⚡ electrical engineering

Trust–reliance alignment is context-sensitive in automated driving: a within-person secondary analysis of open experimental data

This study reanalyzes experimental data to demonstrate that while an individual's moment-to-moment trust in automation generally aligns with their reliance decisions, this relationship is significantly weaker in highway contexts compared to suburban ones, suggesting that interface design should prioritize supporting timely control reassessment rather than simply aiming to increase average trust levels.

Original authors: Xiaoqing Zhao

Published 2026-09-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xiaoqing Zhao

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When a car drives itself, the human inside is not merely a passenger; they are a partner in a constantly shifting dance of responsibility. The vehicle handles the steering and speed, but the person must decide how much to trust the machine at any given moment. This relationship relies on two distinct things. First, there is trust, which is a feeling—a general belief that the technology is safe and reliable. Second, there is reliance, which is the actual choice to let the car take control or to grab the wheel back. In the real world, these two do not always move in lockstep. A driver might generally trust a self-driving system but still hesitate to let it handle a sudden, dangerous situation. Understanding how these feelings and actions connect, especially when the road conditions change, is vital for building vehicles that keep people safe.

Researchers recently looked closely at this connection by re-examining data from a large experiment involving 206 licensed drivers. In the original study, these participants watched video scenes of a self-driving car navigating two very different environments: a busy highway and a quiet suburban neighborhood. Each scenario included a moment of surprise, such as another car merging dangerously close or a vehicle slowing down unexpectedly, followed by a resolution. After watching each scene, the drivers were asked to choose how much control they wanted to give the automated system, ranging from taking over the wheel completely to letting the car drive with their eyes off the road. They also rated how much they trusted the vehicle in that specific moment. The new analysis took these responses and separated the drivers' general, long-term trust from their moment-to-moment changes in trust. This allowed the researchers to see if a single person would hand over more control when their trust spiked, and if that pattern held true across different types of roads.

The findings reveal a clear but nuanced picture. When a driver's trust in the system increased for a specific scene, they were indeed willing to give the car more control. This link between feeling and action was strong and consistent, appearing even after accounting for the driver's age, their general attitude toward robots, and how often they drove. However, the context of the drive mattered just as much as the feeling. In the high-speed highway scenes, drivers consistently chose to keep more control for themselves, regardless of how much they trusted the car. Even when a driver felt very confident in the system, the danger of the highway environment made them pull back, reducing the amount of automation they were willing to accept. Furthermore, the connection between trust and action was weaker on the highway than in the suburbs. In the suburban setting, a rise in trust led to a more direct increase in letting the car drive, whereas on the highway, that link was looser, and drivers remained more cautious.

The study also looked at how drivers reacted when the unexpected events happened and then were resolved. When a surprise occurred, both trust and the willingness to let the car drive dropped. When the situation was resolved, trust often bounced back quickly, but the willingness to let the car drive did not always recover to the same degree. This suggests that a driver might understand that a problem is solved and feel safe again, yet still hesitate to fully hand over control. The researchers found that simply showing drivers more information on the screen did not fix this gap. Providing a detailed display of what the car was seeing and planning did not make the link between trust and reliance any stronger or more consistent. The extra information did not help drivers translate their feelings into better decisions about control.

These results suggest that designing self-driving cars requires more than just trying to make people feel good about the technology. The goal should not be to maximize the average level of trust a driver feels, but rather to support the driver in making the right decision at the right time. A system that feels trustworthy on a calm day might still need to be monitored closely on a busy highway. The study indicates that interfaces and safety protocols should focus on helping drivers reassess control quickly when the situation changes, rather than assuming that a high trust score means the driver is ready to let go. The most important factor is not just whether a driver trusts the car, but whether they can adjust their reliance to match the immediate risks of the road.

Technical Summary: Trust–Reliance Alignment in Automated Driving

Problem Statement
Current research on automated driving often treats trust as a static, person-level attitude, assuming that a user's general willingness to trust an automated system predicts their reliance behavior across all situations. However, reliance decisions unfold dynamically as traffic conditions change. A critical gap exists in distinguishing between between-person differences (stable traits where some individuals generally trust more than others) and within-person dynamics (how a specific individual adjusts their reliance when their trust fluctuates across different scenes). Furthermore, it remains unclear whether the coupling between trust attitude and reliance behavior is consistent across different contexts (e.g., highway vs. suburban) or interface designs (high vs. low information). Aggregate correlations often obscure whether an individual actually updates their control allocation in response to situational changes.

Methodology
This study conducts a secondary analysis of open experimental data from Tanevska et al. (2025a), involving 206 licensed drivers who evaluated six scenes across highway and suburban scenarios. Each scenario included an unexpected event and a resolution.

Design: A 2 × 2 between-participant design manipulating interface information (high vs. low) and scenario order.
Measures: Participants selected a Level of Automation (LoA) from 0 (takeover) to 3 (eyes-off) after each scene, followed by a trust attitude rating.
Analytical Approach: The study employed person-mean decomposition to separate stable between-person trust from scene-specific within-person trust deviations. Cluster-robust ordinary least squares (OLS) models were estimated to predict the conferred LoA.
Controls: Models controlled for scenario type, scene sequence, interface information, order effects, age, negative attitudes toward robots (AV-NARS), and driver behavior scores (DBQ). Sensitivity analyses included comfort levels.
Event Analysis: Spearman correlations were calculated for changes in trust and LoA across event transitions (Scene 1→2 and 2→3) to assess the alignment of dynamic shifts.

Key Results

Within-Person Alignment: There is a positive alignment between within-person trust deviations and conferred automation. When a participant's trust in a specific scene was higher than their personal average, they were willing to confer approximately 0.29 additional LoA units ( $b = 0.291$ , 95% CI [0.219, 0.362]), independent of their general trust level.
Context Sensitivity: The highway context significantly reduced the level of automation participants were willing to confer ( $b = -0.612$ ) compared to suburban settings, regardless of trust levels.
Weakened Coupling in High-Risk Contexts: The coupling between trust and reliance was weaker in highway contexts. The interaction term indicated that for every point increase in trust, the increase in conferred LoA was 0.080 units lower on the highway than in the suburbs ( $b = -0.080$ ).
Interface Information: High-information interfaces did not measurably strengthen the within-person trust–reliance coupling ( $b = -0.021$ , $p = .599$ ).
Event Transitions: Changes in trust and changes in reliance were positively correlated across event transitions but only modestly so (Spearman $\rho = 0.239–0.324$ ). Notably, in suburban scenarios, trust recovered more strongly than reliance after an event resolution, suggesting a lag or asymmetry in behavioral restoration.

Key Contributions

Methodological Distinction: The study explicitly separates stable between-person dispositions from situation-level updating, avoiding the ecological fallacy of interpreting aggregate correlations as individual dynamics.
Dynamic Control Allocation: It demonstrates that trust and reliance are related but non-interchangeable; situational criticality (e.g., highway driving) shifts control allocation independently of concurrent trust attitudes.
Design Implications: The findings challenge the design objective of maximizing average trust. Instead, they suggest that interfaces and procedures should support timely control reassessment, as higher trust does not automatically guarantee appropriate reliance in high-risk contexts.

Significance and Claims
The paper claims that for safety-critical automated systems, the operational question is not merely whether users trust the technology, but how their reliance choices adjust to specific situations. The study concludes that trust level and context-sensitive control allocation are related but distinct constructs. Consequently, human–automation systems should prioritize supporting the dynamics of control allocation—helping users translate changing evidence and risk into proportionate decisions—rather than treating higher average trust as the primary success metric. The author notes that while high information did not improve trust–reliance coupling in this dataset, information quantity alone is not equivalent to effective decision support; information must be diagnostic and timely to aid control reassessment.

Note: The author emphasizes that these findings are based on self-reported actions in static vignettes, not observed on-road behavior, and that the secondary analysis was exploratory, requiring replication for causal claims.

Technical Summary: Trust–Reliance Alignment in Automated Driving

More like this