Data were collected from 55 participants recruited through a third-party agency engaged by the Japan Automobile Research Institute (JARI), where the study took place. All participants reported no prior hands-on experience with Level 2 automation systems, and therefore represented an automation-naïve sample at the time of testing.

The study aimed for approximately 20 participants per educational condition to provide a balanced design and reasonable sensitivity for detecting medium-to-large effects. Participants were randomly assigned to one of three groups: Basic Education (n = 20), capability-focused (Method 1, n = 15), and limitation-focused (Method 2, n = 20). The slight imbalance arose from normal variation in attendance and early withdrawals rather than any planned differences in allocation.

Of the 55 participants in the final sample, 16 self-identified as female and 39 as male (Table 1). Sex distribution was broadly comparable across groups: the Basic group comprised 6 females and 14 males, the capability-focused group 5 females and 10 males, and the limitation-focused group 5 females and 15 males. Age and self-reported annual driving mileage by education group are summarized in Table 2.

All participants provided informed consent prior to the study. The protocol was reviewed and approved by the JARI Research Ethics Committee. No personally identifying data were retained; all datasets were anonymised using numerical identifiers.

The study employed a 3 × 4 mixed factorial designwith Educational Condition (Basic, capability-focused, limitation-focused) as a between-subjects factor and Time (T1, T2, T3, T4) as a within-subjects factor corresponding to four measurement points throughout the session. This structure (Figure 1) enabled examination of both immediate and evolving effects of instructional framing across the automation experience, capturing trust calibration and behavioral outcomes over time.

Educational briefings were designed to examine how different communication strategies shape drivers' understanding of automation and their expectations of system behavior. Each briefing conveyed the same core informational content: the automation supports driving but still requires active human supervision. The briefings differed in instructional emphasis (capability-focused versus limitation-focused framing). The guiding question was: How should information about system capabilities and limitations be conveyed to adjust driver expectations appropriately?

Educational content was delivered through scripted verbal and written briefings that were standardized across participants within each condition. The three conditions held duration, and delivery format constant, while varying the instructional emphasis (capability-focused versus limitation-focused framing) relative to the reference briefing.

Basic education (reference condition): Participants received only the minimum information typically provided in current vehicle manuals or pre-drive explanations. The briefing stated: “The driving automation system you will use today is a partially automated driving system. When the system is activated, the vehicle can drive automatically without pedal or steering operation. However, the driver must monitor the road and surrounding environment. When an alert occurs, you must take over control of the vehicle.” This condition served as a reference point reflecting minimal onboarding information.

Capability-focused education (method 1): This briefing explained how the automation system functions and perceives its surroundings. It described the system's control of steering and speed on highways up to 100 km/h, and how front and rear cameras detect surrounding vehicles and obstacles. Participants were told that system accuracy depends on image quality, which can deteriorate in poor weather. They were informed that the system issues an alert and disengages when reaching its operational limits, although delayed alerts may occur if sensing is impaired. This education aimed to build a general understanding of how the system works and why its performance changes under certain conditions. Visual aids illustrated camera positions and sensor coverage. The full script is provided in Appendix 4.12.2.

Limitation-focused education (method 2): This briefing conveyed the same core message as Method 1 but focused on operating boundaries. It explained that the system controls steering and pedals, functions in normal traffic and weather, and adjusts speed within its lane up to 100 km/h. Drivers were informed that visibility loss in rain or fog could degrade sensor performance and delay alerts. The message emphasized that when the system reaches its limits, it will issue an alert and disengage. Visual aids showed how adverse weather affects sensing. This framing encouraged drivers to monitor conditions and prepare to take control when necessary. The full script is provided in Appendix 4.12.3.

All briefings were delivered by the same experimenter using standardized scripts to ensure consistency. Each lasted approximately 5 min. Participants were unaware that different educational versions existed or that the study examined trust calibration.

These framings were informed by Rasmussen's Skills-Rules-Knowledge framework (Rasmussen, 1983). The capability-focused condition provided conceptual information about how the system perceives the environment and controls speed and steering. The limitation-focused condition emphasized operating boundaries, environmental constraints, and the driver's responsibility to intervene when sensing is impaired. The Basic condition offered only minimal information typical of current onboarding. These distinctions allowed us to test how different types of pre-drive information shape supervision during Level 2 automation.

The experiment was conducted using the Japan Automobile Research Institute's omnidirectional driving simulator, a high-fidelity system featuring a complete vehicle cabin mounted on a motion platform with 360-degree visual projection (Figure 2). The simulator provides realistic vehicle dynamics, including motion cueing for acceleration, braking, and lateral movement, which enhances the ecological validity of driver responses during automation supervision and takeover events.

The simulated environment depicted a two-lane, two-direction, Japanese Expressway divided by a barrier, with moderate traffic density. Traffic consisted of computer-controlled vehicles programmed to maintain realistic speeds and following distances. The Level 2 automation system controlled the simulator vehicle's steering and speed throughout a 15-min driving automation segment, maintaining position in the center lane at approximately 80 km/h and adjusting speed to maintain safe following distance from lead vehicles.

The critical takeover scenario (Figure 3) occurred at the conclusion of the 15-min automated segment. A lead vehicle performed an emergency braking maneuver that exceeded the automation's deceleration capability. The system issued a visual and auditory takeover request requiring the driver to resume manual control and apply braking to avoid collision. This scenario was designed to assess drivers' readiness to intervene during safety-critical events following extended automation supervision.

Each session followed a standardized 80-min protocol designed to capture behavioral, physiological, and self-report data across all phases of the automation experience while ensuring participant comfort. Figure 6 illustrates the complete timeline and timing of trust assessments (T1 to T4).

Upon arrival at the simulator facility, participants provided written informed consent and completed demographic questionnaires. The session then progressed through the following sequence:

T1 (Baseline trust measurement). Participants first rated their initial level of trust in automation before receiving any detailed system information.

Educational intervention and T2 (Post-education trust).

Participants were randomly assigned to one of three instructional conditions (Basic, capability-focused, or limitation-focused; see Section 2.3). The assigned briefing was delivered by the experimenter while participants remained seated in the briefing room. Immediately after the briefing, participants provided a second trust rating (T2) to assess the immediate effect of the educational content.

Simulator familiarization. Participants then moved to the driving simulator and completed a brief manual driving segment to acclimate to the simulator's steering and pedal dynamics and reduce novelty effects.

Practice drive and T3 (Post-practice trust). Participants then engaged the Level 2 automation during a short, non-critical practice segment on a straight motorway section. This phase provided hands-on experience with system operation and the human-machine interface. A third trust rating (T3) was collected immediately afterward.

Eye-tracking calibration. Before the main drive, a 9-point calibration was performed for each participant using the NAC EMR-9 system, following established protocols to ensure accuracy throughout data collection.

Main automated drive. Participants supervised the Level 2 automation for approximately 15 minutes on a simulated highway with moderate traffic. Several non-critical traffic interactions were embedded to sustain ecological validity, and continuous eye-tracking and vehicle control data were recorded. Participants were instructed to monitor the system and remain ready to resume control at any time.

Critical takeover scenario. At the end of the automated segment, the lead vehicle performed sudden braking, triggering a visual and auditory takeover request prompting the driver to resume manual control. Takeover reaction time was recorded automatically by the simulator.

T4 (Post-drive trust) and mind-wandering assessment. Immediately after the takeover event, participants completed the final trust questionnaire (T4) and a retrospective mind-wandering measure. A structured debriefing followed to collect qualitative feedback about the automation experience.

All analyses were performed using standard parametric and non-parametric methods according to data characteristics. Outliers were screened prior to hypothesis testing, and assumption checks were conducted for normality, sphericity, and homogeneity of variance. When Mauchly's test indicated violation of sphericity, Greenhouse-Geisser correction was applied, with adjusted degrees of freedom and epsilon (ε) values reported. Levene's tests verified that the assumption of equal variances was met for these between-group comparisons.

This section addresses RQ1 (trust dynamics) and evaluates H1.1-H1.2.

Findings from the repeated-measures ANOVA (Table 3) address H1.1 (main effect of Time) and H1.2 (Time × Group), both involving the within-subject factor Time. Outcomes from the between-subjects tests (Table 4) address the between-group component of H1.2 (main effect of Group). Levene's tests verified that the assumption of equal variances was met for these between-group comparisons.

This section addresses RQ2 (mind-wandering) and evaluates H2, followed by RQ5 (relationships) for H5.1.

After identifying the general pattern of trust development, we next explored whether the educational interventions affected mind-wandering during automation supervision.

This section addresses RQ3 (visual scanning) and evaluates H3, then addresses RQ5 for H5.2, and finally addresses RQ4 (takeover performance) for H4.

Trust increased across all educational conditions as participants gained exposure to Level 2 automation, supporting H1.1. This pattern is consistent with earlier evidence that direct interaction with automation gradually aligns trust with actual system performance, even when users begin with different initial information (Beggiato and Krems, 2013; Lee and See, 2004; Hoff and Bashir, 2015). Neither capability-focused nor limitation-focused framing appeared to alter this trajectory (H1.2), suggesting that direct exposure to the system's behavior outweighed prior instruction.

This pattern aligns with the CAUSE model (Rowan, 1991), which describes effective communication as a sequential process: establishing credibility, raising awareness, promoting understanding, achieving satisfaction, and supporting enactment. The brief instructional framings in this study were unlikely to complete this process; instead, participants appeared to calibrate trust through their direct observation of system reliability. In this context, performance feedback rather than message framing likely served as the main influence on trust development.

The positive correlation between trust growth and mind-wandering (H3) indicated a potential tension between confidence in the system and supervisory engagement. Participants who showed greater increases in trust also showed lower values on monitoring-related indicators, consistent with a trust-associated reduction in supervisory engagement. This pattern is consistent with Malleable Attentional Resources Theory (MART) (Young and Stanton, 2002; Warm et al., 2008), which proposes that attentional investment contracts under low task demand. The relationship was observed across educational conditions, suggesting that reduced monitoring may emerge generally as drivers gain experience with stable automation.

Although capability-focused education did not substantially alter trust trajectories, it appeared to yield two modest benefits (H8): greater attention to the HMI and shorter system-recognized reaction times during takeover. Participants who received knowledge-based (capability-focused) instruction monitored automation status displays more frequently and responded more promptly when manual control was required. These findings suggest that a conceptual understanding of system operation may have refined monitoring selectivity and readiness for authority transfer.

Within Endsley's situation-awareness framework (Endsley, 1995b), these improvements may reflect an increased ability to interpret what automation feedback signals about system state. Capability-focused instruction might have rendered HMI cues more meaningful and diagnostic, enabling participants to interpret them more efficiently without compromising roadway monitoring. The faster SRRT further suggests somewhat smoother coordination during manual resumption.

Hence, brief conceptual instruction may provide targeted benefits by sharpening monitoring priorities and readiness, even though it does not seem to mitigate trust-related vigilance loss. These effects illustrate how education and exposure can influence distinct aspects of supervision: exposure shaping trust and perceived monitoring demand, and education fine-tuning particular monitoring behaviors.

Rule-based (limitation-oriented) education showed little measurable influence on trust, attention, or takeover performance. Several factors may explain this outcome. One possible interpretation is that emphasizing system limits did not translate into actionable monitoring strategies during a stable scenario, which may have limited observable behavioral effects. The instruction listed boundary conditions such as poor weather or degraded lane markings but offered little procedural guidance on how to detect or respond to them. Without actionable cues, participants had limited opportunity to apply these rules during the stable highway scenario.

The motivational tone may also have contributed. Emphasizing system fallibility might have led drivers to expect that failures would occur only under specific conditions, such as poor weather or unclear lane markings. Because the simulated drive presented no such cues, participants may have focused less on continuous monitoring and more on waiting for visible signs of malfunction. This conditional form of supervision differs from the general vigilance loss associated with growing trust, yet it can be viewed as another expression of trust miscalibration, where expectations about when the system might fail are narrower than its actual range of vulnerabilities (Parasuraman and Riley, 1997). Moreover, the drive remained within the automation's operational domain, so the limitation messages were not especially relevant. Under these conditions, participants may have regarded the Limitation-focused briefing as less meaningful, which could have limited its behavioral impact.

The Skills-Rules-Knowledge (SRK) framework (Rasmussen, 1983) provides a useful lens for understanding these differences. Limitation-focused instruction depends on stored “if-then” procedures, which apply only when matching cues arise. In this study, such cues were absent, and the learned rules were therefore unlikely to be triggered. Capability-focused education, by contrast, fosters conceptual models that apply across diverse conditions. Understanding how the automation senses and acts may have improved the perceived relevance of HMI information and facilitated faster recognition of control transitions. In Rasmussen's framework, capability-focused processing is required when operators face unfamiliar or unexpected situations for which no stored rules apply. Such conditions are analogous to automation takeovers, where drivers must interpret system behavior and re-establish manual control.

These findings can be interpreted within the situation-awareness discrepancy model, δ = α−α′, introduced in earlier work (Chouchane et al., 2022). Within this model, trust is likely to shape perceived monitoring demand (α). As trust increases, drivers may judge that less active supervision is necessary. This change reflects heightened confidence in automation reliability rather than an actual reduction in the monitoring required. The observed association between higher trust and mind-wandering supports this interpretation, indicating that lower perceived demand can promote attentional drift and weaker supervisory engagement.

Educational interventions, by contrast, may affect achieved monitoring (α′) through partially independent pathways. Capability-focused education seemed to enhance certain components of monitoring, such as HMI attention and takeover readiness. These effects represent localized gains in α′, without offsetting the overall decline in perceived demand. Consequently, brief instruction may reinforce specific monitoring habits but is unlikely to counter the wider vigilance contraction associated with rising trust.

Malleable Attentional Resources Theory (MART) (Young and Stanton, 2002; Warm et al., 2008) provides a complementary explanation for this contraction. Under stable, low-demand automation, attentional resources tend to narrow, leading to less structured and less frequent scanning between key visual areas. The observed correlation between mind-wandering and reduced scanning organization (lower SSI and RRC) aligns with this mechanism. The CAUSE model (Rowan, 1991) further clarifies why instructional messages had limited effect: effective risk communication depends on credible and sequenced engagement rather than isolated briefings. Finally, performance-based trust models (Lee and See, 2004; Hoff and Bashir, 2015) help explain why direct exposure to automation performance dominated trust development.

Taken together, these frameworks suggest that trust most plausibly influences perceived monitoring demand, while education can refine selective monitoring behaviors. The combination of rising trust and stable system performance therefore creates conditions in which drivers monitor less actively even though objective supervisory requirements remain unchanged.

The results suggest that brief pre-drive education has limited influence on trust trajectories, which develop mainly through direct exposure to automation performance. Verbal or written instructions, though valuable for setting initial expectations, appear insufficient to shape how trust evolves during actual use. Nonetheless, such education can enhance comprehension and promote more focused monitoring.

The modest benefits of capability-focused instruction indicate that more extensive, scenario-based training could produce broader effects. Programmes combining conceptual explanation with guided practice might help drivers encounter automation in varied conditions and near its operational limits, allowing better calibration of both trust and monitoring strategies. Real-time feedback from instructors could reinforce how and when to intervene, supporting balanced supervision.

The improvement in HMI monitoring following capability-focused instruction also points to the importance of interface transparency. Displays that provide meaningful cues, such as sensor coverage or automation confidence, can enhance monitoring, provided the information is interpretable and not overwhelming. Education and interface design may therefore work best when developed in tandem.

Finally, the link between scanning structure and mind-wandering highlights opportunities for driver-monitoring systems. Beyond tracking gaze-on-road duration, systems might assess how organized scanning patterns remain over time. Identifying reduced mirror-to-road transitions or increased randomness could enable early alerts before performance declines.

Several factors limit interpretation and generalisability. First, between-group sample sizes were modest (n = 15–20 per educational condition), which limits statistical power to detect small effects and interactions; null findings for group differences should therefore be interpreted cautiously. Second, the educational interventions were brief and delivered once; repeated or interactive training might yield stronger and more durable effects. Third, the simulated drive involved only one takeover event after 15 minutes of automation in stable conditions. Real driving includes greater variability in road, traffic, and weather conditions, which may influence both trust and attention. Fourth, participants were inexperienced with hands-on Level 2 automation use, and the findings therefore reflect early exposure rather than longer-term adaptation; future studies should examine whether educational benefits or trust-attention dynamics persist over extended use. Fifth, mind-wandering was measured retrospectively, which could have underestimated momentary fluctuations; incorporating concurrent or physiological measures would strengthen future assessments. Finally, the instructional materials represented one implementation of capability-focused and limitation-focused framing. Although the briefings differed in emphasis, we did not include a direct manipulation check to verify whether participants internalized the framings as intended. Given the shared informational content across the briefings, it is possible that overlap attenuated between-group differences in trust and gaze metrics, and effects should therefore be interpreted cautiously. Variations in content, delivery medium, or realism, as well as brief post-briefing checks of capability versus boundary understanding, would strengthen future validation of framing-based interventions.

Future research should examine how educational and exposure factors interact over time and across contexts. Longitudinal studies could track trust and monitoring behavior over extended periods of real-world driving. Scenario-based training could identify how conceptual understanding and hands-on exposure combine to calibrate trust and attention effectively. Research in diverse driving environments, including urban and adverse-weather conditions, would clarify when limitation-focused education becomes most relevant. Finally, combining educational interventions with adaptive HMIs and monitoring systems could help sustain driver engagement.

Brief pre-drive education modulated specific monitoring behaviors but did not influence the trajectory of trust formation during Level 2 automation. Trust increased over time across all groups and was not affected by instructional framing, consistent with experience-driven adjustment to observed system performance. Capability-focused education was associated with greater attention to automation feedback and faster takeover responses compared with limitation-focused education, indicating that conceptual information may refine how drivers monitor system status without altering overall trust development.

Across participants, increases in trust were accompanied by patterns consistent with reduced structured monitoring, suggesting a potential trade-off between growing confidence and sustained supervisory engagement during routine automation. More structured gaze organization was associated with lower self-reported mind-wandering, supporting the interpretation of gaze metrics as behavioral proxies for attentional allocation rather than direct measures of attention.

These findings suggest the potential value of integrated approaches combining extended experiential training, transparent interface design, and gaze-based monitoring systems to support sustained supervisory engagement in partially automated driving.