Available via license: CC BY 4.0
Content may be subject to copyright.
A Transparency Paradox? Investigating the Impact of Explanation
Specificity and Autonomous Vehicle Perceptual Inaccuracies on
Passengers
Daniel Omeizaa,∗,Raunak Bhattacharyyaa,c,1,Marina Jirotkaa,Nick Hawesaand Lars Kunzea,b
aUniversity of Oxford, Oxford, United Kingdom
bUniversity of the West of England, Bristol, United Kingdom
cIndian Institute of Technology, Delhi, India
ARTICLE INFO
Keywords:
Explanations
Transparency
Autonomous Driving,
Perceived Safety
Visual Attention
ABSTRACT
Transparency in automated systems could be afforded through the provision of intelligible expla-
nations. While transparency is desirable, might it lead to catastrophic outcomes (such as anxiety),
that could outweigh its benefits? It’s quite unclear how the specificity of explanations (level of
transparency) influences recipients, especially in autonomous driving (AD). In this work, we
examined the effects of transparency mediated through varying levels of explanation specificity
in AD. We first extended a data-driven explainer model by adding a rule-based option for
explanation generation in AD, and then conducted a within-subject lab study with 39 participants
in an immersive driving simulator to study the effect of the resulting explanations. Specifically,
our investigation focused on: (1) how different types of explanations (specific vs. abstract)
affect passengers’ perceived safety, anxiety, and willingness to take control of the vehicle when
the vehicle perception system makes erroneous predictions; and (2) the relationship between
passengers’ behavioural cues and their feelings during the autonomous drives. Our findings
showed that passengers felt safer with specific explanations when the vehicle’s perception system
had minimal errors, while abstract explanations that hid perception errors led to lower feelings
of safety. Anxiety levels increased when specific explanations revealed perception system errors
(high transparency). We found no significant link between passengers’ visual patterns and their
anxiety levels. Our study suggests that passengers prefer clear and specific explanations (high
transparency) when they originate from autonomous vehicles (AVs) with optimal perceptual
accuracy.
1. Introduction
The automotive industry is witnessing an increasing level of development in the past decades, from manufacturing
manually operated vehicles to manufacturing vehicles with a high level of automation. Despite these technological
strides, accidents involving AVs continue to undermine public trust [59,5,38,62,44]. As highly automated vehicles
make high-stake decisions that can significantly affect end-users, the vehicles should explain or justify their decisions
to meet set transparency guidelines or regulations, e.g., GDPR Article 12 [63] and the [31].
Accompanying AVs’ driving decisions with natural language explanations is one promising approach for better
vehicle transparency [51,26,34]. This transparency, obtained through intelligible explanations, can help to reassure
passengers of safety and also assist them in effectively calibrating their trust in an AV [32]. The specificity level of
∗Corresponding author
danielomeiza@robots.ox.ac.uk (D. Omeiza); raunak@robots.ox.ac.uk (R. Bhattacharyya); marina.jirotka@ox.ac.uk (M.
Jirotka); nickh@robots.ox.ac.uk (N. Hawes)
ORCID (s): 0000-0002-1232-346X (D. Omeiza)
1This author was with the University of Oxford at the time this work was done. This author’s affiliation is now the Indian Institute of Technology,
Delhi
D. Omeiza et al.: Preprint submitted to Elsevier Page 1 of 28
arXiv:2408.08785v1 [cs.HC] 16 Aug 2024
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
Figure 1: Driving simulation setup for the study. The setup included a VR headset, steering wheel, brake and acceleration pedals,
screen, and arcade seat. The screen shows a pedestrian crossing at a crosswalk.
explanations is, however, an important factor in achieving the aforementioned benefits. In real-world deployments,
AVs may not always achieve perfect scene understanding due to the limitations of their perception systems. Depending
on the specificity level, explanations might reflect these flaws even when they are inconsequential. Informing operators
about these inherent imperfections could enhance safety by helping them recognise when to remotely take control of the
vehicle [36] and it would equally help developers optimise for higher accuracy. While this transparency is helpful for
these groups, it remains uncertain whether in-vehicle passengers would prefer high level transparency (if at all useful)
that reveals such minor errors (such as mistaking a van for a bus). This is what we refer to the transparency paradox
in this paper. It is therefore important to determine the appropriate level of transparency for in-vehicle passengers,
mediated through the specificity of the explanations.
Furthermore, as passengers are likely to engage in other activities during their ride, relying solely on visual cues
to communicate awareness may be ineffective when passengers’ attention is desired. Hence, auditory and vibrotactile
feedback[37] are also necessary to ensure passengers are adequately informed and can respond appropriately in critical
situations.
In this study, we investigate the effects of explanation specificity on AV passengers’ perceived safety, feeling of
anxiety, and desire to takeover control. We test two levels of explanation specificity: abstract and specific. We use the
term abstract to describe the provision of vague auditory explanations that conceal details about a driving situation
(including perception system errors). In contrast, specific refers to the provision of very detailed and fine-grained
explanations about a situation. We found non-significant effect of explanation specificity on perceived safety. However,
we noticed a significant influence with respect to the amount of perception errors unveiled in the explanations (i.e.,
between Specific (5) — Specific (50). Perception of safety dropped in the Specific (50) scenario compared to Specific
(5). The feeling of anxiety also significantly increased with the increase in perception errors. There was a significant
D. Omeiza et al.: Preprint submitted to Elsevier Page 2 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
effect of explanations specificity on takeover feeling (Abstract — Specific (50)), but no significant difference noticed
between Specific (5) and Specific (50). Lastly, no strong correlation was observed between the participants’ behavioural
cues and the aforementioned factors.
1.1. Contribution Statement
Overall, this research makes three contributions to the fields of explainable autonomous driving and human-
machine interaction:
1. It sets out a new case study of explanation specificity in the presence of perception systems errors in the
autonomous driving context;
2. It provides an enhanced interpretable technique for generating textual and auditory natural language explanations
for AV navigation actions;
3. It reveals experimental findings on whether high AV transparency, though critical to other stakeholders, is helpful
to AV passengers.
2. Background
This section provides a background on explanations for automated decisions and in autonomous driving from the
literature.
2.1. Explanations for Automated Decisions
The concept of explanations has been studied extensively by a multidisciplinary group of scholars, ranging from
Philosophy to Psychology. Each adopts an idiosyncratic lens, characterising explanations in terms relevant to the goals
of their respective disciplines. Therefore, to guide our study, we align with a definition proposed in an extensive survey
of explanations in autonomous driving [51]: explanations are a piece of information presented in an intelligible way
as a reason or part of a reason for an outcome, event or an effect in text, speech, or visual forms.
Recent efforts around explanations have been mainly channelled towards complex AI systems (explainable
AI [25,1,16]) to understand and communicate the reasons for the systems’ decisions. Techniques developed for
this purpose fall under different categories based on their mode of operation. Some are model-specific [7] in that
they investigate the underlying AI algorithm in detail to support debugging tasks [43]. These types of explainers are
intrinsic and model specific meaning that they are inherently coupled with the underlying algorithm. Other explanation
approaches are classified as model-agnostic as they can assess the properties of an output independent of the algorithms
used to realise the output [55,41]. These explainers aim to help enhance user’s knowledge of a system and potentially
foster trust.
D. Omeiza et al.: Preprint submitted to Elsevier Page 3 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
Explanations must possess certain properties to be effective to their intended recipients. Mittelstadt et al. [47]
argued that the risk of conflicts in communicating explanations when the explainer (explanation provider) and the
explainee (explanation recipient) have different motives may be mitigated through social, selective, and contrastive
explanations. Social in the sense that the explanation process involves different parties and the explainer is able to
model the expectations of the explainee. The explanation is selective if it can select explanations from several competing
hypotheses. It is considered contrastive if it can differentiate the properties of two competing hypotheses. Kment [33]
further emphasised the value of counterfactual explanations in enhancing understanding. These explanations describe
how changes in input can lead to a shift from one fact to a competing hypothesis (foil) [46]. Explanations should be
intelligible [50] and strike a delicate balance between providing sufficient detail and respecting the cognitive capacity
of the explainee. This balance presents a significant challenge, as accurately gauging an individual’s real-time cognitive
capacity remains difficult.
Different methodologies have been adopted in the XAI literature. Wang et al. [64] categorised these research
methodologies into three groups: First, the existence of unvalidated guidelines for the design and evaluations of
explanations was highlighted. The authors claimed that these kinds of guidelines are based on authors’ experiences
with no further substantial justification. Second, researchers suggested (in [66]) that understanding users’ requirements
is helpful in XAI research. It is on this premise that previous research on explanation design has been thought to be
empirically derived. This type of XAI research elicits explanation requirements from user surveys to determine the
right explanation for a use case with explanation interfaces. Third, some explanation design methods are derived from
psychological constructs from formal theories in the academic literature. Some of these methods (e.g., in [30]) draw
on theories from cognitive psychology to inform explanation design for explanation frameworks. Our work is heavily
grounded on empirical studies.
2.2. Explanations in Autonomous Driving
Explanations have been found useful in enhancing user experience [56], trust [34,26], and improved situational
awareness [49,40] in automated driving. Recent works have explored human factors in the application of explainable
AI in AD. For instance, in [50,48], a socio-technical approach to explainability was proposed. An interpretable
representation and algorithms for explanations based on a combination of actions, observations, and road rules were
designed. Regarding explanations depths, the notion that explanations with higher levels of abstraction and correctness
are superior has been argued in the literature [6,24]. Additionally, Ramon et al. (2021) argued that the specificity
of explanations should be tailored to the application context, noting that low-level specificity is often preferred by
individuals with a more deliberative cognitive style.
D. Omeiza et al.: Preprint submitted to Elsevier Page 4 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
In this paper, the term explanation specificity is used to refer to two specificity levels of explanations, abstract
(low transparency) and specific (high transparency). Explanations can be used to convey different information in AD,
e.g., vehicle uncertainties and intentions, and communicated through different modalities. For example, Kunze et
al. [36] conveyed visual uncertainties with multiple levels to operators using heartbeat animation. This information
helped operators calibrate their trust in automation and increased their situation awareness. Similarly, Kunze et al. [37]
used peripheral awareness display to communicate uncertainties to alleviate the workload on operators simultaneously
observing the instrument cluster and focusing on the road. This uncertainty communication style decreased workload
and improved takeover performance. In addition, the effects of augmented reality visualisation methods on trust,
situation awareness, and cognitive load have been investigated in previous studies using semantic segmentation [11],
scene detection and prediction [12], and pedestrian detection and prediction [10]. These deep vision-based techniques
applied to automated driving videos and rendered in augmented reality mode were a way of calling the attention of
operators to risky traffic agents in order to enhance safety. While under-explored, auditory means of communicating
explanations are important to calling in-vehicle participants’ attention to critical situations in AD. We thus used
an auditory communication style in this study to convey explanations to passengers. Some existing works around
human-machine interaction [40] have leveraged theoretical models (e.g., mental and situational models [21]) to study
explanations. We based our work on behavioural cues and subjective feedback from participants while drawing
connections to such existing works.
2.3. Research Questions
From the preceding literature review, we find the need to gain a deeper understanding of the effects of transparency,
brought about by natural language explanations of varying specificity, especially under imperfect AV perception
systems.
1. Given varying levels of perception system errors, how do natural language explanations influence passengers’
perceived safety?
•H1.1 - Perceived Safety. Low transparency yields a higher perception of safety in an AV with perception
system errors. We hypothesise that passengers feel safer in a low transparency AV, despite receiving
abstract explanations. While individuals often seek the truth, many prefer information that aligns with
their expectations [27]. Consequently, specific explanations may reveal perception system errors that
contradict passenger expectations. Additionally, research has shown that placebo explanations can have
similar positive effects on people as real explanations [18].
•H1.2 - Feeling of Anxiety. Passengers’ feeling of anxiety increases with increasing perception system
errors in a highly transparent AV. We posit that there is a connection between perceived safety and the
D. Omeiza et al.: Preprint submitted to Elsevier Page 5 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
feeling of anxiety [13,53]. Therefore, explanations that frequently reference misclassified actors are likely
to create a sense of insecurity, leading to increased anxiety.
•H1.3 - Takeover Feeling. In highly transparent AVs, passengers are more likely to develop the feeling to
take over navigation control from the AV with higher perception system errors. Although passengers are
not able to take control in this study, we anticipated that they might nurse the thought to do so if they
repeatedly received illogical explanations from the AV.
2. Do passengers’ behavioural cues correlate with their feelings?
•H2.1 - Visual Feedback Visual feedback from participants correlates with their feeling of anxiety.
Individuals with the feeling of anxiety might be usually hyper-aroused and sensitive to environmental
stimuli. They may have difficulties concentrating, performing tasks efficiently, and inhibiting unwanted
thoughts and distractions [28,8]. Participants’ fixation points and saccades should correlate with anxiety.
3. Passenger Study
In this section, we describe the participants’ demographic, experiment apparatus setup, experiment design, and the
procedure of the experiment. The necessary approval to conduct the study was obtained from our University’s Research
Ethics Committee.
3.1. Participants
We conducted a power analysis to estimate the number of subjects required for the study. Afterward, calls for
participants were placed on various online platforms, such as the callforparticipants platform, university mailing
groups, university Slack channels, the research group website, and social media to recruit subjects. Upon screening, the
final sample consisted of 𝑁= 39 participants (28 male, 11 female) ranging in age from 18 to 59 years. The participants
comprised students, university employees, and members of the callforparticipants platform. Although prior driving
experiences were not required, 28 (71.79 %) of the participants were licensed drivers. Only 2 of the 39 participants
(5.13%) had experience with autonomous drives, however, in a research context. 6 (15.38%) of the participants had
used a virtual reality headset for a driving game or driving experiment in the past.
3.2. Apparatus
3.2.1. Hardware
The hardware setup is shown in Fig. 1. We conducted the experiment in a driving simulator that comprised a GTR
arcade seat, Logitech G29 steering wheel with force feedback, turn signal paddles, brake and accelerator pedals, and
an ultra-wide LG curved screen to display the experiment. A state-of-the-art virtual reality (VR) headset (with an
immersive 360◦FoV and an eye tracker) was also used to provide an immersive experience and high visual fidelity.
D. Omeiza et al.: Preprint submitted to Elsevier Page 6 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
Table 1
Description of a subset of the events (5 out of 9) and corresponding explanations provided during the study. Observations and causal
explanations are announced to passengers. AV’s action (text in red), other agent’s class & action (text in blue), and the agent’s
location (text in green) are determined by the explainer algorithm described in Algorithm 1.
Event Description Observation Causal Explanation
FollowLeadingVehicle AV follows a leading actor. At
some point, the leading actor
slows down and finally stops. The
AV has to react accordingly to
avoid a collision.
vehicle ahead on my lane.Stopping because cyclist
stopped on my lane.
VehicleTurning AV takes a right or a left turn
from an intersection where an
actor suddenly drives into the way
of the AV, AV stops accordingly.
After some time, the actor clears
the road, AV continues driving.
motorbike crossing my
lane.
Stopping because motor-
bike is crossing my lane.
LaneChangeObstacle AV follows a leading actor, and
at some point, the leading actor
decelerates. The AV reacts ac-
cordingly by indicating and then
changing lanes.
vehicle ahead on my lane.Changing lane to the
[right/left] because vehi-
cle stopped on my lane.
StopSignalNoActor No actor ahead of the AV at a
signalised intersection with a red
traffic signal. AV decelerates and
stops.
red traffic light ahead on
my lane.
Stopping because traffic
light is red on my lane.
MovSignalNoActor No actor ahead of the AV. AV
starts moving from a stop state
at a signalised junction or inter-
section.
None Moving because traffic
light is green on my lane.
Figure 2: High-level architecture of our simulation software. DReyeVR uses Unreal engine and extends CARLA simulator, which also
builds on Unreal engine. DReyeVR extends CARLA by adding VR functionalities, vehicular and ambience sounds, eye tracker data
logging, and additional sensors, among others. Our explainer model, which is both rule-based and data-driven, receives ground truth
data from CARLA or DReyeVR and generates explanations for predicted actions. The post-processing script allowed us to modify the
generated explanations as we desire.
3.2.2. Driving Software
Software architecture is illustrated in ??. We adapted the DReyeVR [57], an open-source VR-based driving
simulation platform for behavioural and interaction research involving human drivers. DReyeVR was built atop
CARLA [17], an open-source driving simulator for AD and Unreal Engine 4. DReyeVR provides a very realistic
D. Omeiza et al.: Preprint submitted to Elsevier Page 7 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
experience with naturalistic visuals (e.g., in-vehicle mirrors) and auditory (e.g. vehicular and ambient sounds)
interfaces allowing for an ecologically valid setup. It also provides an experimental monitoring and logging system to
record and replay scenarios, as well as a sign-based navigation system.
3.2.3. Explainer Software
As shown in Fig. 2, we developed an explainer system based on previous work [48]. This system utilises a tree-based
model trained on an annotated AV driving dataset that we collected in a prior project. While the original algorithm
in [48] is primarily data-driven, we incorporated a rule-based technique to serve as a fallback when the data-driven
method fails or makes an incorrect ego action prediction. The data-driven method employs a trained tree-based model
to predict and generate explanations from detections obtained from CARLA, a driving simulator. In contrast, the
rule-based approach relies on CARLA’s ground truth data and follows predefined rules to determine which agents to
reference in the explanations. By comparing predictions from the data-driven method with ground truth observations
from CARLA, we can identify incorrect predictions. This enhanced explainer system, combining both data-driven
and rule-based approaches, was used to generate preliminary explanations for our created scenarios. Wintersberger
et al. (2020) suggested types of traffic elements to be included in visual explanations based on user preferences. Our
proposed explainer, however, selects traffic elements deemed important (feature importance [2]) by the driving model
for its decisions (see Algorithm 1).
We performed post-processing operations on the generated explanations, including fine-tuning some of the content
and adjusting timestamps to ensure the explanations were delivered at the appropriate moments.
Algorithm 1: Intelligible Explanation Generation
Input: tree model for ego’s action prediction, input vector 𝐗describing ego’s environment
Output: intelligible auditory explanation
1Select a representative tree 𝑚∈from tree model .
2Predict action 𝑦∈given 𝐗.
3Compare prediction 𝑦with CARLA ground truth 𝑦𝐺𝑇 .
4if 𝑦=𝑦𝐺𝑇 then
5Trace the decision path 𝑦for the prediction 𝑦in tree 𝑚.
6Compute the importance score 𝐼(𝐗𝑖)of the attributes 𝐗𝑖in each node 𝑖along the decision path 𝑦.
7Select attributes 𝐗𝑖with importance scores 𝐼(𝐗𝑖)≥𝑘, where 𝑘is a predefined threshold.
8Merge the conditions/inequalities in the selected attributes 𝐗𝑖.
9Translate merged attributes 𝐗𝑖to natural language following the template in Tab. 1.
10 else
11 Use CARLA ground truth information 𝑦𝐺𝑇 and predefined rules to generate explanation following the
template in Tab. 1.
D. Omeiza et al.: Preprint submitted to Elsevier Page 8 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
Table 2
Description of Scenarios: The Independent Variables
Scenario Specificity/Transparency Perception Errors (%) Scenario Length Examples of Events Environment
Abstract Low 0 ∼ 4 minutes All events from Table 1 CARLA Town10HD
Specific (5) High 5 ∼ 4 minutes All events from Table 1 CARLA Town10HD
Specific (50) High 50 ∼4 minutes All events from Table 1 CARLA Town10HD
4. Experiment Design
Before the start of the trials, participants were asked to manually drive a vehicle for about two minutes in CARLA
Town03—a complex town, with a 5-lane junction, a roundabout, unevenness, and a tunnel. 30 vehicles and 10
pedestrians were spawned in this town. This preliminary drive aimed to familiarise participants with the driving
simulation environment and to allow them to experience manual driving within the simulation.
We employed a within-subject design due to the limited sample size, which was insufficient for a between-subject
study. Additionally, this design helped mitigate the potential co-founding factor of between-individual differences in a
between-subject design.
4.1. Independent Variable
Combinations of transparency level (low and high) and AV perception errors (low and high) were done to obtain
the independent variable Scenarios. The first scenario (Abstract scenario) comprises abstract explanations indicating
low transparency and an undefined amount of perception system errors. The second scenario (Specific(5) scenario)
comprises specific explanations indicating high transparency and 5% amount of perception system errors indicating low
error degree. The third scenario (Specific(50) scenario) comprises specific explanations indicating high transparency
and 50% amount of perception system errors indicating high error degree. The driving events that made up the different
Figure 3: Scenario routes. Red: Abstract, Green: Specific(5), Blue: Specific(50). Each route is a loop and overlaps with
others at some points.
scenarios were carefully designed to include different driving conditions that are obtainable in the real world (See
Table 1). Note that scenarios The scenario routes are shown in Figure 3.
D. Omeiza et al.: Preprint submitted to Elsevier Page 9 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
i. Abstract A scenario in CARLA Town10HD, which is about 4 minutes long (330 secs). Town10HD is an urban city
environment with different infrastructures, such as an avenue or promenade, and realistic textures. Driving conditions
are a combination of the events in Table 1. The perception system in this scenario might contain some errors, but
the explanations provided in this scenario were post-processed to always provide surface information which is vague
enough to conceal perception errors. The rules governing explanations for this scenario were:
•all traffic lights are referred to as ‘traffic sign’ without specifying the state (e.g., red, green, amber, off) of the
traffic light;
•pedestrians are referred to as ‘road users’;
•all non-human moving actors are referred to as ‘vehicle’. This includes cycles, motorbikes, cars, etc.
An example explanation is ‘stopping because of the traffic sign on my lane’. This obfuscates the type and colour of the
traffic sign.
ii. Specific(5) A scenario in CARLA Town10HD, which was about 4 minutes in length (256 seconds). Driving
conditions in this scenario were a combination of the events in Table 1. The explanations generated in this scenario were
specific and detailed, exposing all errors. The perception system of the AV in this scenario was about 5% inaccurate.
This error value was estimated following the dynamic traffic agent classification model and confusion matrix provided
by [4] and the traffic light classification model and confusion matrix by [45]. We were only interested in the confusion
matrices (and not the models). The confusion matrices helped us to systematically introduce the 5% perception system
errors during the post-processing stage of the explanations. In this scenario, the 5% error resulted in one explanation
(1 out of the 22) being erroneous as the explanation exposed the misclassification errors from the perception system.
An example of an erroneous explanation is: ‘van ahead on my lane’. Here, a car was misclassified as a van.
iii. Specific(50) A scenario in CARLA Town10HD, which was 4 minutes in length (274 seconds). Driving conditions
were a combination of the events in Table 1. The explanations generated in this scenario were as fine-grained/specific
and detailed as those in the Specific(5) scenario. The perception system error of the AV in scenario Specific(5)
was significantly noised to reach a reduced accuracy of 50%. We assumed that this reduction in accuracy might be
sufficient to influence peoples’ behaviour. Therefore, half of the explanations in this scenario (12 out of 24) reflected
misclassification of actors or actor states. An example of an erroneous explanation is ‘moving because traffic light is
switched off on my lane’. In this case, the perception system failed to identify a green light accurately.
Note that all three scenarios were designed so that the AV perception errors were insignificant to the AV’s navigation
actions. Hence, the AV respected all road rules and avoided collisions. This was important as the state-of-the-art AVs
would likely not make obvious navigation errors. Moreover, we were interested in the effects of the awareness of
D. Omeiza et al.: Preprint submitted to Elsevier Page 10 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
inconsequential perceptual errors in AVs. Hence, it was necessary to introduce artificial errors of varying degrees
(low and high). The non-influence of AV perception errors on navigation control also helped to avoid the confounding
factors of route navigation problems. Further, we counterbalanced the routes across scenarios. That is, the AV’s route
was different in each scenario. This design decision was made to reduce carry-over effects on the participants. With
this setup, the scenarios were still comparable as they were all within the same town, and the routes shared similar
features. Each scenario also had a balanced combination of the events listed in Table 1. In all the scenarios, the AV
maintained a speed below 30𝑚𝑝ℎ, the recommended speed limit in urban areas in the UK. See Figure 4for sample
scenes from each scenario and their corresponding explanations.
4.2. Dependent Variables
There were six dependent variables: Perceived Safety, Feeling of Anxiety, Takeover Feeling, Fixation Divergence,
Saccade Difference, and Button Presses. These variables were categorised into two (psychological factors and
behavioural cues) for easy analysis and reporting.
Psychological Factors These factors include Perceived Safety, Feeling of Anxiety, and Takeover Feeling. They
were mainly measured using items from the Autonomous Vehicle Acceptance Model Questionnaire (AVAM) [29].
AVAM is a user acceptance model for autonomous vehicles, adapted from existing user acceptance models for generic
technologies. It comprises a 26-item questionnaire on a 7-point Likert scale, developed after a survey conducted to
evaluate six different autonomy scenarios.
Items 24—26 were used to assess the Perceived Safety factor, while items 19—21 were used to assess the Feeling of
Anxiety factor. Similar to [56], we introduced a new item to assess participants’ feelings to takeover navigation control
from the AV during the ride (Takeover Feeling). Specifically, participants were asked to rate the statement ‘During the
ride, I had the feeling to take over control from the vehicle’ on a 7-point Likert scale. Actual navigation takeover by
participants was not permitted because we wanted to be able to control the entire experiment and have all participants
experience the same scenarios. Moreover, we were dealing with L4 automation. Though participants were not expected
to drive or take over control, they might have nursed the thought to do so. This is what the Takeover Feeling variable
measures.
We added a free-response question related to explanations with the aim of obtaining qualitative data for triangu-
lating quantitative results. Participants were asked the following question: ‘What is your thought on the explanations
provided by the vehicle, e.g., made you less/more anxious, safe, feeling to take over control?’. We refer to the resulting
questionnaire as the APT Questionnaire (i.e., A-Anxiety, P-Perceived Safety, T-Takeover Feeling).
D. Omeiza et al.: Preprint submitted to Elsevier Page 11 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
(a) Abstract - Observation: ‘vehicle is
crossing on my lane’
(b) Specific(5) - Observation: ‘motor
bike ahead on my lane’
(c) Specific(50) - Causal Explanation:
‘stopping because cyclist is crossing
my lane.’
Figure 4: Sample screenshots and the generated explanations (including observations announcement and causal
explanations) from the three driving scenarios. Heatmaps of gaze points from all the participants are plotted over the
images, indicating areas of interest. In the Abstract scenario (Figure 4a), all movable/dynamic non-human actors are
referred to as ‘Vehicle’. Thus, a cyclist was referred to as a vehicle. Figure 4b depicts a scene from the Specific(5)
scenario in which the AV’s perception system accurately identified and classified a motorbike and provided a fine-grained
explanation for this. In the Specific(50) scene (Figure 4c), the AV’s perception system misclassified a pedestrian as a
cyclist. The fine-grained/specific explanation provided exposed this error.
Figure 5: Study procedure. Eye calibration was done with the VR headset; participants drove for two minutes, participants
experienced each of the 4 mins scenarios in counterbalanced order and completed the Feeling of Anxiety, Perceived Safety,
and Takeover Feeling Questionnaire (APT Scale) in between each scenario. Participants were debriefed.
Behavioural Cues We also used Button Presses, Fixation Divergence, and Saccade Difference as additional metrics.
Button Presses were used to express unsafe, anxious or confused feelings.
Fixation Divergence is the Euclidean distance between mean participants’ fixation points and reference fixation
points. This provides information to draw inferences about participants’ distractions.
For Saccade Difference, we estimated participants’ saccade velocity over time following the method in [23] and
found the difference from a reference saccade velocities. Saccade is the rapid movement of the eye between fixation
points. Saccade velocity is the speed of such movements. The fixation and saccade reference points (or ground truths)
were the fixation and saccade records obtained from the researcher, who also participated in the study.
4.3. Procedure
The procedure of the experiment is illustrated in Fig. 5. After all preliminary form completions and briefings, we
introduced the physical driving simulator and explained the subsequent steps, which involved a pre-experiment manual
driving session in VR mode lasting for 2 minutes. Participants were informed that the purpose of this pre-experiment
D. Omeiza et al.: Preprint submitted to Elsevier Page 12 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
exercise was to familiarise them with the simulation environment and to identify individuals prone to motion sickness,
who would then be excluded from the main experiment.
Upon completion of the manual driving exercise, the researcher removed the VR headset from the participant
and explained the aim and procedure of the main experiment. The instructions included the following statements:
‘you would experience 3 autonomous rides by different vehicles, [...] and after each ride, you would complete a short
survey. The vehicle drives along a predefined path for about 4 minutes and provides explanations for its planned driving
decisions and announces relevant objects in its environment. [...]. The vehicle tells you its next direction at a junction
or an intersection using its right or left red light indicators on its dashboard accordingly.[...] Simply click any of these
buttons if the decision or the explanation of the vehicle makes you feel confused, anxious or unsafe [...]’. The researcher
then placed the VR headset back on the participant and initiated the scenarios. Complete counterbalancing was applied
to the scenario treatments, resulting in six different orders of scenarios. Each participant experienced the scenarios in
one of these six orders, with approximately six participants per order.
Participants were encouraged to rest briefly after each driving experience, with the VR headset removed. A short
debriefing session followed the study, and participants were given a £10 Amazon gift card. The entire experiment lasted
approximately 50 minutes.
The researcher also participated in the experiment, experiencing all three scenarios. Throughout, the researcher
focused on the lane ahead and the actors referenced by the explanations. Neither erroneous nor abstract explanations
influenced the researcher’s focus, as the researcher remained attentive to the lane and the actors or obstacles impacting
the AV’s actions, regardless of the explanations. This consistency was due to the researcher’s familiarity with all
scenarios. The data from the researcher served as a reference or ground truth. Notably, the researcher’s eye movements
matched normal human saccadic velocity, which reaches 300—400°/seconds [54,65].
5. Quantitative Results
5.1. Psychological Factors Analysis
To test our hypotheses listed in Section 2.3, we analyzed data from the three APT questionnaires. We created a latent
variable (Feeling of Anxiety) by averaging responses to AVAM Items 19—21, and another latent variable (Perceived
Safety) by averaging responses to AVAM Items 24—26. We calculated Cronbach’s Alpha (𝛼) for the independent
variables that formed the latent dependent variables to ensure they had adequate internal consistency. Results with
an adjusted p-value less than 0.05 (𝑝<.05) were considered significant. P-values were adjusted using Bonferroni
corrections, where the calculated p-values were multiplied by the number of scenarios, to reduce the likelihood of Type
I errors (false positives). Normality tests, including the Kolmogorov-Smirnov, Shapiro-Wilk, and Anderson-Darling
D. Omeiza et al.: Preprint submitted to Elsevier Page 13 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
Table 3
Descriptive statistics from APT questionnaire analysis.
Perceived Safety
Cronbach 𝛼∶ 0.87,
𝐻(2) =
8.17,𝒑=.𝟎𝟏𝟕
Feeling of Anxiety
Cronbach 𝛼∶ 0.86
𝐻(2) =
13.32,𝒑=.𝟎𝟎𝟏
Takeover Feeling
𝐻(2) =
6.27,𝒑=.𝟎𝟒𝟒
Mean SD Mean Rank Mean SD Mean Rank Mean SD Mean Rank
Abstract 4.89 1.35 2.15 2.81 1.34 1.72 2.79 1.91 1.68
Specific(5) 4.93 1.13 2.22 2.79 1.2 1.81 3.31 1.79 2.10
Specific(50) 3.86 1.58 1.63 3.93 1.68 2.47 3.87 1.94 2.22
tests, indicated a violation of normality in the Feeling of Anxiety,Perceived Safety, and Takeover Feeling factors. Hence,
we performed a Friedman test for these dependent variables, see Table 3and Figure 6.
Figure 6: Perceived safety, feeling of anxiety, and takeover feeling distribution. Perceived safety is highest in the Specific(5)
scenario, the feeling of anxiety is highest in the Specific(50), and takeover feeling is lowest in the Abstract scenario.
H1.1 - Perceived Safety Low transparency yields a higher perception of safety in an AV with perception system
errors.
A Friedman test was conducted. No significant difference was found in the scenario pair: Abstract —Specific(5),
and the pair: Abstract —Specific(50). In fact, the perceived safety mean rank in the Specific(5) scenario (2.22) was
higher than that in the Abstract scenario (2.15), see Table 3. Therefore, there was no sufficient evidence in support of
hypothesis H1.1.
D. Omeiza et al.: Preprint submitted to Elsevier Page 14 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
H1.2 - Feeling of Anxiety Passengers’ feeling of anxiety increases with increasing perception system errors in a
highly transparent AV. A Friedman test indicated a significant difference in the Feeling of Anxiety across scenarios,
𝐻(2) = 13.32, 𝑝 =.001. The pairwise scenario comparisons of Abstract -Specific(50) and Specific(5) -Specific(50)
resulted in an adjusted p-value of .003 and .01 respectively (see Table 3). Hence, there is strong evidence in support of
hypothesis H1.2.
H1.3 - Takeover Feeling In highly transparent AVs, passengers are more likely to develop the feeling to take over
navigation control from the AV with higher perception system errors. A Friedman test showed a significant difference
in Takeover Feeling across scenarios, 𝐻(2) = 6.27, 𝑝 =.044. While the pairwise scenario comparison of Abstract
-Specific(50) resulted in an adjusted p-value of .017, the pairwise comparison of Specific(5) -Specific(50) resulted
in an adjusted p-value of 0.61. Hence, there is no significant difference in Takeover Feeling between Specific(5) and
Specific(50) scenarios, and therefore, no evidence in support of hypothesis H1.3 (see Table 3).
Figure 7: Fixation divergence across scenarios. While Specific(5) had the highest mean fixation divergence, Specific(50)
had more frequent high fixation divergences. Red vertical bars represent the positions in time where causal explanations
were provided.
5.2. Behavioural Cues Analysis
H2.1 - Visual Responses Visual feedback from passengers correlates with passengers’ anxiety. At this stage, we
utilised the reference data from the researcher. We estimated the Euclidean distances between participants’ fixation
points and the reference fixation points over time for each participant.
Results from Spearman correlation showed that there was no significant association between the Feeling of Anxiety
and Fixation Divergence,𝑟(115) = −0.07, 𝑝 =.442. See the fixation divergence plot in Figure 7. Results from
Spearman correlation showed that there was no significant association between the Feeling of Anxiety and saccade
difference,𝑟(115) = 0.1, 𝑝 =.281. However, there was a significant association between perceived safety and saccade
difference,𝑟(115) = −0.25, 𝑝 =.007., indicating a weak negative correlation between perceived safety and saccade
difference. Hypothesis H2.1, therefore, has no sufficient support. See the saccade difference plot in Figure 8.
In addition to correlation, we checked for significant differences. There was a significant difference in Fixation
Divergence between Abstract and Specific(5) with an adjusted p-value of .028, and between Specific(5) and Specific(50)
D. Omeiza et al.: Preprint submitted to Elsevier Page 15 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
Figure 8: Saccade velocity difference across scenarios. Specific(5) had the lowest mean saccade velocity difference while
the Abstract scenario had the highest. Red vertical bars represent the positions in time where causal explanations were
provided.
Table 4
Descriptive statistics from the haptic and Visual responses.
ButtonPress
𝐻(2) =
15.44,𝒑<.𝟎𝟎𝟏
Fixation Divergence
𝐻(2) =
20.67,𝒑<.𝟎𝟎𝟏
Saccade Difference
𝐻(2) =
15.35,𝒑<.𝟎𝟎𝟏
Mean SD Mean Rank Mean SD Mean Rank Mean SD Mean Rank
Abstract 2.26 3.35 1.72 4.37 1.84 1.95 1.25 0.43 2.42
Specific(5) 1.64 1.63 1.77 7.66 5.21 2.54 1.06 0.42 1.54
Specific(50) 4.9 4.33 2.51 3.2 1.23 1.51 1.17 0.45 2.04
with an adjusted p-value < .001. See Table 4for descriptive statistics. Also, there was a significant difference between
Abstract and Specific(5) with respect to Saccade Difference (adjusted p-value of < .001). See Figure 4for sample
scenes from each scenario with the generated explanations. All the participants’ gaze points are plotted as heatmaps
over the screenshots.
5.2.1. Haptic Response
Participants were asked to press a button on the Logitech wheel when they felt confused, anxious, or unsafe by
the explanations or the decision of the AV during the ride. Spearman rank correlation was used as a measure to
investigate monotonic associations. There was a weak negative correlation between the variables Perceived Safety
and ButtonPress (𝑟(115) = −0.31, 𝑝 =.001), a weak positive correlation between the Feeling of Anxiety and
ButtonPress (𝑟(115) = 0.31, 𝑝 =.001), and insignificant correlation between the Feelings to Takeover and ButtonPress
(𝑟(115) = 0.15, 𝑝 =.099).
We also checked for statistically significant differences in Button Presses across scenarios. There was a significant
difference in ButtonPresses,𝐻(2) = 15.44, 𝑝 < .001. This was specifically in the pairs: Abstract -Specific(50) with
adjusted p-value .002, and Specific(5) -Specific(50) with adjusted p-value .005. See Figure 9for behavioural cues
results.
D. Omeiza et al.: Preprint submitted to Elsevier Page 16 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
Figure 9: Button presses, fixation divergence and saccade difference distribution. Button presses are fewer in the Specific(5)
scenario. Fixation divergence is highest in the Specific(5) scenario, and saccade difference is lowest in the Specific(5).
6. Qualitative Results: Themes and Reflections
We obtained qualitative data from the APT questionnaire administered after every scenario. Participants were asked
to describe their feelings regarding the explanations they received during the ride. Table 5and Figure 10 describe the
themes obtained from the inductive thematic analysis of the comments. Themes are broadly categorised based on the
participants’ feelings, their assessment of the explanations, and the vehicle dynamics.
Perceptual errors in the Specific(50) scenario evoked negative emotions of anxiety, feeling to takeover navigation
control and distrust. CAND1 expressed a feeling of anxiety: ‘The explanations made me feel a bit anxious, it says many
things that were not right and misleading. I had the urge to look at the buildings and the environment but could not
really do that because I wanted to be sure the vehicle is taking the right decision.’. CAND39 expressed the urge to
takeover navigation control: ‘When the explanations are false, e.g. ’a cyclist is crossing my lane’, and it is actually
a pedestrian, it made me slightly anxious and likely to want to take over. But nevertheless, I felt safe in the vehicle’.
CAND5 expressed distrust in the AV: ‘anxious as the vehicle did not correctly understand the environment and the
types of vehicles around it, which made me trust its judgement less’. More participants expressed a feeling of safety in
the Specific(5) scenario: ‘felt safe that the vehicle understood the road and what was going on around us’. About the
same number of participants expressed a decline in their feeling of anxiety in the Abstract and Specific(5) scenarios.
An example is CAND34’s comment about the abstract scenario: ‘When the explanations provided are more general,
e.g. ’vehicle’ instead of ’van’ and ’road user’ instead of ’cyclist’, it feels like the vehicle has a better understanding of
D. Omeiza et al.: Preprint submitted to Elsevier Page 17 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
Table 5
Themes derived from the thematic analysis of the qualitative data from participants. Freq. = Frequency of occurrence, SP
= Scenario Percentage
Abstract Specific(5) Specific(50)
Category Theme Freq. SP (%) Freq. SP (%) Freq. SP (%)
Feelings Anxious 2 5 2 5 8 21
Less Anxious 5 13 513 1 3
Safe 9 23 12 31 7 18
Unsafe 0 0 1 313
Takeover 2 5 2 5 7 18
Confident 2 5 5 13 3 8
Trust 2 51 3 2 5
Distrust 1 3 0 0 6 15
Reassuring 5 13 2 5 0 0
Uncomfortable 2 51 3 0 0
Explanations Good Timing 1 30 0 0 0
Bad Timing 7 18 1 3 1 3
Plausible 2 5 10 26 1 3
Implausible 5 13 3 8 25 64
Unintelligible 6 15 0 0 0 0
Repetitive 3 8 4 10 2 5
Vague 5 13 0 0 0 0
Vehicle Dynamics Careful Manoeuvre 3 8 2 5 4 13
Aggressive Manoeuvre 1 3 3 8 5 13
Vehicle Feature 0 0 3 81 3
Figure 10: Themes derived from the thematic analysis of the qualitative data from participants. Frequency is expressed in percentage
of the total number of responses in each scenario.
the surroundings because it gives a correct explanation, so I felt less anxious and unsafe’. The abstract explanations
might have concealed some errors, in turn, reducing the feeling of anxiety.
There were specific comments about the explanations across the three scenarios. Many participants thought that
the explanations in the Specific(5) were plausible in that they sounded correct and aligned with what the participants
saw. For example: ‘Explanations were clear and made sense. Still don’t feel some of the reactions were as quick as
D. Omeiza et al.: Preprint submitted to Elsevier Page 18 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
I might have made them’—CAND14. There were a good number of comments around the implausible nature of the
explanations in the Specific(50) scenario. For example, CAND20 said, ‘The vehicle this time had difficulty giving
the correct reason for stopping/going. Couldn’t tell the difference between a pedestrian and a cyclist sometime or
thought that traffic lights were off instead of green. I feel that this time I would have wanted more control over the car,
particularly at traffic lights as I could determine better if a traffic light was ’working’ or not’.
A couple of candidates thought that the explanations in the Abstract scenarios were either too early or late. For
example, ‘The explanations should have arrived a bit earlier, like a few meters before the vehicle actually stops so that
I will know that it is planning to stop. Also, I would be more comfortable if the explanation ’traffic sign’ was ’traffic
light is red/green’. when referring to a traffic light.’—CAND19.
Some interesting comments were made about the vehicle’s driving style and its interior. For example, CAND31
made a comment about the careful manoeuvre of the vehicle in the Specific(50) scenario: ‘I was calm throughout the
journey. There was no feeling of anxiety as the vehicle did not speed too much to make me feel that way.’—CAND31.
There was a comment relating to aggressive manoeuvre in the Abstract scenario: ‘Seemed like oncoming vehicles
were going to collide with me. It seems to sometime drive on pavements when negotiating corners.’—CAND35. The
rotating steering wheel of the vehicle made some of the participants uncomfortable: ‘The steering wheel moving
abruptly startled me sometimes.’—CAND21 (Specific(5) scenario). Some participants liked the vehicle indicators and
the sound they made when indicating the next directions. ‘The indicator sound was nice to hear. [...]’—CAND6
(Specific(50) scenario).
7. General Discussion
We examined the effects of explanation specificity (abstract and specific) in AD, while accounting for varying
degrees of perception system errors (low and high). We focused on how this setup would impact passengers’
perceived safety and related factors, such as the feeling of anxiety and the thought to takeover control. Our results
not only corroborate but also extend previous findings in the field, among others, demonstrating that while intelligible
explanations generally create positive experiences for AV users [50,26,56,42], this effect is predominantly observed
when the AV’s perception system errors are low.
7.1. Psychological Effects
Hypothesis 1.1 - Low transparency yields higher perception of safety Contrary to expectations, participants
expressed a greater sense of safety in the Specific(5) scenario, indicating a preference for specific explanations in
an AV with notably reduced perception system errors. This finding challenges our initial hypothesis and suggests a
more nuanced relationship between transparency and perceived safety. It appears that users may prefer more detailed
D. Omeiza et al.: Preprint submitted to Elsevier Page 19 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
information when the system demonstrates high reliability. However, overly detailed explanations were perceived as
verbose and repetitive by some participants in the specific scenarios. Therefore, achieving a balanced approach between
explanation specificity (or transparency in general) and the cognitive load imposed on passengers is crucial [52].
This observation underscores the importance of carefully calibrating the level of transparency to optimise user
comprehension and trust.
As observed from Figure 6, AVs characterised by high transparency and high perception system errors generally
elicited lower perceptions of safety among passengers. This result aligns with previous work suggesting that
transparency can sometimes amplify the negative effects of errors on trust in automated systems [14]. In our study,
the combination of detailed explanations and frequent errors may have heightened passengers’ awareness of the
system’s limitations, potentially exacerbating their safety concerns. Nonetheless, qualitative responses revealed that
a minority of participants reported positive sentiments despite these errors. They praised the vehicle for its ability to
detect obstacles and respond appropriately, suggesting that for these individuals, the AV’s correct decision-making
outweighed concerns about the specific type of obstacle encountered.
The diverse reactions we observed highlight the complex nature of human-AV interaction and underscore the need
for personalised approaches to transparency and explanation design. Adaptive explanation strategies that consider
individual differences in information processing and risk perception may be necessary to optimise user experience and
safety perceptions in autonomous vehicles.
Hypothesis 1.2 - Feeling of Anxiety increases with increasing perception system errors Drivers’ anxiety has
been noted to increase when utilising AVs [35]. In our study, we anticipated that passengers’ feelings of anxiety would
correspondingly increase with higher levels of perception system errors in an AV. This hypothesis was supported by a
significant difference in anxiety levels observed between the Specific(5) and Specific(50) scenarios. In the context of
AVs, perception system errors represent a critical factor that can directly impact passenger safety and, consequently,
their psychological state. As the frequency of errors increases, passengers may experience a heightened sense of
uncertainty and loss of control. This might be contributors to the reported anxious feelings.
Given that participants perceived the highest safety in the Specific(5) scenario (as seen Hypothesis 1.1), we
expected the lowest levels of anxiety in this scenario, assuming a relationship between perceived safety and anxiety,
as suggested by prior research [15]. This inverse relationship between perceived safety and anxiety aligns with the
broader psychological concept of risk perception and its impact on emotional states [60,58]. Our findings extend
beyond the mere confirmation of this relationship, highlighting the nuanced interplay between system transparency,
error frequency, and the feeling of anxiety. The significant difference in anxiety levels between the Specific(5) and
Specific(50) scenarios suggests that there may be a threshold of error frequency beyond which anxiety levels sharply
D. Omeiza et al.: Preprint submitted to Elsevier Page 20 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
increase. This notion is supported by research in risk perception, which indicates that individuals often have a non-linear
response to increasing risk levels [22]. Investigating this threshold is an interesting topic for future research.
While Dillen et al., [15] primarily examined how AV driving styles affect driver anxiety and comfort, they noted
that certain in-vehicle features, such as a rotating steering wheel, could also influence feelings of anxiety among
participants. This experience was reported by participants (e.g., CAND21) in our study. This observation underscores
the multifaceted nature of anxiety in AV contexts, where both system performance and physical design elements play
crucial roles.
Furthermore, our results suggest that the relationship between perception system errors and anxiety may be
moderated by the level of explanation specificity. In scenarios with high transparency (i.e., specific explanations),
the impact of errors on anxiety might be more pronounced, as passengers are more acutely aware of the system’s
limitations. On the other hand, uncertainty about the workings of algorithms and perceived lack of control has been
noted to cause what researchers term ‘algorithmic anxiety’ [19]. This is also observed from our result in Table 6as the
feeling of Anxiety is higher in the Abstract scenario compared to Specific(5).
It is worth noting that individual differences in technology acceptance and risk tolerance may also influence the
anxiety response to perception system errors [9]. Future research could explore these individual factors to develop a
more comprehensive model of anxiety in AV contexts.
Hypothesis 1.3 - Takeover feeling increases with the increase in perception system error Contrary to
our hypothesis, the data did not support an increase in takeover feeling with increased perception system errors.
While we observed a significant difference between the Abstract and Specific(50) scenarios, the lack of significant
difference between Specific(5) and Specific(50) scenarios suggests a complex relationship between system errors and
user responses. This finding indicates that mere disclosure of perception system errors does not necessarily escalate
passengers’ desire for control, pointing to a possible threshold effect in how passengers perceive and respond to
AV errors. These results challenge the notion that awareness of system imperfections automatically erodes trust in
automation [39], suggesting instead that passengers may be more tolerant of disclosed errors than previously thought,
especially with regards to seizing control in autonomous driving.
Our empirical findings contrast with the conceptual analysis presented by [61], who advocated for shared control
between vehicle and user. Our results suggest that users might be more comfortable with full automation than theoretical
analyses have predicted, even when aware of system limitations. It is crucial to reconcile these findings with our
previous results showing increased anxiety levels as error rates increased. This apparent contradiction suggests a
complex interplay between emotional responses and behavioural intentions in AV contexts. While passengers may
experience heightened anxiety feelings when aware of higher error rates, this emotional response does not necessarily
D. Omeiza et al.: Preprint submitted to Elsevier Page 21 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
translate into a stronger desire to take control of the vehicle. This disconnect between anxiety and takeover feeling
could be attributed to factors such as perception of one’s own ability to manage the situation, cognitive dissonance
between acknowledging risks and maintaining comfort with automation [3], or trust in the system’s overall capability.
It is important to note on the other hand that our qualitative results (Table 5) indicate higher number of comments
relating to takeover intents in Specific (50) compared to Specific (5). It might be the case that participants might have
struggled to calibrate their thoughts or feelings on a 5-point Likert scale. It might also be the case that this difference
in frequency of the said comments are statistically insignificant. This would be investigated in future work.
7.2. Behavioural Cues
Hypothesis 2.1 - Visual signal correlates with anxiety Our study’s findings challenge the established link
between anxiety and visual distraction proposed by Hepsomali et al. [28]. We found no significant correlation between
fixation point divergences and anxiety levels across the different scenarios. This surprising outcome might be attributed
to the varied individual priorities in visual attention. For instance, while some participants might have focused on
the cityscape, others may have concentrated on areas highlighted by the explanations. The Specific(50) scenario
exhibited higher divergences in fixation points compared to the Specific(5) and Abstract scenarios, possibly due to
misclassifications in explanations directing attention to incorrect elements. However, the Specific(5) and Abstract
scenarios showed similar fixation effects, suggesting that explanations might have been more effective in these
scenarios than in the Specific(50) scenario. Contrary to Dillen et al., [15]’s findings on the relationship between eye
movement entropy and anxiety, our study did not reveal a significant correlation between saccade differences and
anxiety levels. Interestingly, the Specific(5) scenario demonstrated the lowest saccade difference, potentially indicating
reduced distraction or confusion. This assumption is based on the understanding that saccade velocity reflects the
speed of gaze shifts between fixation points. In contrast, the Abstract scenario exhibited the highest saccade difference,
which could be interpreted as more active visual searching due to the non-specific nature of the explanations failing to
effectively guide participants’ gaze.
These findings amplifies the existing complexities in the relationship between visual behaviour, anxiety, and the
nature of explanations in automated driving scenarios. Further research to better understand the interplay between
cognitive states, visual attention, and the effectiveness of different types of explanations in automated driving contexts
is in order. Additionally, investigating the potential impact of individual differences in visual processing and attention
allocation could provide valuable insights into designing more effective and personalised explanation systems for
automated vehicles.
D. Omeiza et al.: Preprint submitted to Elsevier Page 22 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
7.3. Practical Implications
Our findings challenge the initial assumption that passengers may not desire specific explanations detailing error
information from automated vehicles (AVs). Contrary to this presupposition, the study reveals a preference among
passengers for specific explanations, particularly when the AV’s perception system demonstrates near-perfect accuracy.
This insight has significant implications for both manufacturers and regulators in the AV industry. The observed inverse
relationship between perception system errors and passenger anxiety underscores the critical need for highly transparent
AVs with exceptional perception and decision-making capabilities. This finding aligns with the broader literature
on trust in automation, which emphasises the importance of system reliability and transparency in fostering user
acceptance [39]. Manufacturers should prioritise the development of robust perception systems that minimise errors,
while simultaneously implementing transparent communication mechanisms to convey system status and decision-
making processes to passengers. Regulators, in turn, should consider establishing or strengthening existing standards
for AV perception accuracy and mandating clear and usable interfaces for conveying this information to passengers.
While our study did not observe direct consequences of misclassification errors on AV actions, it is crucial to
recognise the potential ramifications of such errors in more complex, real-world scenarios. The accurate classification
of obstacle types is paramount in determining appropriate vehicle responses, particularly when dealing with dynamic
obstacles possessing varied manoeuvrability characteristics. This aligns with research efforts on situation awareness
in automated systems, which emphasises the importance of accurate environmental perception for effective decision-
making [20]. In intricate traffic scenarios, even minor perceptual inaccuracies could lead to sub-optimal navigation
decisions, potentially compromising safety. Therefore, AV developers must strive for highly accurate environmental
estimations to ensure appropriate responses to diverse obstacle types, such as bicycles or motorcycles, with very similar
attributes but differing capabilities.
The study’s findings highlights the link between transparency and accuracy in AV systems. This relationship echoes
the concept of ‘calibrated trust’ in automation, where user trust is appropriately aligned with system capabilities [39].
To foster this calibrated trust, AV interfaces should not only provide accurate information but also communicate
the system’s confidence levels and potential limitations. This approach can help manage passenger expectations and
maintain appropriate levels of situational awareness.
While visual feedback from experimental studies provides valuable insights into participants’ psychological and
behavioural responses, it is useful to simultaneously implement multiple methodologies (e.g., mixed methods) for more
accurate conclusions. Complementing observational data with other measurement techniques, such as standardised
surveys, physiological measurements, or qualitative interviews, can provide a more comprehensive understanding of
user experiences and perceptions. This multi-faceted approach allows for triangulation of data and more confident
interpretations of complex human-machine interactions.
D. Omeiza et al.: Preprint submitted to Elsevier Page 23 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
Future research should explore the long-term effects of exposure to AV explanations on passenger trust, anxiety,
and overall acceptance. Longitudinal studies could reveal how user preferences and responses evolve over time,
informing the design of adaptive explanation systems that cater to changing user needs and expectations. Additionally,
investigating the impact of cultural differences and individual variability in risk perception on AV explanation
preferences could yield insights for developing globally applicable, yet locally tailored, AV communication strategies.
In conclusion, these findings have far-reaching implications for AV design, regulation, and user experience. By
prioritising both transparency and accuracy, and adopting a holistic approach to user research, the AV industry can
work towards creating systems that not only meet technical performance standards but also address the complex
psychological needs of their users, thereby avoiding the transparency paradox and ultimately fostering greater public
acceptance and trust in automated transportation technologies.
8. Conclusion
In this study, we conducted a rigorous within-subject laboratory investigation (𝑁= 39) utilising an immersive
driving simulator to examine passengers’ preferences for explanation specificity levels and the impact of varying
perception system errors on perceived safety, and related constructs in autonomous vehicles (AVs). Our findings reveal
a link between explanation specificity, perception system errors, and passenger anxiety, contributing to the growing
body of literature on trust calibration and transparency in automated vehicles. We observed a significant increase in
anxiety levels when fine-grained, intelligible explanations exposed perception system errors. Interestingly, passengers
preferred highly detailed (or specific) explanations when AV perception system errors were low, despite the potential
for vague explanations to conceal such errors. Hence, as desirable as transparency is, the more we have it beyond
certain thresholds, the more anxious we might be. We refer to this phenomenon as the transparency paradox in the
context of this paper. While our study provides valuable insights, we acknowledge its limitations, particularly, the
absence of explanation personalisation, and the inability to select an AV driving style. Future research should address
these limitations by investigating the implications of providing passengers with personalisation options, drawing upon
theories of user-centred design and adaptive interfaces. Additionally, longitudinal studies and exploration of individual
differences could yield valuable information for tailoring AV explanations to diverse user populations. Our findings
have significant implications for AV explanation interface design, emphasising the need for adaptive, context-sensitive
explanation systems that balance transparency with user comfort, ultimately fostering greater public acceptance and
trust in autonomous transportation systems.
D. Omeiza et al.: Preprint submitted to Elsevier Page 24 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
CRediT authorship contribution statement
Daniel Omeiza: Conceptualisation of idea, writing and editing the original draft, software implementation, lab
experiment, data collection, data analysis. Raunak Bhattacharyya: Writing and editing, review, data collection,
lab experiment. Marina Jirotka: Conceptualisation, bi-weekly supervision and guidance, resources. Nick Hawes:
Supervision, provision of resources. Lars Kunze: Conceptualisation of idea, weekly supervision, and guidance,
resources provision, review.
Acknowledgements
This work was supported by the EPSRC RAILS project (grant reference: EP/W011344/1) and the EPSRC project
RoboTIPS (grant reference: EP/S005099/1).
References
[1] Adadi, A., Berrada, M., 2018. Peeking inside the black-box: a survey on explainable artificial intelligence (XAI). IEEE Access 6, 52138–52160.
[2] Anjomshoae, S., Omeiza, D., Jiang, L., 2021. Context-based image explanations for deep neural networks. Image and Vision Computing 116,
104310.
[3] Aronson, E., 1969. The theory of cognitive dissonance: A current perspective, in: Advances in experimental social psychology. Elsevier.
volume 4, pp. 1–34.
[4] Bin Issa, R., Das, M., Rahman, M.S., Barua, M., Rhaman, M.K., Ripon, K.S.N., Alam, M.G.R., 2021. Double deep Q-learning and faster
R-Cnn-based autonomous vehicle navigation and obstacle avoidance in dynamic environment. Sensors 21, 1468.
[5] Board, N.T.S., 2018. Collision between a sport utility vehicle operating with partial driving automation and a crash attenuator mountain view,
california. URL: https://www.ntsb.gov/investigations/AccidentReports/Reports/HAR2001.pdf. Accessed: Oct. 30, 2020.
[6] Buijsman, S., 2022. Defining explanation and explanatory depth in XAI. Minds and Machines 32, 563–584.
[7] Chakraborti, T., Sreedharan, S., Kambhampati, S., 2020. The Emerging Landscape of Explainable Automated Planning & Decision Making,
in: Bessiere, C. (Ed.), Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, International Joint
Conferences on Artificial Intelligence Organization. pp. 4803–4811.
[8] Chen, N.T., Clarke, P.J., Watson, T.L., Macleod, C., Guastella, A.J., 2014. Biased saccadic responses to emotional stimuli in anxiety: An
antisaccade study. PloS One 9, e86474.
[9] Choi, J.K., Ji, Y.G., 2015. Investigating the importance of trust on adopting an autonomous vehicle. International Jour nal of Human-Computer
Interaction 31, 692–702.
[10] Colley, M., Bräuner, C., Lanzer, M., Walch, M., Baumann, M., Rukzio, E., 2020. Effect of Visualization of Pedestrian Intention Recognition
on Trust and Cognitive Load, in: 12th International Conference on Automotive User Interfaces and Interactive Vehicular Applications, pp.
181–191.
[11] Colley, M., Eder, B., Rixen, J.O., Rukzio, E., 2021. Effects of Semantic Segmentation Visualization on Trust, Situation Awareness, and
Cognitive Load in Highly Automated Vehicles, in: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, pp.
1–11.
D. Omeiza et al.: Preprint submitted to Elsevier Page 25 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
[12] Colley, M., Rädler, M., Glimmann, J., Rukzio, E., 2022. Effects of Scene Detection, Scene Prediction, and Maneuver Planning Visualizations
on Trust, Situation Awareness, and Cognitive Load in Highly Automated Vehicles. Proceedings of the ACM on Interactive, Mobile, Wearable
and Ubiquitous Technologies 6, 1–21.
[13] Davidson, M.M., Butchko, M.S., Robbins, K., Sherd, L.W., Gervais, S.J., 2016. The mediating role of perceived safety on street harassment
and anxiety. Psychology of Violence 6, 553.
[14] De Visser, E.J., Pak, R., Shaw, T.H., 2018. From ‘automation’to ‘autonomy’: the importance of trust repair in human–machine interaction.
Ergonomics 61, 1409–1427.
[15] Dillen, N., Ilievski, M., Law, E., Nacke, L.E., Czarnecki, K., Schneider, O., 2020. Keep calm and ride along: Passenger comfort and anxiety
as physiological responses to autonomous driving styles, in: Proceedings of the 2020 CHI conference on human factors in computing systems,
pp. 1–13.
[16] Došilović, F.K.,Brčić, M., Hlupić, N., 2018. Explainable artificial intelligence: A survey, in: 2018 41st International Convention on Information
and Communication Technology, Electronics and Microelectronics (MIPRO).
[17] Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V., 2017. CARLA: An open urban driving simulator, in: Conference on Robot
Learning.
[18] Eiband, M., Buschek, D., Kremer, A., Hussmann, H., 2019a. The impact of placebic explanations on trust in intelligent systems, in: Extended
abstracts of the 2019 CHI conference on human factors in computing systems, pp. 1–6.
[19] Eiband, M., Völkel, S.T., Buschek, D., Cook, S., Hussmann, H., 2019b. When people and algorithms meet: User-reported problems in
intelligent everyday applications, in: Proceedings of the 24th international conference on intelligent user interfaces, pp. 96–106.
[20] Endsley, M.R., 1995. Toward a theory of situation awareness in dynamic systems. Human factors 37, 32–64.
[21] Endsley, M.R., 2000. Situation models: An avenue to the modeling of mental models, in: Proceedings of the Human Factors and Ergonomics
Society Annual Meeting, SAGE Publications Sage CA: Los Angeles, CA. pp. 61–64.
[22] Fischhoff, B., Bostrom, A., Quadrel, M.J., et al., 1993. Risk perception and communication. Annual review of public health 14, 183–203.
[23] Gibaldi, A., Sabatini, S.P., 2021. The saccade main sequence revised: A fast and repeatable tool for oculomotor analysis. Behavior Research
Methods 53, 167–187.
[24] Guidotti, R., Monreale, A., Ruggieri, S., Turini, F., Giannotti, F., Pedreschi, D., 2018. A survey of methods for explaining black box models.
ACM computing surveys (CSUR) 51, 1–42.
[25] Gunning, D., 2017. Explainable Artificial Intelligence (XAI). Defense Advanced Research Projects Agency (DARPA), nd Web 2, 1.
[26] Ha, T., Kim, S., Seo, D., Lee, S., 2020. Effects of explanation types and perceived risk on trust in autonomous vehicles. Transportation
Research Part F: Traffic Psychology and Behaviour 73, 271–280.
[27] Hart, W., Albarracín, D., Eagly, A.H., Brechan, I., Lindberg, M.J., Merrill, L., 2009. Feeling validated versus being correct: a meta-analysis
of selective exposure to information. Psychological Bulletin 135, 555.
[28] Hepsomali, P., Hadwin, J.A., Liversedge, S.P., Garner, M., 2017. Pupillometric and saccadic measures of affective and executive processing
in anxiety. Biological Psychology 127, 173–179.
[29] Hewitt, C., Politis, I., Amanatidis, T., Sarkar, A., 2019. Assessing public perception of self-driving cars: The autonomous vehicle acceptance
model, in: Proceedings of the 24th International Conference on Intelligent User Interfaces.
[30] Hoffman, R.R., Klein, G., 2017. Explaining explanation, part 1: Theoretical foundations. IEEE Intelligent Systems 32, 68–73.
[31] Information Commissioner’s Office, . What goes into an explanation? URL: https://ico.org.uk/for-organisations/
guide-to- data-protection/key- dp-themes/explaining-decisions- made-with- artificial-intelligence/
D. Omeiza et al.: Preprint submitted to Elsevier Page 26 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
part-1- the-basics- of-explaining-ai/what- goes-into- an-explanation/. Accessed: Jul, 2021.
[32] Khastgir, S., Birrell, S., Dhadyalla, G., Jennings, P., 2018. Calibrating trust through knowledge: Introducing the concept of informed safety
for automation in vehicles. Transportation Research Part C: Emerging Technologies 96, 290–303.
[33] Kment, B., 2006. Counterfactuals and explanation. Mind 115, 261–310.
[34] Koo, J., Kwac, J., Ju, W., Steinert, M., Leifer, L., Nass, C., 2015. Why did my car just do that? Explaining semi-autonomous driving actions to
improve driver understanding, trust, and performance. International Journal on Interactive Design and Manufacturing (IJIDeM) 9, 269–275.
[35] Koo, J., Shin, D., Steinert, M., Leifer, L., 2016. Understanding driver responses to voice alerts of autonomous car operations. International
journal of vehicle design 70, 377–392.
[36] Kunze, A., Summerskill, S.J., Marshall, R., Filtness, A.J., 2019a. Automation Transparency: Implications of Uncertainty Communication for
Human-Automation Interaction and Interfaces. Ergonomics 62, 345–360.
[37] Kunze, A., Summerskill, S.J., Marshall, R., Filtness, A.J., 2019b. Conveying Uncertainties using Peripheral Awareness Displays in the
Context of Automated Driving, in: Proceedings of the 11th International Conference on Automotive User Interfaces and Interactive Vehicular
Applications, pp. 329–341.
[38] Lavrinc, D., 2016. This is how bad self-driving cars suck in rain. URL: https://www.wired.com/2016/02/
googles-self- driving-car- may-caused-first- crash/. Accessed: Jul., 2020.
[39] Lee, J.D., See, K.A., 2004. Trust in automation: Designing for appropriate reliance. Human factors 46, 50–80.
[40] Liu, H., Hirayama, T., Watanabe, M., 2021. Importance of instruction for pedestrian-automated driving vehicle interaction with an external
human machine interface: Effects on pedestrians’ situation awareness, trust, perceived risks and decision making, in: 2021 IEEE Intelligent
Vehicles Symposium (IV), pp. 748–754.
[41] Lundberg, S.M., Lee, S.I., 2017. A unified approach to interpreting model predictions, in: Proceedings of the 31st International Conference
on Neural Information Processing Systems.
[42] M. Faas, S., Kraus, J., Schoenhals, A., Baumann, M., 2021. Calibrating Pedestrians’ Trust in Automated Vehicles: Does an Intent Display in
an External HMI Support Trust Calibration and Safe Crossing Behavior?, in: Proceedings of the 2021 CHI Conference on Human Factors in
Computing Systems.
[43] Magnaguagno, M.C., Fraga Pereira, R., Móre, M.D., Meneguzzi, F.R., 2017. Web planner: A tool to develop classical planning domains and
visualize heuristic state-space search, in: 2017 Workshop on User Interfaces and Scheduling and Planning (UISP@ ICAPS).
[44] McFarland, M., 2016. Who’s responsible when an autonomous car crashes? URL: https://money.cnn.com/2016/07/07/technology/
tesla-liabilityrisk/index.html. Accessed: Jul. 24, 2020.
[45] Michael, M., Schlipsing, M., 2015. Extending traffic light recognition: Efficient classification of phase and pictogram, in: 2015 International
Joint Conference on Neural Networks (IJCNN).
[46] Miller, T., 2019. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence 267, 1–38.
[47] Mittelstadt, B., Russell, C., Wachter, S., 2019. Explaining explanations in AI, in: Proceedings of the conference on fairness, accountability,
and transparency.
[48] Omeiza, D., Anjomshoae, S., Webb, H., Jirotka, M., Kunze, L., 2022a. From spoken thoughts to automated driving commentary: Predicting
and explaining intelligent vehicles’ actions, in: 2022 IEEE Intelligent Vehicles Symposium (IV).
[49] Omeiza, D., Kollnig, K., Webb, H., Jirotka, M., Kunze, L., 2021a. Why not explain? effects of explanations on human perceptions of
autonomous driving, in: IEEE International Conference on Advanced Robotics and its Social Impacts.
D. Omeiza et al.: Preprint submitted to Elsevier Page 27 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
[50] Omeiza, D., Web, H., Jirotka, M., Kunze, L., 2021b. Towards accountability: Providing intelligible explanations in autonomous driving, in:
2021 IEEE Intelligent Vehicles Symposium (IV).
[51] Omeiza, D., Webb, H., Jirotka, M., Kunze, L., 2022b. Explanations in autonomous driving: A survey. IEEE Transactions on Intelligent
Transportation Systems 23, 10142–10162. doi:10.1109/TITS.2021.3122865.
[52] Poursabzi-Sangdeh, F., Goldstein, D.G., Hofman, J.M., Wortman Vaughan, J.W., Wallach, H., 2021. Manipulating and measuring model
interpretability, in: Proceedings of the CHI Conference on Human Factors in Computing Systems.
[53] Quansah, F., Hagan Jr, J.E., Sambah, F., Frimpong, J.B., Ankomah, F., Srem-Sai, M., Seibu, M., Abieraba, R.S.K., Schack, T., 2022. Perceived
safety of learning environment and associated anxiety factorsdur ing COVID-19 in Ghana: Evidence from physical education practical-oriented
program. European Journal of Investigation in Health, Psychology and Education 12, 28–41.
[54] Raab, E.L., 1985. Normal saccadic velocities. Journal of Pediatric Ophthalmology & Strabismus 22, 20–22.
[55] Ribeiro, M.T., Singh, S., Guestrin, C., 2016. "Why should I trust you?" Explaining the predictions of any classifier, in: Proceedings of the
22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.
[56] Schneider, T., Hois, J., Rosenstein, A., Ghellal, S., Theofanou-Fülbier, D., Gerlicher, A.R., 2021. Explain yourself! transparency for positive
ux in autonomous driving, in: Proceedings of the CHI Conference on Human Factors in Computing Systems.
[57] Silvera, G., Biswas, A., Admoni, H., 2022. DReyeVR: Democratizing Virtual Reality Driving Simulation for Behavioural & Interaction
Research, in: ACM/IEEE Human Robot Interaction Conference.
[58] Slovic, P., 2010. The feeling of risk,. earthscan.
[59] Stanton, N.A., Salmon, P.M., Walker, G.H., Stanton, M., 2019. Models and methods for collision analysis: a comparison study based on the
uber collision with a pedestrian. Safety Science 120, 117–128.
[60] Stapel, J., Gentner, A., Happee, R., 2022. On-road trust and perceived risk in Level 2 automation. Transportation research part F: traffic
psychology and behaviour 89, 355–370.
[61] Terken, J., Pfleging, B., 2020. Toward shared control between automated vehicles and users. Automotive Innovation 3, 53–61.
[62] Tilley, A., 2016. Google’s self-driving car caused its first crash. URL: https://www.forbes.com/sites/aarontilley/2016/02/29/
googles-self- driving-car- caused-its-first- accident/?sh=5ae097b0538d. Accessed: Jun. 21, 2021.
[63] Voigt, P., Von dem Bussche, A., 2017. The EU General Data Protection Regulation (GDPR). A Practical Guide, 1st Ed., Cham: Springer
International Publishing .
[64] Wang, D., Yang, Q., Abdul, A., Lim, B.Y., 2019. Designing theory-driven user-centric explainable AI, in: Proceedings of the CHI Conference
on Human Factors in Computing Systems.
[65] Wilson, S.J., Glue, P., Ball, D., Nutt, D.J., 1993. Saccadic eye movement parameters in normal subjects. Electroencephalography and Clinical
Neurophysiology 86, 69–74.
[66] Zhu, J., Liapis, A., Risi, S., Bidarra, R., Youngblood, G.M., 2018. Explainable AI for designers: A human-centered perspective on mixed-
initiative co-creation, in: IEEE Conference on Computational Intelligence and Games (CIG).
A. Online Survey Questionnaire
Note that the same questionnaire was administered in all scenarios.
D. Omeiza et al.: Preprint submitted to Elsevier Page 28 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
1/4
ExplainableAutonomousDriving1
Page1
Canyoupleasecarefullyreadandrespondtothequestions?Yourresponsesarevaluableto
ourresearch.Thankyou!
Pleasedon'tselectmorethan1answer(s)perrow.
Pleaseselectatleast3answer(s).
Strongly
Disagree Disagree Somewhat
Disagree
Neither
Agree
Nor
Disagree
Somewhat
Agree Agree Strongly
Agree
Iwouldhave
concerns
aboutusing
thevehicle.
Thevehicle
wouldbe
somewhat
frighteningto
me.
Iamafraid
thatIwould
not
understand
thevehicle.
Canyoupleaseevaluatethefollowingstatementsbasedonyourexperiencefromthejust
concludeddrivingexercise? Required
D. Omeiza et al.: Preprint submitted to Elsevier Page 29 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
2/4
Pleasedon'tselectmorethan1answer(s)perrow.
Pleaseselectatleast3answer(s).
Strongly
Disagree Disagree Somewhat
Disagree
Neither
Agree
Nor
Disagree
Somewhat
Agree Agree Strongly
Agree
Ibelievethat
usingthe
vehicle
wouldbe
dangerous.
Iwouldfeel
safewhile
usingthe
vehicle.
Iwouldtrust
thevehicle.
Canyoupleaseevaluatethefollowingstatementsbasedonyourexperiencefromthejust
concludeddrivingexercise? Required
Pleasedon'tselectmorethan1answer(s)perrow.
Pleaseselectatleast1answer(s).
Strongly
Disagree Disagree Somewhat
Disagree
Neither
Agree
Nor
Disagree
Somewhat
Agree Agree Strongly
Agree
Duringthe
ride,Ihadthe
feelingto
takeover
controlfrom
thevehicle.
Canyoupleaseevaluatethefollowingstatementsbasedonyourexperiencefromthejust
concludeddrivingexercise? Required
D. Omeiza et al.: Preprint submitted to Elsevier Page 30 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
3/4
Whatisyourthoughtontheexplanationsprovidedbythevehicle,e.g.,madeyouless/more
anxious,safe,feelingtotakeovercontrol?
D. Omeiza et al.: Preprint submitted to Elsevier Page 31 of 28
A Transparency Paradox? Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies
4/4
Page2:Finalpage
Thankyouforcompletingthissurvey!
D. Omeiza et al.: Preprint submitted to Elsevier Page 32 of 28
