AI as the Phantom Limb: The Asymmetry of Attribution in Human vs. AI Delegation

Paper Title

AI as the Phantom Limb: The Asymmetry of Attribution in Human vs. AI Delegation

Publication Info

• Topic area: Human-AI interaction in task delegation and responsibility attribution.
• Keywords: Human-AI interaction, delegation, trust, responsibility attribution, psychological distance, feedback valence, AI Phantom Limb, competence.

Background and Problem

• Problem / challenge: Existing research on delegation in Human-AI interaction has not systematically explored how people perceive responsibility, trust, and feedback differently when delegating tasks to AI versus human assistants.
• Significance: Understanding these dynamics is critical as AI systems increasingly act as collaborators in workplaces, influencing interpersonal dynamics and perceptions of accountability.
• Motivation and related work: Prior studies have examined trust in AI and delegation decisions but have not addressed how feedback and responsibility attribution differ between human and AI delegates. This paper builds on gaps in understanding the psychological and social implications of Human-AI delegation.

Solution

• Proposed approach: A 2×2×2 experimental study examining how participants perceive trust, psychological distance, feedback, and responsibility attribution when delegating tasks to human versus AI assistants under varying conditions of competence (high vs. low) and feedback valence (positive vs. negative).
• Novelty:
  1. Identification of the "AI Phantom Limb" effect: participants internalize negative feedback toward AI assistants as self-directed but do not internalize positive feedback symmetrically.
  2. Differentiation between responsibility and accountability in Human-AI collaboration.
  3. Empirical evidence on how competence and feedback valence shape trust, psychological distance, and responsibility attribution.
• Procedure and key techniques:
  1. Participants delegated a scheduling task to either a human or AI assistant.
  2. Competence was manipulated via negotiation latency and outcomes (high competence = faster, securing ideal slots; low competence = slower, securing non-ideal slots).
  3. Feedback valence was manipulated as positive or negative client feedback.
  4. Dependent variables included trust (via a validated scale), psychological distance, responsibility attribution, and self-directed feedback perception.

Results

• Concrete findings:
  • Trust was higher for human assistants (M = 3.09) than AI assistants (M = 2.69), for high-competence assistants (M = 3.57) compared to low-competence ones (M = 2.22), and for assistants receiving positive feedback (M = 3.28) compared to negative feedback (M = 2.50).
  • Psychological distance was lower for human assistants (M = 3.62) than AI assistants (M = 3.87), for high-competence assistants (M = 3.37) compared to low-competence ones (M = 4.11), and for assistants receiving positive feedback (M = 3.42) compared to negative feedback (M = 4.07).
  • Participants internalized negative feedback toward AI assistants as more self-directed (Mnegative = 3.81) than positive feedback (Mpositive = 3.36), an asymmetry not observed with human assistants.
  • Responsibility was more often attributed to assistants (mean ~50%) than to participants themselves (~22%), with higher attribution to low-competence assistants and those receiving negative feedback.
• Advantage over baselines:
  • The study reveals unique asymmetries in feedback perception and responsibility attribution specific to AI, which were not evident in human delegation scenarios.
• Experiments / evaluation:
  • Participants: 355 individuals recruited via Prolific.
  • Design: 2×2×2 between-subjects experiment manipulating assistant type (human vs. AI), competence (high vs. low), and feedback valence (positive vs. negative).
  • Metrics: Trust (via a validated scale), psychological distance, responsibility attribution, and self-directed feedback perception.
• Limitations and future work:
  • The simplified online experiment lacked ecological validity and did not capture real-world complexities like trade-offs in competence.
  • The sample was limited to English-speaking participants from individualist cultures.
  • Future work should explore richer, context-aware paradigms, cultural differences, and longitudinal dynamics in Human-AI collaboration.

Summary

This study investigates how people perceive trust, psychological distance, feedback, and responsibility attribution when delegating tasks to human versus AI assistants. Key findings include the "AI Phantom Limb" effect, where participants internalize negative feedback toward AI assistants as self-directed but do not internalize positive feedback symmetrically. Trust and psychological distance were higher for human and high-competence assistants, while responsibility was more often attributed to assistants than participants themselves, especially under low competence or negative feedback. These findings have significant implications for designing AI systems that clarify responsibility and accountability, fostering healthier Human-AI collaboration.