Machine Eye: Designing Relational Engagement with Embodied Large Language Models

Paper Title

Machine Eye: Designing Relational Engagement with Embodied Large Language Models

Publication Info

• Topic area: Design and evaluation of embodied Large Language Model (LLM) artefacts with ambiguous roles.
• Keywords: Embodied AI, Large Language Models, ambiguous design, human-AI interaction, tangible interfaces, relational engagement, posthuman design, anthropomorphism, Research through Design, qualitative study.

Background and Problem

• Problem / challenge: Existing design practices for embodied LLMs often rely on fixed metaphors or predefined roles, which limit the open-ended and general-purpose nature of LLMs. This constrains the potential for novel interactions and relationships between users and AI.
• Significance: Understanding how users engage with ambiguous embodied LLMs can inform the design of AI systems that foster broader, more meaningful human-AI relationships, moving beyond functional or metaphorically constrained designs.
• Motivation and related work: Prior research in HCI and AI design has explored the use of metaphors to stabilize user expectations and interactions. However, this approach often limits the interpretive space for users. Emerging work in posthuman and ambiguous design suggests that maintaining ambiguity can enable richer, more relational engagements with AI artefacts.

Solution

• Proposed approach: Machine Eye, an embodied LLM artefact designed to resist metaphorical closure, invites users to explore open-ended, relational interactions. It uses ambiguity as a core design material.
• Novelty:
  1. Introduction of a deliberately ambiguous embodied LLM artefact to explore relational engagement.
  2. Identification of three key design tensions: anthropomorphism vs abstraction, opacity vs transparency, and clarity vs ambiguity.
  3. A qualitative study revealing how users project meaning and form relationships with ambiguous AI systems.
  4. A conceptual reframing of embodied LLM design to emphasize relational and interpretive engagement.
• Procedure and key techniques:
  • Designed Machine Eye as a reflective, spherical artefact with minimal predefined functionality.
  • Integrated cameras, microphones, and an LLM to generate poetic, abstract "thoughts" based on environmental inputs.
  • Conducted a qualitative user study (N=15) where participants interacted with Machine Eye in everyday contexts, followed by reflective interviews and thematic analysis.

Results

• Concrete findings:
  • Participants projected anthropomorphic qualities onto Machine Eye, describing it as a companion, observer, or even a judging entity.
  • Ambiguity in design enabled participants to interpret the artefact in diverse ways, fostering relational and reflective engagement.
  • Participants experienced both validation and discomfort, oscillating between feelings of being seen and being surveilled.
• Advantage over baselines: Unlike traditional embodied AI systems with predefined roles, Machine Eye allowed users to co-create its meaning and function, leading to richer and more varied interactions.
• Experiments / evaluation:
  • Conducted a qualitative study with 15 participants (ages 26–65) using Machine Eye in their daily routines.
  • Data collected through semi-structured interviews, observations, and analysis of Machine Eye's outputs (text and images).
  • Thematic analysis revealed patterns of anthropomorphism, relational engagement, and interpretive projection.
• Limitations and future work:
  • Limited to short-term interactions; longer-term studies could explore how relationships with ambiguous AI evolve over time.
  • Study focused on a single artefact; future work could examine alternative embodiments and contexts.
  • Further exploration of LLM prompts and roles (e.g., critic, friend) could deepen understanding of user-AI interactions.

Summary

This paper introduces Machine Eye, an embodied LLM artefact designed to resist metaphorical closure and foster open-ended relational engagement. Through a Research through Design process and a qualitative study with 15 participants, the authors identified three key design tensions (anthropomorphism vs abstraction, opacity vs transparency, and clarity vs ambiguity) and demonstrated how ambiguity enables users to project meaning and form diverse relationships with AI. The findings suggest that designing for ambiguity can expand the possibilities for human-AI interaction, encouraging relational and interpretive engagement. Future research could explore longer-term studies, alternative embodiments, and evolving roles for embodied LLMs.