[Paper Review] Trust and Reliance in XAI -- Distinguishing Between Attitudinal and Behavioral Measures

Nicolas Scharowski, Sebastian A. C. Perrig+2 more|arXiv (Cornell University)|Mar 23, 2022

Personal Information Management and User Behavior17 citations

TL;DR

This paper argues for a critical distinction between attitudinal trust and behavioral reliance in XAI research, demonstrating that conflating these constructs leads to inconsistent findings on transparency's impact. By advocating for separate operationalization—using surveys for trust and behavioral metrics for reliance—it provides a theoretical foundation to improve measurement rigor and resolve ambiguities in XAI evaluation.

ABSTRACT

Trust is often cited as an essential criterion for the effective use and real-world deployment of AI. Researchers argue that AI should be more transparent to increase trust, making transparency one of the main goals of XAI. Nevertheless, empirical research on this topic is inconclusive regarding the effect of transparency on trust. An explanation for this ambiguity could be that trust is operationalized differently within XAI. In this position paper, we advocate for a clear distinction between behavioral (objective) measures of reliance and attitudinal (subjective) measures of trust. However, researchers sometimes appear to use behavioral measures when intending to capture trust, although attitudinal measures would be more appropriate. Based on past research, we emphasize that there are sound theoretical reasons to keep trust and reliance separate. Properly distinguishing these two concepts provides a more comprehensive understanding of how transparency affects trust and reliance, benefiting future XAI research.

Motivation & Objective

• To address the inconsistent empirical findings on transparency's effect on trust in XAI by identifying conceptual confusion between trust and reliance.
• To argue that trust (an attitudinal, subjective construct) and reliance (a behavioral, objective outcome) are theoretically distinct and should not be operationally conflated.
• To highlight that using behavioral measures (e.g., reliance rates) to assess trust—while claiming to measure trust—leads to misinterpretation and flawed conclusions.
• To promote a more rigorous scientific foundation in XAI by advocating for clear operational definitions and measurement frameworks.
• To encourage future research to measure both trust and reliance simultaneously, informed by established theoretical models of human-AI interaction.

Proposed method

• Proposes a theoretical framework distinguishing trust as a psychological attitude from reliance as observable behavior, grounded in social psychology and human-automation interaction theory.
• Reviews empirical XAI studies (e.g., Lai & Tan, 2019; Cheng et al., 2019) to illustrate how reliance is often mislabeled as trust in operationalization.
• Applies Ajzen and Fishbein’s (1980) attitude-behavior relationship model, emphasizing that attitudes (like trust) do not deterministically predict behavior (like reliance).
• Draws on established models of trust and reliance (e.g., Lee & See, 2004; Hoff & Bashir, 2015) to justify the conceptual separation and guide measurement design.
• Recommends a dual-measurement approach: using validated survey scales for trust and direct behavioral observation or task performance metrics for reliance.
• Calls for a systematic literature review to assess the extent of conceptual inconsistency in current XAI research and to develop a standardized measurement framework.

Experimental results

Research questions

• RQ1 To what extent are trust and reliance conflated in current XAI research, and what are the consequences of this conflation?
• RQ2 Why is it theoretically and empirically problematic to use behavioral measures of reliance as proxies for attitudinal trust?
• RQ3 How can a clear distinction between trust and reliance improve the validity and reproducibility of XAI evaluation studies?
• RQ4 What are the implications of measuring both trust and reliance simultaneously for understanding transparency’s impact on human-AI interaction?
• RQ5 How can future XAI research adopt a more rigorous measurement framework that differentiates between subjective attitudes and objective behaviors?

Key findings

• Empirical studies in XAI frequently conflate behavioral reliance with attitudinal trust, leading to ambiguous and inconsistent findings about transparency’s effects.
• Lai and Tan (2019) operationalized trust as reliance rate (behavioral), while Cheng et al. (2019) used survey-based trust scales—demonstrating divergent measurement approaches for the same construct.
• There is no deterministic link between trust and reliance: individuals may express high trust but exhibit low reliance, and vice versa, due to situational, cognitive, or workload-related factors.
• Theoretical models such as Lee and See’s (2004) trust-reliance framework show that trust is a necessary but not sufficient condition for reliance, underscoring the need for separate measurement.
• Mislabeling reliance as trust leads to misinterpretation of results and undermines the scientific rigor of XAI evaluation, particularly in comparative studies.
• A dual-measurement approach—assessing both trust (via validated scales) and reliance (via behavioral metrics)—is essential for accurate, theory-grounded XAI research.