Trust Formation in AI Delegation: The Interplay of Explainability and Anthropomorphism

Paper Title

Trust Formation in AI Delegation: The Interplay of Explainability and Anthropomorphism

Publication Info

• Topic area: Human-AI interaction and trust formation mechanisms
• Keywords: Explainable AI (XAI), anthropomorphism, trust formation, cognitive engagement, eye-tracking, human-computer interaction, dual-process theory, cognitive load, user experience, AI delegation

Background and Problem

• Problem / challenge: The interaction between explainability (XAI) and anthropomorphism in AI trust-building is unclear. Prior research has studied these strategies in isolation, leaving gaps in understanding their combined effects, particularly whether they synergize or interfere.
• Significance: Understanding this interaction is critical for designing AI systems that foster calibrated trust, especially in high-stakes applications like financial planning, medical diagnosis, and e-commerce.
• Motivation and related work: XAI enhances cognitive trust by making AI logic transparent, while anthropomorphism fosters emotional trust through human-like cues. However, these strategies may compete for user attention, leading to cognitive overload. Existing studies rely heavily on self-reports, failing to capture real-time cognitive mechanisms, which this paper addresses using eye-tracking.

Solution

• Proposed approach: A mixed-methods investigation combining a large-scale online experiment and a controlled lab study with eye-tracking to examine the interaction effects of XAI and anthropomorphism on trust.
• Novelty:
  1. Identification of context-dependent interactions between XAI and anthropomorphism, showing interference in low-engagement settings and synergy in high-engagement settings.
  2. Use of eye-tracking to uncover real-time cognitive mechanisms driving these effects.
  3. Introduction of a Context-Dependent Interaction Framework to guide AI interface design.
• Procedure and key techniques:
  • Study 1: Online experiment (N=900) using a 2x2 factorial design to test XAI and anthropomorphism effects on trust in a low-engagement context.
  • Study 2: Lab experiment (N=57) with eye-tracking to replicate findings in a high-engagement context and explore attentional mechanisms.
  • Key metrics: budget allocation (behavioral trust), perceived ability, fixation durations, gaze transitions, and pupil dilation.

Results

• Concrete findings:
  • Study 1: In low-engagement online settings, XAI and anthropomorphism independently increased trust, but their combination led to interference, reducing trust relative to XAI alone.
  • Study 2: In high-engagement lab settings, the hybrid (XAI + anthropomorphism) agent elicited the highest trust, showing a synergistic effect.
  • Eye-tracking revealed that XAI increased cognitive engagement (e.g., longer fixation durations), which redirected attention to anthropomorphic cues, enabling their integration into trust judgments.
• Advantage over baselines:
  • XAI-only agents consistently outperformed control and anthropomorphic-only agents in both studies.
  • The hybrid agent outperformed all others in high-engagement settings, increasing budget allocation by 34.81 percentage points and perceived ability by 0.88 points.
• Experiments / evaluation:
  • Study 1: Online experiment with a shopping task, measuring trust through budget allocation and trust scales.
  • Study 2: Lab experiment with eye-tracking, analyzing fixation patterns, gaze transitions, and pupil dilation to uncover cognitive mechanisms.
• Limitations and future work:
  • Limited generalizability due to specific tasks (shopping) and modalities (text and static avatars).
  • Need for replication in broader populations and high-stakes domains (e.g., healthcare, finance).
  • Future work should manipulate engagement levels within a single paradigm and explore trust dynamics over repeated interactions.

Summary

This paper investigates the interaction between explainability (XAI) and anthropomorphism in AI trust formation, revealing context-dependent effects. In low-engagement online settings, these cues interfere, while in high-engagement lab settings, they synergize. Eye-tracking data show that XAI activates cognitive engagement, enabling users to integrate anthropomorphic cues into trust judgments. The proposed Context-Dependent Interaction Framework provides actionable insights for designing AI systems tailored to user engagement levels, with implications for diverse applications in human-AI interaction.