Source-linked AI summary
Look Who's Talking Now: Implications of AV's Explanations on Driver's Trust, AV Preference, Anxiety and Mental Workload
Na Du, Jacob Haspiel, Qiaoning Zhang, Dawn Tilbury, Anuj K. Pradhan, X. Jessie Yang, Lionel P. Robert
TL;DR
The paper examines whether AV explanations promote acceptance, addressing uncertainty about their effects on AV users. In a within-subject driving-simulator experiment, it compares explanation conditions across trust, preference, anxiety, and workload, finding that before-action explanations were associated with higher trust and AV preference. The study’s scope is limited by simulator-based measurement and overlap between preference and trust ratings.
Problem
It remains unclear whether explanations promote acceptance of AVs, despite their reported benefits for trust in other automation.
Method
The study examines AV explanations and their timing and autonomy conditions across trust, AV preference, anxiety, and mental workload.
Results
Explanations provided before AV action were associated with higher trust in and preference for the AV, while anxiety and workload did not differ.
Takeaways & Limitations
Explanation is not enough; AVs need to provide explanations before acting to promote trust and preference.
Takeaways & Limitations
The high-fidelity driving simulator still reduced the realism of the study setting, and the preference measure was highly correlated with trust and lacked high discriminant validity.
Abstract
from arXiv · showhide
Explanations given by automation are often used to promote automation adoption. However, it remains unclear whether explanations promote acceptance of automated vehicles (AVs). In this study, we conducted a within-subject experiment in a driving simulator with 32 participants, using four different conditions. The four conditions included: (1) no explanation, (2) explanation given before or (3) after the AV acted and (4) the option for the driver to approve or disapprove the AV's action after hearing the explanation. We examined four AV outcomes: trust, preference for AV, anxiety and mental workload. Results suggest that explanations provided before an AV acted were associated with higher trust in and preference for the AV, but there was no difference in anxiety and workload. These results have important implications for the adoption of AVs.
1. Introduction
The study addresses whether AV explanations promote acceptance and how explanation timing and autonomy affect drivers’ responses. It examines trust, AV preference, anxiety, and mental workload as outcomes.
- Trust is a central challenge because drivers may hesitate to cede complete control of driving to AVs.
- Prior research suggests explanations can facilitate trust in automation, but their effects on AV trust and preference have been mixed.
- The study examines whether AV explanations influence trust, AV preference, anxiety, and mental workload.
- It asks whether explanations improve drivers’ responses and how explanation timing and degree of AV autonomy influence effectiveness.
- Based on Uncertainty Reduction Theory, the authors hypothesize that explanations before action outperform explanations after action across trust, preference, anxiety, and workload.
- They also hypothesize that requiring driver approval before AV action increases trust and preference while decreasing anxiety and increasing workload.
2. Background
The background reviews how explanations and autonomy shape responses to automation and AVs. It highlights inconsistent AV explanation findings and the lack of systematic research on explanation timing.
- Explanations and Automated Vehicles: Explanations expose users to automated systems’ logic and have been linked generally to trust, acceptance, and reduced concern.
- Explanations and Automated Vehicles: AV explanation studies have produced mixed outcomes, with some increasing acceptance or trust and others showing no significant trust improvement.
- Explanations and Automated Vehicles: Before-action explanations have been associated with lower anxiety and higher control, alertness, preference, trust, anthropomorphism, and usability.
- Explanations and Automated Vehicles: Körber et al. found that after-action explanations did not significantly increase AV trust, although they increased perceived understanding of the system and takeover reasons.
- Explanations and Automated Vehicles: No study had systematically explored explanation timing, despite prior comparisons of explanation types.
- Degree of Autonomy: Degree of autonomy concerns whether automation makes decisions and acts independently, and lower or adaptive autonomy has been linked to greater trust.
2.3 Automated Vehicle Explanation Outcomes
The paper evaluates trust, preference, anxiety, and mental workload as outcomes of AV explanations. These measures connect explanation design to acceptance, usability, and drivers’ responses during automated driving.
- Together, the four outcomes support comparisons with prior AV-explanation research and can inform designers and policymakers concerned with AV adoption.
- Trust: Trust is important because it is associated with intentions to use and effective use of automation, while misaligned trust can create problems.
- Trust: Accurate information about automation can help users develop an appropriate level of trust.
- Preference and Anxiety: Anxiety reflects fear, worry, apprehension, or concern about using an AV and is included to assess explanation effectiveness.
- Preference and Anxiety: Preference captures how much someone likes or favors a particular AV technology and complements trust and anxiety measures.
- Mental Workload: Mental workload was included because AV explanations may influence users’ effort during automated driving and potentially affect ease of use.
3. Hypotheses Development
The hypotheses use Uncertainty Reduction Theory to predict how explanations, their timing, and driver approval affect AV responses. The authors expect information and control to shape trust, preference, anxiety, and workload.
- Uncertainty Reduction Theory proposes that people reduce uncertainty through information and communication, with lower uncertainty associated with greater trust.
- The authors predict that explanations, regardless of timing, increase trust and preference while decreasing anxiety and mental workload.
- They predict that before-action explanations outperform after-action explanations because they provide preparation time and reduce the chance of startling reactions.
- They hypothesize that allowing drivers to disapprove an AV action produces the highest trust, preference, and workload and the lowest anxiety.
- The predicted workload increase under driver approval reflects the attention and effort required to decide about AV actions.
- H1, H2, and H3 formalize predictions for explanation presence, timing, and lower-autonomy approval conditions across the four outcomes.
3. Method
The study used a within-subject driving-simulator experiment to examine how explanation timing and AV autonomy affect trust, preference, anxiety, and mental workload. Thirty-two participants experienced four explanation conditions in simulated SAE level 4 driving.
- Apparatus: The simulator represented an SAE level 4 AV that controlled vehicle movement, navigation, and responses to traffic elements without requiring active driver monitoring.Participants engaged automation after beginning each drive and were not asked to retake control.
- Experimental design: The within-subject design manipulated whether explanations occurred before or after AV actions and whether driver approval was required.The four conditions were no explanation, before explanation, after explanation, and permission request.
- Experimental design: Before explanations occurred 7 seconds before AV actions, whereas after explanations occurred within 1 second after actions.In the permission-request condition, the AV explained its upcoming action and requested approval 7 seconds later.
4. Results
Across four explanation conditions, explanation timing affected trust and preference, while anxiety and mental workload did not differ significantly. Before-action explanations generally produced the most favorable outcomes, but explanations overall did not consistently outperform no explanation.
- Trust: Trust rankings were higher for before-action explanation and permission request than for no explanation and after-action explanation.Before-action explanation and permission request did not differ significantly in trust ranking.
- Trust: Trust ratings differed significantly across explanation conditions (F(3,93) = 4.814, p = .008), with before-action explanation producing the highest trust.Before-action explanation exceeded no explanation, after-action explanation, and permission request in post hoc comparisons.
- Trust: Averaged explanation conditions did not significantly differ from no explanation on trust ratings (F(1,31) = .506, p = .482).Thus, the trust advantage depended on explanation timing rather than explanations uniformly increasing trust.
- Workload and anxiety: Mental workload did not significantly differ across conditions (F(3,93) = 2.233, p = .09), although before-action explanation had the lowest mean.Anxiety likewise showed no significant condition differences (F(3,93) = .525, p = .666).
- Autonomy: The authors found no evidence that permission-based autonomy increased preference or mental workload or lowered anxiety relative to no permission.Trust ratings were not significantly higher under the lower-autonomy permission condition.
5. Discussion
The discussion indicates that explanation timing matters: before-action explanations were associated with higher trust and AV preference than no or after-action explanations, while effects on anxiety and mental workload were generally limited. The study also finds mixed support for giving drivers more control and notes that automation level and trust measurement may shape the results.
- Explanation timing: Before-action explanations were associated with higher trust than no- and after-explanation conditions.The authors conclude that merely explaining an action is insufficient; timing before the AV acts matters.
- Explanation timing: Before-action explanations were associated with greater preference for the AV, whereas after-action explanations showed no such effect.This pattern was consistent with prior AV studies and correlated highly with trust.
- Driver outcomes: Explanations did not significantly reduce anxiety, and both explanation conditions were not significantly different from no explanation.The discussion contrasts this result with prior work and suggests that differing automation levels may help explain the discrepancy.
- Degree of autonomy: Giving drivers permission to approve or disapprove AV actions produced little evidence of benefits for trust, preference, anxiety, or mental workload.Participants viewed the permission-seeking AV as less intelligent than the before-explanation AV (p = .041), and reported higher mental workload, although the latter was not significant at .05.
- Measurement and interpretation: Trust rankings differed from attitude-based trust measures, suggesting that how trust is measured may contribute to inconsistent findings.The authors call for further investigation into when and why these measures diverge.
- Theoretical interpretation: The study provides mixed support for using explanations and degree of autonomy to improve AV outcomes, and URT did not explain all overall null effects.The discussion specifically notes that URT explained the advantage of before-explanations but not why greater control lacked broader benefits.
6. Limitations and Future Work
The study’s findings are bounded by simulator realism, overlapping trust and preference measures, aggregate analysis, and limited explanation formats and outcome measures. Future work proposes adaptive interfaces, multimodal explanations, and qualitative measures.
- The driving simulator reduced the risk of unexpected events, which might have influenced participants’ anxiety and perceived safety.
- The preference survey was highly correlated with trust ratings and lacked high discriminant validity.The authors suggest similar meanings between trust and preference may explain this overlap.
- Results were general findings averaged across participants and did not consider individual differences such as desirability of control and personality.
- Future studies could adapt explanation timing and degree of autonomy to drivers’ characteristics.
- Future studies could examine multimodal explanations because the study used auditory explanations בלבד.
- Future studies should include qualitative measures to provide additional insights into the outcomes.
7. Conclusion
The study examined how explanation timing and autonomy degree affect drivers’ responses to automated vehicles. It contributes evidence on explanation timing and autonomy in SAE level 4 driving, while noting that further research is needed.
- The study investigated explanation timing and degree of autonomy in relation to drivers’ trust, AV preference, anxiety, and mental workload.
- The study focused on SAE level 4, where drivers may take their eyes off the road for extended periods during highly automated driving.
- The authors identified and demonstrated the importance of the timing of AV explanations.
- The study extended prior AV-explanation research by incorporating the impacts of degree of autonomy.
- Further research is needed to build on these ideas and provide new insights.
Appendix 1
The appendix presents driving scenarios in which the AV responds to roadway hazards, traffic, police vehicles, construction, and route conditions by changing speed, lane, route, or position.
- The AV rerouted or changed lanes when road construction, roadway obstruction, or a road hazard appeared ahead.
- The AV slowed down for a swerving vehicle or an oversized vehicle blocking the roadway.
- The AV changed lanes to avoid collision with a police vehicle stopped on the shoulder.
- The AV rerouted when heavy traffic was reported ahead.
- The AV pulled over when a police vehicle approached from behind with its siren activated.
Debriefing before training session
Before training, participants were told they would act as drivers of a fully autonomous vehicle that could drive independently and obey traffic laws. They were also instructed about automated-mode constraints and activation procedures.
- Participants were told the vehicle could drive safely on its own in all driving situations and obey traffic laws.
- The vehicle could receive external navigation information and change routes to reach a destination more quickly when appropriate.
- Once automated driving mode was engaged, participants could not control the vehicle and no longer needed to monitor the roadway.
- Participants were instructed to activate automated mode with the steering-wheel button, remove their hands, and respond vocally if prompted.
Debriefing before each drive
Participants were briefed that each drive differed in how the autonomous vehicle presented information about roadway events and actions. After each drive, surveys assessed trust, perceived AV traits, anxiety, and mental workload, followed by rankings of the four information methods.
- Drive procedure: Each study session began with four driving sections, each followed by a survey evaluating the participant’s experience in the autonomous vehicle.The drives differed in how the vehicle presented information about the environment and other vehicles’ actions.
- Survey measures: Anxiety was assessed with adjectives including anxious, fearful, afraid, and uneasy describing how participants felt while the AV drove itself.Participants rated how well each adjective described their feelings during the drive.
- Survey measures: Mental workload was measured with NASA TLX, including mental, physical, and temporal demand, effort, performance, and frustration.The survey asked participants to rate how demanding the task was and how hard they worked to achieve their performance.
- Information conditions: The four information methods were no information, information before events, information after events, and information before events with driver input requested.Participants later ranked these methods by how much they trusted the vehicle.