermeneutic drawing strategy that enables building competency with the model and by proxy the data it is trained on. We suggest that the resulting interaction may contribute to an "education of attention" helping museum visitors to become attentive to certain visual qualities in Munch’s drawing practice. Finally, we discuss how the concepts of technological mediation and relationality are useful for designing how the output of a DGM is understood by its users.