Through the Looking-Glass: AI-Mediated Video Communication Reduces Trust and Confidence in Judgement

Paper Title

Through the Looking-Glass: AI-Mediated Video Communication Reduces Trust and Confidence in Judgement

Publication Info

• Topic area: Effects of AI-mediated video communication on trust, judgment accuracy, and confidence.
• Keywords: AI-mediated communication, video retouching, avatars, trust, deception detection, judgment confidence, expectancy violations, uncertainty reduction, human-computer interaction.

Background and Problem

• Problem / challenge: AI-mediated video tools, such as retouching, background replacement, and avatars, may alter how people evaluate trustworthiness, credibility, and honesty in online communication. Prior research lacks clarity on how these tools affect deception detection, trust, and confidence in judgments.
• Significance: Understanding these effects is critical for designing trustworthy AI tools, especially in high-stakes contexts like hiring, legal proceedings, and telemedicine, where trust and confidence are essential.
• Motivation and related work: Previous studies in computer-mediated communication (CMC) and AI-mediated communication (AI-MC) suggest that reduced or altered interpersonal cues can erode trust and complicate impression formation. However, the impact of everyday AI-mediated video tools on deception detection and interpersonal evaluations remains underexplored, particularly in dynamic video contexts.

Solution

• Proposed approach: The study investigates how varying levels of AI-mediated video communication—original recordings, weak mediation (retouching and virtual backgrounds), and strong mediation (avatars)—affect trust, deception detection accuracy, and judgment confidence.
• Novelty:
  1. Examines the effects of everyday AI-mediated video tools (e.g., retouching, avatars) on trust and confidence.
  2. Tests these effects in both homogeneous (single mediation type) and mixed (multiple mediation types) environments.
  3. Explores shifts in cue reliance (e.g., from nonverbal to content-based cues) under AI mediation.
  4. Provides empirical evidence challenging cue-based deception detection theories.
• Procedure and key techniques:
  • Conducted two online experiments (N = 2,000) using prerecorded videos of truthful and deceptive statements.
  • Videos were processed into three conditions: original, weak AI mediation, and strong AI mediation.
  • Participants judged trustworthiness, truthfulness, and confidence after watching each video.
  • Statistical analysis included linear mixed-effects models to evaluate the effects of AI mediation and environment type.

Results

• Concrete findings:
  • Trust in speakers decreased under AI-mediated conditions, especially for avatars (strong mediation) and in mixed environments.
  • Judgment accuracy (52–54%) remained stable across all mediation types, aligning with prior findings of near-chance accuracy in deception detection.
  • Confidence in judgments decreased significantly in mixed environments, particularly for avatar-mediated videos.
  • Participants shifted reliance from nonverbal cues (e.g., gaze, expressions) to content-based cues (e.g., story consistency, fluency) when evaluating AI-mediated videos.
• Advantage over baselines:
  • Demonstrates that AI-mediated video does not impair deception detection accuracy but reduces trust and confidence, especially in mixed settings.
  • Challenges cue-based theories of deception detection by showing stable accuracy despite reduced nonverbal cues.
• Experiments / evaluation:
  • Study 1: Homogeneous environment with a single mediation type per participant.
  • Study 2: Mixed environment with varying mediation types per participant.
  • Metrics: Trust ratings, truth judgment rates, judgment accuracy, and confidence levels.
  • Dataset: Miami University Deception Detection Database (MU3D).
• Limitations and future work:
  • Low-stakes experimental context may not generalize to high-stakes scenarios.
  • Results may differ in live, interactive video settings or with familiar participants.
  • Future research should explore high-stakes contexts, longitudinal effects, and cross-cultural differences.

Summary

This study investigates the impact of AI-mediated video communication on trust, deception detection accuracy, and judgment confidence. While AI mediation did not affect accuracy or truth judgment rates, it reduced trust and confidence, particularly for avatar-mediated videos in mixed environments. Participants shifted from relying on nonverbal cues to content-based cues under AI mediation. These findings highlight the need for context-sensitive AI tools and raise concerns about the erosion of trust in mediated communication, especially in high-stakes or mixed environments.