Neural correlates of the out-of-the-loop phenomenon during the supervision of an automated system in aeronautic context

Neural correlates of the out-of-the-loop phenomenon during the supervision of an automated system in aeronautic context | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 24 December 2025 V1 Latest version Share on Neural correlates of the out-of-the-loop phenomenon during the supervision of an automated system in aeronautic context Authors : Salomone. M. 0009-0005-5418-9632 [email protected] , Bruno Berberian 0000-0002-3908-4358 , and Aurélie Campagne Authors Info & Affiliations https://doi.org/10.22541/au.176656330.08996904/v1 159 views 74 downloads Contents Abstract Introduction Method Supervision task during training sessions Familiarization tasks Measures and statistical analysis Results Electroencephalographic data Influence of vigilance level on supervisory activity Electroencephalographic data Correlations between subjective and EEG measures Discussion Limits Conclusion Acknowledgements Data Availability Funding information Conflict of interest Author Contributions Reprints References Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Automation is now widespread in many fields, particularly in aeronautics. It enhances performance and safety but transforms active operators into supervisors. This transition can trigger the “out-of-the-loop” (OOTL) phenomenon, which is linked to impaired supervision and accidents. In an aeronautical context, our EEG study aimed to better understand and detect OOTL by analyzing associated brain activity while supervising an obstacle-avoiding automated system. The study focused on two key factors: trust (manipulated via reliability), and vigilance. We examined event-related spectral perturbations during obstacle detection (information exploration) and system decision-making (supervision and performance monitoring). Additionally, we assessed how the sense of agency (control felt) evolved with system reliability and examined its relationship with brain activity. Our results revealed that supervising a reliable system did not lead to disengagement but altered supervisory strategies. Monitoring reliable systems elicited an increased frontal theta activity during obstacle exploration and a reduced parieto-occipital alpha activity during system’s decision, suggesting participants re-engaged before the system’s decision and prioritized it due to its relevance for performance. Conversely, monitoring unreliable systems resulted in decreased parieto-occipital beta activity during obstacle exploration and reduced scalp-wide gamma activity during the system’s decision, indicating participants processed and integrated information earlier during the exploration phase, and then paid less attention to decisions made by the unreliable system. Interestingly, interacting with a perfectly reliable system did not weaken the sense of agency compared to interacting with an unreliable system. However, alpha activity reduction correlated with felt control only for reliable systems, suggesting participants relying on automation (as evidenced by lower alpha) processed less information during obstacle exploration. Heightened attention to system’s decisions may reflect a search for control, compensating for reduced upstream supervision. These findings clarify how OOTL factors shape automated system supervision and offer insights for real-time detection that can improve safety in human-automation collaboration. Introduction For many years, complex aeronautical (e.g., airplanes or Unmanned Aerial Vehicles - UAVs), rail, or automobile systems have integrated functionalities that automate specific actions (e.g., vehicle positioning warning system in automobile transportation, ground and traffic collision warning systems, and autopilot in the railway and aeronautics sectors). The integration of artificial intelligence has considerably increased their autonomy level. The automation of systems has always served the same purposes, namely increasing performance and safety (Norman, 1990). But this evolution has generated many changes in the activity of the operators as they have shifted from manual control to a supervisory role. Although the cooperation between humans and automated systems is generally successful, in some situations, automation integration can have negative consequences on safety and more globally on performance (Parasuraman & Riley, 1997; Sarter, Woods, & Billings, 1997). It can lead to the appearance of the out-of-loop (OOTL) phenomenon, a cognitive state that results in difficulty in understanding operations of the system, supervising it and detecting its errors, even when the operator is in physical control of the system (Endsley & Kiris, 1995; Lee & See, 2004; Berberian et al., 2017; Merat et al., 2018). According to the literature, this OOTL phenomenon would reflect operator disengagement, the origin of which could be overconfidence or even complacency towards the automated system, a deterioration in the vigilance level or attentional mechanisms, or engagement in a secondary task to the detriment of the main one (Endsley & Kiris, 1995; Merat et al., 2018). Cognitive changes linked to OOTL limit the operator’s ability to regain control of the system and can be a critical issue, especially outside of nominal situations (e.g., failures), in high-risk environments, as demonstrated by examples of UAV accidents (e.g., Rio-Paris flight crash in 2012; Bureau d’Enquête et d’Analyse, 2012). Determining in real time when an operator is out of the loop would help to avoid dramatic events and to identify countermeasures. The use of electroencephalography (EEG) is a relevant tool for achieving this goal, thanks to its temporal resolution and its ability to finely detect different cognitive states through time-frequency power analysis. However, the brain correlates of the OOTL phenomenon are not fully identified. The aim of this study is to better characterize these correlates by manipulating two causal factors of the OOTL phenomenon, i.e., the trust in the system and the vigilance level, and to assess their influence on the brain activity associated with the supervision of an automated system. Effective supervision of an automated system requires an appropriate level of trust, i.e., one that allows the human agent to correctly perform their tasks through effective cooperation with the system (Lee, 2008). Given the importance of this factor, its brain correlates and influence on supervisory activity have been increasingly studied (e.g., Kohn et al., 2021; Vereschak, Bailly & Caramiaux, 2021). Some of these studies have shown that a lack of trust in automated systems is generally associated with more cautious behavior, resulting in increased supervision with greater cognitive effort and attentional engagement. In EEG activity, this is reflected in increased fronto-central theta activity (Jung, Dong, & Lee, 2019) or alpha desynchronization (Blais et al., 2019; Oh et al., 2020). However, high trust can also lead to high attentional engagement. Indeed, in the context of a collaborative gaming task with a trusted or untrusted third party, Blais and collaborators (2019) showed greater suppression of alpha prior to both trusting decisions with a trusted third party and distrusting decisions with an untrusted third party, suggesting greater attentional engagement when decision is consistent with the third party’s trust profile. Moreover, Hu and collaborators observed an increase in fronto-median theta activity when participants were evaluating the outcome of a trusted decision (Hu et al., 2018). Apart from theta and alpha activity, changes in other frequency bands were also found. During a decision phase, Oh and collaborators (2020) observed higher gamma activity when participants were not confident in their system and higher beta activity when they were (Oh et al., 2020). Taken together, these findings suggest that evaluating the activities and decisions of others, whether human or non-human, and the associated neural correlates are influenced by the level of trust placed in them. But how does this level of trust and its influence evolve when supervising a perfectly reliable system? The previously cited works were conducted in contexts where automated systems made errors. Studies have shown that even the slightest error can drastically reduce confidence in a system despite its initially high reliability level (e.g., Dzindolet et al., 2003; Oh et al., 2020). In the case of a perfectly reliable system, some studies suggest that excessive confidence may arise in relation to the system’s actual performance. This overconfidence is sometimes described as complacency when supervisors to unreservedly rely on the automated system’s actions and decisions, completely delegating tasks to the automaton (Rice, 2009; Whan et al. 2023). In this context, however, what about the quality of the system’s supervisory activity? Delegating tasks from operators to an automated system can induce a passive state in the operator, which can reduce their level of vigilance, particularly over time. Declines in alertness are commonly observed with time on task during monotonous and repetitive tasks and are reflected at the brain level by an increased alpha activity spanning occipital to frontal regions, as well as increased frontal theta activity (Borghini et al., 2014; Washer et al. 2014; Wascher, Getzmann, & Karthaus, 2016; Maille et al., 2022). Somon and collaborators (2022) also observed such an increase in alpha activity over time in the frontal and occipital regions when monitoring the correct responses of an aircraft equipped with an automated obstacle avoidance system. The gradual decline in alertness can hinder the ability to effectively supervise human and artificial agents and their errors. The negative consequences of this deterioration are well documented (Mackworth, 1948; Parasuraman, Warm, & Dember, 1987). Error detection is classically associated with increased frontal theta activity and decrease parieto-occipital alpha activity, reflecting greater control and attentional engagement (Washer et al., 2014; Chuang et al., 2018; Somon et al., 2022). This has been observed both when supervising humans (Washer et al., 2014; Arnau et al., 2021) and automated or non-automated artificial systems (Chuang et al., 2018; Somon et al., 2022). With time on task, theta activity and alpha suppression decrease, indicating a decline in the efficiency of supervision and error processing (Washer et al., 2014; Arnau et al., 2021; Somon et al., 2022). But how does the effect of vigilance level on supervisory activity evolves with the operator’s trust in the supervised agent, particularly when interacting with automated systems of different reliability? To our knowledge, no study has evaluated the supervisory activity of an automated system and its associated neural correlates depending on trust and vigilance levels jointly. It should be noted that a degraded state, such as reduced vigilance, can also modulate trust, complacency, and dependence on highly autonomous systems. However, the literature on this effect is controversial. Some studies suggest that decreased vigilance leads to increased trust and dependence on automated systems (e.g., Neubauer et al., 2012), while others report the opposite (Reichenbach et al., 2011; Wohleber et al. 2019; Lopes et al., 2022). These discrepancies may arise from the fact that trust is influenced by internal factors such as mood or cognitive resources, which are themselves modulated by time on task (Hoff & Bashir, 2015). In addition to trust and vigilance, a decline in the sense of agency (SoA) has been proposed as another factor affecting supervisory performance and contributing to the OOTL phenomenon (Berberian, 2019; Wen, Kuroki, & Asama, 2019). The SoA refers to the sense of control we have over our actions or external events (Gallagher, 2000; Synofzik, Vosgerau & Newen, 2008). This feeling arises when the expected results of our actions are consistent with those observed. Its emergence can be based on sensorimotor information as well as on high-level processes, such as goal achievement. This sense of control guides our attention to certain elements of our environment. According to Wen and Haggard (2018), the level of control over a specific element relative to other elements in the environment determines its attentional capture. Moreover, the sense of control influences individuals’ involvement in an activity by motivating the effort and cognitive resources they committed to it (Eitam, Kennedy, & Higgins, 2013; Vantrepotte, Chambon, & Berberian, 2023). The intelligibility and predictability of a system’s goals are essential for developing a SoA. However, the absence of these characteristics in an opaque design of automated systems can limit users’ ability to understand and predict system’s intentions (Berberian et al., 2012; Sahai, et al., 2019). Consequently, a decrease in sense of control has been reported several times alongside an increase in level of automation and result in a deterioration of performance (Berberian et al., 2012; Berberian et al., 2017; Ueda, Nakashima, & Kumada, 2021). Nevertheless, different factors could increase the sense of control when interacting with an automated system (Berberian, 2019; Wen, Kuroki, & Asama, 2019). For example, several studies showed that participants can still feel in control if the system is predictable or if the help it provides improves performance, even if this help is actually outside their control (e.g., Dzindolet et al., 2002; Metcalfe & Greene, 2007; Wen, Yamashita, & Asama, 2015; Ueda, Nakashima, & Kumada, 2021; Vantrepotte et al., 2023). Given its influence on both information processing and task engagement, the perceived level of control must be taken into account. Our study aims to characterize the brain correlates of supervisory activity of an automated system and their evolution depending on trust and vigilance levels, two of the main causal factors of the OOTL phenomenon. These correlates will be studied in an ecological context during a realistic task involving the monitoring and (if necessary) correction of an UAV’s autonomous obstacle avoidance decisions during two experimental sessions. Since very few studies have used ecological tasks, a better understanding of these correlates in conditions that approximate real-life will help improve the generalizability of literature data obtained in less ecological laboratory situations. EEG will be used to measure brain activity associated with supervisory activity during the system decision and the preceding obstacle exploration phase. In this study, trust will be manipulated through two reliability levels of the supervised system: perfectly reliable system (0% errors) versus unreliable system (30% errors). Time on task will modulate participants’ vigilance levels. Since the sense of control can vary depending on the reliability of the supervised automated system and the level of trust associated with it, we assessed it by measuring the control felt and used during decision-making (Potts & Carlson, 2019). Then, we explored its relationship with the brain correlates of supervisory activity. We assumed that supervision activity should be lower for the reliable system than for the unreliable system, especially during the second session, when the level of reliability is most internalized. This high level of reliability should boost participants’ trust and possibly lead to complacency toward the system. Consequently, task engagement and vigilance in the task should be reduced compared to the supervision of an unreliable system. According to the literature, this decrease in supervisory activity for the reliable system would result in lower fronto-central theta activity and higher alpha activity. Both were expected to intensify with time on task. These results were expected during both the system’s decision phase and the preceding obstacle exploration phase. Modulations in gamma and beta activity were also expected. In particular, higher gamma activity should be observed primarily for the unreliable system when participants have low confidence in the system (Oh et al., 2020). This activity would increase the allocation of attention and facilitate the integration and perceptual binding of sensory information (Fries et al., 2001; Friese et al., 2016; Ni et al., 2016; Ghiani et al., 2021) as well as the integration of sensory evidence (Donner et al., 2009). This would thus provide a reliable representation of the information needed for accurate prediction of directional choice and, ultimately, for effective detection of decision errors in an unreliable system. Conversely, beta activity should be reduced for the unreliable system (Oh et al. 2020). A decrease in beta activity is closely related to the preparation and execution of voluntary motor movements and this decrease is even greater when observing the action than when executing it (Koelewijn et al., 2008). Because the unreliable system requires trajectory corrections due to its numerous errors, it should therefore promote motor preparation in participants so they can react to these errors. This effect should be more pronounced at lower levels of vigilance. Finally, we will examine the relationship between supervisory brain activity and the level of felt control. Felt control partly determines the attention and engagement devoted to the task performed by the system and can be influenced by the system’s reliability as well as by the perceived usefulness of automation (Metcalfe & Greene, 2007; Ueda et al., 2021; Vantrepotte et al., 2023). This is especially true when the system’s actions or outcomes align with the participants’ intentions or goals (Dzindolet et al., 2002; Wen, Yamashita, & Asama, 2015). Therefore, when felt control is low, we expect participants to seek to control and monitor the system more (Wen & Haggard, 2018). Method Participants Fifty participants were randomly divided into two experimental groups, each distinguished by the reliability of the supervised system. We selected 25 participants per group based on the results of previous studies (e.g., Somon et al., 2022). We excluded five participants from all analyses for technical reasons, misunderstanding of the instructions or excessive signal noise. Therefore, the group using a ”reliable” (RS) system was composed of 22 participants (11 women, mean = 22.7 years; SD = 4.6 years) and the ones using an ”unreliable” (US) system included 23 participants (12 women, mean = 25.6 years; SD = 8.2 years). All participants had normal or corrected-to-normal vision, had no past or present psychiatric or neurological pathologies and were not taking any medications that affect the nervous system. The participants provided written informed consent and received 70€ compensation for taking part of the experiment. This research was approved by a local ethics committee (Comité Ethique pour la recherche Grenoble Alpes, n°CERGA-Avis-2021-25). Procedure The experiment comprised four experimental sessions including two 90-minute EEG sessions interspersed with two 30-minute training sessions. The two experimental EEG sessions were approximately performed one week apart (reliable: mean = 8.9 days; SD = 5.9 days; unreliable: mean = 8 days; SD = 5.3 days). Each session consisted of four blocks of 40 trials (14-17 min./block) and included questionnaires between blocks (see subjective measures). Prior to the first EEG session, participants performed familiarization tasks on 40 trials that were not used during the experiment. A debriefing was conducted after the last EEG session, during which the manipulated experimental conditions (i.e., system reliability) and the study objectives were explained. Experimental tasks We used node.js to program, collect the behavioral results and display the stimuli on an LCD screen located at 60 cm from the participant (33° x 27°) and whose refresh rate was 60 Hz. Responses were recorded using the keyboard and a mouse. Supervision task during EEG sessions Both EEG sessions were similar. The performed task was adapted from the DAA-Displays simulation engine (Masci & Muñoz, 2020) and consisted of an UAV supervision task, close to the one used by Somon and collaborators (2022). Participants had to supervise an UAV, always positioned in the center of the screen (in the center of a radar zone), moving through air traffic represented by obstacles. This UAV was equipped with an obstacle avoidance system whose function was to detect obstacles and implement trajectories to avoid them. Participants were asked to monitor the functioning of this system and to correct the implemented trajectories if they thought they might be incorrect (i.e., if they induced a conflict with one of the obstacles). Two levels of system reliability (reliable vs. unreliable) were tested between subjects. Depending on the participant group, 100% (reliable system) or 70% (± 4% maximum - unreliable system) of avoidance maneuvers were successful. For the unreliable system, the error rate varied by up to 4% between trial blocks, as did the error distribution to prevent any anticipation by participants during each block. Specifically, a block could contain between 66% and 74% of successful avoidance maneuvers. For each system, each trial lasted between 18.2s and 22.7s and took place in several stages (Figure 1): - Initialization : at the beginning of the trial, indicated by the UAV passing through a waypoint (represented by a white circle), the UAV was moving in a straight line for 2,000 ms to 4,000 ms. During this time period, eleven obstacles (primary and secondary) gradually appeared. The co-called primary obstacle was located on the UAV’s trajectory and had to be avoided. Secondary obstacles were positioned in equal numbers on either side of the primary obstacle, and could condition the system’s decision. Each obstacle corresponded to an aircraft represented by an arrow surrounded by a yellow circle. The distance between obstacles and their distribution varied between trials and were defined pseudo-randomly. - Conflict detection: once all the obstacles were visible, the system warned that a conflict was approaching by displaying the message ”CONFLICT AHEAD” and a yellow circle around the drone to avoid for 200 ms. - System decision: Then the simulation was frozen for 1,500 ms. During this time period, the UAV was surrounded by an orange circle and the ”AVOID” message and the direction taken by the system to avoid the obstacles (represented by an arrow pointing to the left or right) were displayed. In addition, a blue rectangle was displayed on the outer line of the compass to give another indication about the followed direction by the UAV. The UAV avoided obstacles by turning to the left for half the trials and to the right for the other half. During each trial, only one of the two possible directions was correct. - Participant response: The ”AVOID” message disappeared and a blue circle appeared around the UAV. From then on, if the participants judged that the chosen trajectory would lead to a collision with an obstacle, they could make the UAV change altitude. To do this, they had to press the ‘E’ key on the keyboard as quickly as possible. The phase ended either as soon as the participant responded or after 2,500 ms had elapsed without any response from the participant. - Avoidance and return to the initial trajectory: After the response phase was complete, the blue circle disappeared, and the UAV started moving again along the chosen obstacle-avoidance trajectory. If an obstacle was on this trajectory as a result of a system error and the participant did not correct it, a red circle would appear around the UAV for 1,000 ms when the obstacle was hit. If the participant made a correct altitude change, a green circle appeared around the UAV to indicate successful obstacle avoidance. If there was no obstacle on the chosen trajectory following a correct system decision, no feedback was displayed. After clearing the obstacles, the UAV returned to its initial trajectory and the next waypoint. This maneuver lasted 14.5 seconds. Figure 1.Timeline of a trial during the task performed in an EEG session. Supervision task during training sessions The participants conducted two training sessions to reinforce learning regarding the system’s reliability level and to instill confidence in the system based on this level of reliability. These sessions also increased the participants’ expertise in performing the task. The task performed during these sessions was virtually identical to that of the EEG sessions. The only differences were that (i) participants had to respond to each trial to either maintain the system’s trajectory or change altitude using separate buttons, and (ii) the simulation was never frozen, so participants responded while the UAV was in motion. Each training session included 40 trials. Measures and results recorded during these training sessions will not be detailed in this article. Familiarization tasks Before the first EEG session, participants completed two familiarization tasks. Each task consisted of 20 trials. In the first familiarization task, participants performed a protocol identical to that used in the training sessions. The number of system errors in this task depended on the participant’s experimental condition (i.e., 100% or 70%). The design of the second familiarization task was identical to the one performed during the EEG sessions, with a 50% error rate for both experimental groups, so that both groups begin the supervision task with an identical level of trust in the UAV. Measures and statistical analysis Subjective and behavioral measures Several questionnaires were used to assess variables that might affect the supervision activity. At the beginning of the experiment, participants rated their tendency to be complacent with an automated system using the Automation-Induced Potential Complacency scale (AICP-R; Merritt et al., 2019). Participants’ sleepiness levels were also assessed at the beginning of the experiment and after each block using the Karolinska Sleepiness Scale (KSS, Åkerstedt & Gillberg, 1990; on a Likert scale ranging from 1, “extremely alert”, to 9 “really sleepy”). During the experiment, after each block, we also measured trust in the UAV’s decisions, as well as their felt and used control (two aspects of the sense of agency, Potts & Carlson, 2019), on a visual analog scale (VAS, from 0 to 100), by asking respectively the following questions: 1) How confident are you in the system’s ability to make the right decision to avoid obstacles? (On a scale from 0, “not at all”, to 100, “completely”); 2) To what extent did you feel in control during the aircraft’s avoidance maneuvers (on a scale from 0 “not at all”, to 100 “completely”); and 3) To what extent was the avoidance task cognitively demanding (in terms of resource use)?” (On a scale from not at all” to “extremely”). Response times (RT) and error rates were recorded. However, only the error rates were analyzed because the RT were temporally biased by the imposed response time window and the small number of corrections made by the participants due to the system’s reliability. An error was identified when a participant failed to execute an avoidance when the system’s decision was erroneous, or when they executed an avoidance despite a correct system decision. All statistical analyses of these measures were performed in R (R Core Team, 2023). We used an analysis of variance (ANOVA) when the distribution of residuals did not violate the assumptions of homogeneity (Levene’s test, p < .05) and normality (Shapiro–Wilk test, p < .05, and by visual inspection using Q–Q plots). When the assumption of sphericity was violated, the Greenhouse–Geisser correction was applied. Post-hoc tests were conducted using the emmeans package with Holm correction for multiple comparisons. Partial eta-squared (η²ₚ) was reported as a measure of effect size. When these assumptions were not met, we applied the nonparametric Aligned Rank Transform Analysis of Variance (ART-ANOVA) using the ARTool package (Elkin et al., 2021; Wobbrock et al., 2011). Post-hoc tests (with Holm correction) were performed using the ART-C method, which is adapted for aligned and ranked data. Partial eta-squared was also used to indicate effect size. For ART-ANOVA analyses, partial eta-squared values were estimated from the F-statistics of the ANOVA on the aligned rank data. First, we wanted to ensure that the two experimental groups were not more inclined to trust the automated system before manipulating their trust level and did not differ in their sleepiness level before starting the task. The AICP-R and KSS scores were each analyzed using an ANOVA with the reliability level of the supervised system (reliable and unreliable) as a between-subjects factor. The influence of trust on supervisory activity was assessed by considering only the first two blocks of each experimental EEG session in order to limit the impact of other confounding factors such as decreased vigilance or motivation likely to be more present during the last blocks of each session. Similar analyses were performed for felt and used control measures. Differences were expected between the two experimental EEG sessions, due to an increase in participants’ knowledge of the system’s reliability (the two training sessions between the two EEG sessions also contributing to this). For each of these subjective measures (trust, sleepiness felt and used controls) assessed during the supervisory task, the mean score of the first two blocks of each session was calculated per participant and analyzed using the method ART with a linear mixed model considering the reliability level of the supervised system as a between-subjects factor and the experimental session (Session 1 and Session 2) as a within-subjects factor. We included participants as a random effect. The effect of vigilance level on trust and felt and used controls was only assessed during the second session, once the participants had a full understanding of the reliability level of the supervised system. The vigilance level was based on sleepiness score, and its effect was assessed through the time-on-task. We restricted the analyses of the effect of time on task to the comparison of Block 1 and Block 4 (of session 2), as this effect was expected to be maximal between these blocks. The mean scores of the subjective measures (sleepiness, trust, felt and used controls) across the session 2 were statistically analyzed using the method ART with a linear mixed model considering the reliability level of the supervised system as a between-subjects factor and the block (1 and 4) as a within-subjects factor. Participants was again included as a random effect. Behavioral data were analyzed using the same procedures as those used for subjective data. Electroencephalographic data Electroencephalographic (EEG) activity was continuously measured using an ActiCAP (Brain Products GmbH) which was equipped with 92 Ag/AgCl unipolar active electrodes. These electrodes were positioned according to the extended 10-20 system (i.e., the 5% system, Oostenveld & Praamstra, 2001). The ground electrode used for EEG data acquisition was that of ActiCAP, i.e., AFz. No reference had been used during the recording (related to the equipment used). Electrical activity related to blinking and vertical eye movements was recorded using electrodes from the ActiCap (TPP7h and TPP8h) positioned above and below the right eye on the median axis. Electrical activity related to horizontal eye movements was estimated from the EEG electrodes F9 and F10. Signal impedance was kept below 10 kΩ for all electrodes. The data were recorded and amplified using an ActiCHamp TM system (Brain Products, Inc.) and the LabRecorder software from Lab Streaming Layer (Kothe et al., 2014), with a sampling rate of 1,000 Hz and a resolution of 0.01 μV. All EEG data analyses were performed using the Fieldtrip Matlab toolbox (Oostenveld et al., 2011; R2019b; MathWorks, Inc., Natick, MA). Since no reference was used during the recording, we have therefore re-referenced the data relative to FCz for performing the following data preprocessing steps. The raw EEG data were band-pass filtered between 0.5 Hz and 80 Hz using a non-causal Butterworth bandpass filter. Then, a notch filter was applied between 48 Hz and 52 Hz. The data were then segmented into epochs ranging from 1,000 ms before trial onset to 1,000 ms after system decision. The epoch duration ranged from 5,700 ms to 7,700 ms, depending on the trial initialization duration (see Figure 1). All segments containing muscle artifacts and/or non-physiological artifacts were rejected offline after visual inspection. Signals from electrodes with too much noise were interpolated using spherical spline interpolation (Perrin et al., 1989) over the entire recording time. The signals were then re-referenced offline to the average of the EEG electrodes (after excluding TPP7h, TPP8h, TP9 and TP1011 1 Note that electrodes TP9 and TP10 have been excluded here because they were specifically relocated to the mastoids during recording for possible re-referencing of the data at these sites.) and down-sampled to 200 Hz offline. Artifacts related to eye movements (saccades and blinks) and cardiac activity were visually identified and manually removed after an Independent Component Analysis (ICA) was performed using the extended infomax algorithm. Only correct trials were included in the following analyses (i.e., trials in which participants gave correct responses regardless of system accuracy), with an average of 128 trials (SD = 19.7 trials). Data were then re-segmented into three interest periods: (1) the baseline period, ranging from 0 ms to 1,000 ms (with 0 ms marking the beginning of the initialization period); (2) the obstacle appearance period, ranging from -1,200 ms to -200 ms and (3) the system decision period ranging from 0 ms to 1,500 ms (with 0 ms marking the presentation of the system decision for the periods (2) et (3)). To limit edge effects, 750 ms (for the system decision period) or 1,000 ms (for the obstacle appearance period) were added to the ends of the epoch. The EEG data were analyzed in the time-frequency domain, between, 3 Hz and 80 Hz, using event-related spectral perturbation (ERSP) measures identified from a Morlet wavelet decomposition applied to each trial. For an optimal frequency resolution, wavelet cycles were fixed at 3 cycles at the lowest frequency (3 Hz) and increased linearly until they reached a maximum of 30 cycles for the highest frequency (80 Hz). The data were then corrected against a baseline period using the gain model procedure (Grandchamp & Delorme, 2011). The correction was applied differently depending on the studied factor during supervision activity: system reliability level (and the associated trust) and vigilance level. To study the influence of the reliability levels on supervisory activity during each period of interest (the obstacle appearance and system decision periods), we divided the spectral power at each time-frequency point by the average spectral power of the baseline period at the same frequency for each trial. Then, we applied a decibel transformation. This correction accounts for the person’s initial state (e.g., vigilance and attention) during the trial, and allows us to evaluate supervisory activity independently of the initial state. All trials were then averaged for each participant and each experimental condition. To study the influence of the vigilance levels on supervisory activity during each period of interest (the obstacle appearance and system decision periods), we applied the same procedure to each trial as above, but dividing by the mean baseline spectral activity of all trials. This procedure avoids eliminating the effects of time on task, and is therefore better suited to assessing vigilance levels (Arnau et al., 2021). All statistical analyses described in this paragraph were performed using cluster permutation tests (Maris & Oostenveld, 2007). The ecological and dynamic nature of the task may introduce greater variability in the measured EEG signals, particularly due to the sequence of cognitive processes involved. To limit this variability and the number of multiple comparisons, the time-frequency data were statistically analyzed by frequency band using the following four classical frequency bands based on the literature: Theta 3-8 Hz; Alpha 8-14 Hz; Beta 15-30 Hz; and Gamma 30-80 Hz. To investigate the effect of trust on brain correlates associated with supervisory activity, multiples analyses were restricted using the following procedure: For each interest time period (obstacle appearance and system decision periods), the mean time-frequency spectral activity (of the first two blocks) of the first experimental session was subtracted from that (of the first two blocks) of the second experimental session per participant for each reliability level to assess the interaction between reliability level and session. This difference was then statistically analyzed depending on the reliability level of the supervised system (between-subject analysis) for each frequency band. For each interest time period, when the interaction was statistically significant for a given frequency band and a given time-frequency window, the initial spectral data from the two separate sessions for each reliability level (within-subject analysis) and those from the two reliability levels for each session (between-subject analysis) were statistically compared pairwise. These within- and between-subject analyses were performed only in the frequency band and time window in which a cluster was identified as statistically significant during the interaction analysis22When a significant result was observed, we also compared the baseline periods depending on the reliability level of the supervised system, considering the frequency band and electrodes selected in the cluster analysis. This additional analysis was conducted to ensure there were no pre-existing differences between the groups at baseline. We averaged the brain activity recorded during the first two blocks of Session 2, focusing on the frequency bands and electrodes that were identified in the cluster analyses. We then performed ANOVAs using reliability level as the between-subjects factor. None of these analyses yielded statistically significant results, Fs 0.1. Conversely, when the interaction was not statistically significant for a given frequency band, the considered time-window for analyzing the obstacle appearance and system decision periods was shortened. This minimized variability and the number of multiple comparisons before proceeding cluster-based analyses again. The new time windows were defined as follows: we computed for each participant the mean difference values of time-frequency activity between the two sessions. We averaged these values by reliability level and calculated the difference between the two reliability levels. This yields a mean time-frequency difference between the different conditions. Next, we visually identified the peak power difference and defined its temporal bounds, which determined the position and width of the new analysis time-window for cluster permutation tests in the initially insignificant frequency bands. To evaluate the effect of vigilance level on brain correlates associated with supervisory activity, we replicated the analyses performed to assess the influence of trust, but with two differences: as with the subjective and behavioral data, we only considered the second session and took the different blocks as an intra-subject factor (1 and 4). Correlations between subjective and EEG measures The aim of these correlational analyses was to explore the relation between SoA and supervisory activity for each reliability level of the automated system. We extracted the average spectral power in the selected frequency band, time window, and electrodes for each significant cluster identified from analyses of the influence of trust and vigilance as a function of system reliability level, and assessed the correlations between these neural activities and felt control. Results First, we confirmed that there was no statistical difference between the two experimental groups in their tendency to be complacent toward an automated system or their sleepiness level before the two EEG sessions, F < 1. Influence of trust level on supervisory activity The reliability effect was analyzed based only on the first two blocks of each experimental EEG session in order to limit the impact of confounding factors such as decreased alertness or motivation, which are likely to be more prevalent in the last blocks of each session. Subjective and behavioral data The results of the subjective and behavioral data are illustrated in Figures 2 and 3 respectively. Sleepiness - No main effects of the system’s reliability level or session, nor any interaction effect, were observed on the participants’ subjective sleepiness level. Trust - A higher level of trust was observed for the supervision of a reliable system (90.2) compared to an unreliable system (63.8; F(1, 43) = 41, p < 0.001, \(\eta_{p}^{2}\) = .49). No interaction effect between the level of reliability and the experimental session was observed on the trust scores, F < 1. Felt control – The participants reported a higher level of control during the second session (1: 55; 2: 61, F(1, 43) = 4.26, p < 0.05, \(\eta_{p}^{2}\) = .09). No main effects of the system’s reliability level or interaction effect were observed on felt control. Used control - Used control was significantly modulated by the reliability level (F(1, 43) = 6.5, p < 0.05, \(\eta_{p}^{2}\)= .13), the session (F(1, 43) = 213.4, p < 0.001,\(\eta_{p}^{2}\) = .33) and the reliability x session interaction (F(1, 43) = 4.09, p < 0.05, \(\eta_{p}^{2}\) = .09). Mean comparisons showed that the level of used control was higher for the unreliable system (compared to reliable system) during session 1 (US: 48; RS: 30, t(62.5) = 2.58, p < .05) and session 2 (US: 31; RS: 20, t(62.5) = 2.4, p < .05). It was also higher in the first session than in the second session for the reliable (+10, t(43) = 2.77, p < .05) and unreliable system (17 increase, t(43) = 3.12, p < .05). The difference between the two levels of reliability during Session 1 was greater than that observed during Session 2 (t(43) = 2, p < .05). Figure 2. Average scores of (A) Trust, (B) Sleepiness, (C) Felt control, and (D) Used control depending on the reliability level of the system, and the experimental session. Error bars represent the standard error. Performance – The accuracy was significantly modulated by the reliability level (F(1, 43) = 48.6, p < 0.001,\(\eta_{p}^{2}\) = .53), the session (F(1, 43) = 37.9, p < 0.001, \(\eta_{p}^{2}\) = .43) and the reliability x session interaction (F(1, 43) = 9.2, p < 0.01, \(\eta_{p}^{2}\) = .18). The accuracy was better for the reliable system compared to unreliable system in the session 1 (RS: 94%; US: 77%; t(66.7) = 5.8, p < .0001) and the session 2 (RS: 97%; US: 86%; t(66.7) = 6, p < .0001). A significant increase in performance was observed between the two sessions for both the ’reliable system’ group (3% increase, t(43) = 4.2, p < .001), and for the unreliable system (9% increase; t(43) = 3.8, p < .001). This increase was significantly greater for unreliable system compared to reliable system (t(43) = 3, p < .01). Figure 3. Mean percentage of correct responses depending on the reliability level, and the experimental session. Error bars represent the standard error. Electroencephalographic data Obstacle appearance period Analysis over the entire period of the alpha and gamma activities revealed no interaction between reliability level and experimental session (p > 0.1). Analyses focusing on the maximum power difference between conditions (using the procedure outlined in the Methods section) for alpha (p > 0.05) and gamma activities (p > 0.05) gave a similar result. Concerning the theta activity, an interaction effect between reliability level and session specifically in the 3-6 Hz frequency band (p < 0.05) was only observed by focusing on the maximum power between conditions. The decomposition of the interaction revealed two results. First, the two reliability levels differed statistically only in session 2 (p < 0.05). During this session, participants supervising a reliable system had greater theta activity (5-6 Hz) than those monitoring an unreliable system, between -780ms and -630ms in an 18-electrode frontal cluster (Figure 4). A significant increase in theta activity between session 1 and 2 for the ”reliable system” group (p < 0.01) may have contributed to this difference. Indeed, for this group, theta activity (4-6 Hz) was significantly higher during session 2 than during session 1 (between -875ms and -540ms) in a larger fronto-temporal cluster (26 electrodes). Figure 4. (A) Time course of the mean spectral power of the theta activity (5-6 Hz) recorded in an 18- electrode frontal cluster depending on the experimental session and the reliability level of the system during the obstacle appearance and baseline periods. The black line represents the time window where the statistical difference was significant (-780 to -630ms; 0: system decision). (B). Topographies representing the mean theta activity within the 18-electrode frontal cluster (black crosses). Regarding beta activity, analysis of the entire period revealed an interaction effect between experimental conditions in the 21-30Hz frequency band (p < 0.05). Its decomposition showed than only session 2 differed between the two reliability levels, with significantly lower beta activity (25-28Hz) for the “unreliable system” group compared to the “reliable system” group, between -505 ms and -380 ms, in a 24-electrode occipito-parietal cluster (Figure 5). A decrease in beta activity (21-30Hz) between session 1 and session 2 for the “unreliable system” group (p < 0.05) may also have contributed to this difference. Indeed, for this group, lower beta activity was also identified in session 2 compared to session 1 (between -635ms to -365ms) in a widespread cluster covering the whole scalp (42 electrodes). Figure 5. (A) Time course of the mean spectral power of the beta activity (25-28 Hz) recorded in a 24- electrode occipito-parietal cluster depending on the experimental session and the reliability level of the system during the obstacle appearance and baseline periods. The black line represents the time window where the statistical difference was significant (-505 to -380ms; 0: system decision). (B). Topographies representing the mean beta activity in the 24-electrode occipito-parietal cluster (black crosses). System decision period Analysis of the theta and beta activities throughout the system decision period revealed no interaction between reliability level and experimental session (p > 0.1 and p > 0.05 respectively). Analyses focusing on the maximum power difference between conditions in the theta (p > 0.05) and beta frequency bands (p > 0.05) gave a similar non-significant result. Regarding alpha activity, only the analysis focusing on the maximum power difference revealed an interaction between the two variables (p < 0.05) (Figure 6). Its decomposition showed than only session 2 differed between the two reliability levels, with higher alpha activity for the “unreliable system” group than for the “reliable system” group between 550ms and 995ms in a 31-electrode occipito-parietal cluster (p < 0.01). Figure 6. (A) Time course of the mean spectral power of the alpha activity (8-14 Hz) recorded in a 31- electrode occipito-parietal cluster depending on the experimental session and the reliability level of the system during the system decision and baseline periods. The black line represents the time window where the statistical difference was significant (550 to 995 ms; 0: system decision). (B). Topographies representing the mean alpha activity in the 31-electrode occipito-parietal cluster (black crosses). Regarding gamma activity, analysis of the entire period revealed an interaction effect in the 53-80Hz frequency band (p < 0.05). Its decomposition showed than only session 2 differed between the two reliability levels, with significantly greater gamma activity (56-75Hz) for the “reliable system” group than the ”unreliable system” group between 0 ms and 250 ms in a large occipital, frontal and right temporal cluster of 50 electrodes (p < 0.05; Figure 7). A decrease in gamma activity (63-78Hz) between session 1 and session 2 for the “unreliable system” group (p < 0.05) may have contributed to this difference. Indeed, for this group, lower gamma activity was also identified in session 2 compared to session 1 (between 0ms and 400ms) in large cluster spanning the entire scalp (51 electrodes). Figure 7. (A) Time course of the mean spectral power of the gamma activity (56-75 Hz) recorded in a 50-electrode large occipital, frontal, and right temporal cluster depending on the experimental session and the reliability level of the system during the decision and baseline periods. The black line represents the time window where the statistical difference was significant (0 to 250ms; 0: system decision). (B). Topographies representing the mean alpha activity in the 50-electrode cluster (black crosses). Influence of vigilance level on supervisory activity As a reminder, the influence of vigilance level was examined only from the second session, after participants had fully learned the reliability of the supervised system. Analyses focused on comparing blocks 1 and 4 (of session 2), as the time on task effect was expected to be most pronounced between these two blocks. Subjective and behavioral data The results of the subjective and behavioral data obtained are shown in Figures 8 and 9 respectively. Sleepiness - Sleepiness also increased over time between the blocks 1 (5.73) and 4 (6.44) (F(1, 43) = 4.9, p < 0.05,\(\eta_{p}^{2}\) = .1), regardless of the system reliability level. Sleepiness was not influenced by either system reliability level or the reliability x block interaction, Fs < 1. Trust - Trust was significantly modulated by reliability (F(1, 43) = 39.4, p < 0.001, \(\eta_{p}^{2}\) = .48), blocks (F(1, 43) = 29.7, p < 0.001, \(\eta_{p}^{2}\) = .41), and the reliability x block interaction (F(1, 43) = 14.3, p < 0.001,\(\eta_{p}^{2}\) = .25). Mean comparisons confirmed a higher trust for the reliable system (compared to unreliable system) during the first (RS: 92; US: 62, t(50.5) = 6.6, p < .0001) and last block (RS: 93; US: 69, t(50.5) = 6, p < .0001). Only participants supervising an unreliable system increased their trust over time (1: 62; 4: 69, t(43) = 2.6, p < .05). Felt control - No main effects of the system’s reliability level or blocks, nor any interaction effect were observed on the felt control and the used control. Used control – Used control was higher when the system was unreliable (US: 29; RS: 21, F(1, 43) = 5.5, p < 0.05,\(\eta_{p}^{2}\) = .11), No main effects of block nor interaction effect were observed. Figure 8. Mean scores of: (A) Trust, (B) Sleepiness, (C) Felt control, and (D) Used control depending on the reliability level of the system, and the block of the session 2. Error bars represent the standard error. Performance - The accuracy was significantly modulated by the reliability level (F(1, 43) = 33.5, p < 0.0001,\(\eta_{p}^{2}\) = .44), the block (F(1, 43) = 28.3, p < 0.0001, \(\eta_{p}^{2}\) = .42) and the reliability x block interaction (F(1, 43) = 17.3, p < 0.0001, \(\eta_{p}^{2}\) = .29) (Figure 9). Better performance was observed for participants supervising a reliable system compared to those monitoring an unreliable system in the blocks 1 (RS: 97%; US: 84%; t(55.2) = 6.2, p < .0001) and 4 (RS: 97%; US: 90%; t(55.2) = 4.9, p < .0001). Only participants supervising an unreliable system were more accurate in the last block compared to the first block (6% increase; t(43) = 3.7, p < .01). The difference between the two levels of reliability in block 1 was greater than that observed in block 4 (t(43) = 4.2, p < .001). Figure 9.Mean percentage of correct responses depending on the reliability level, and the block of the session 2. Error bars represent the standard error. Electroencephalographic data Obstacle appearance period We observed no interaction effect in the theta, beta and gamma activities (p > 0.1), even when focusing on the maximum power difference between conditions (p > 0.1). Analysis of alpha activity revealed a significant interaction (p 0.05). In contrast, this alpha activity was significantly reduced in the last block (compared to the first block) only when the system was unreliable. This decrease was observed in an extended 51-electrode cluster over the same period (between -1,110ms and -775ms) and the whole alpha frequency band (p < 0.05; Figure 10). Figure 10. (A) Time course of the mean spectral power of the alpha activity (8-14 Hz) recorded in a 51- electrode cluster depending on the reliability level of the system and block during the obstacle appearance and baseline periods. The black line represents the time window where the statistical difference was significant (-1,100 to -775ms; 0: system decision). (B). Topographies representing the mean alpha activity in the 51-electrode cluster (black crosses). System decision period In contrast to the analyses conducted during the obstacle appearance period, analysis of the theta activity over the entire system’s decision period revealed an interaction effect between reliability level and session 2 blocks over the entire frequency band (p < 0.05; Figure 11). Its decomposition did not reveal any differences between the two reliability levels during the first or last block. However, a decrease in theta activity between blocks 1 and 4 was observed only for the group supervising an unreliable system between 0ms and 415ms in a very large cluster widely distributed over the whole scalp (61 electrodes; p 0.1. Figure 11. (A) Time course of the mean spectral power of the theta activity (3-8 Hz) recorded in a 61- electrode cluster depending on the reliability level of the system and the block during the system decision and baseline periods. The black line represents the time window where the statistical difference was significant (0 to 415ms; 0: system decision). (B). Topographies representing the mean theta activity in the 61-electrode cluster (black crosses). We also observed an interaction effect (p < 0.05) between the two variables on the entire alpha frequency band. Its decomposition showed a decrease in activity over time only when the system was unreliable (p < 0.05). This decrease occurred over the same period and frequency band in an 84-electrode cluster widely distributed over the whole scalp (Figure 12). No interaction effect was observed on the beta and gamma activities, even when the analysis windows were reduced (ps > 0.1). Figure 12. (A) Time course of the mean spectral power of the alpha activity (8-14 Hz) recorded in an 84- electrode cluster depending on the reliability level of the system and block during the system decision and baseline periods. The black line represents the time window where the statistical difference was significant (810 to 1,500ms; 0: system decision). (B). Topographies representing the mean alpha activity in the 84-electrode cluster (black crosses). Correlations between subjective and EEG measures We observed a positive correlation only between alpha activity measured within the cluster identified during the system decision period in session 2 (occipito-parietal cluster, 550–995ms) and control felt when the system was reliable (Pearson r = 0.527; p < 0.05; Figure 13). Figure 13. Correlation between alpha activity and felt control depending on reliability level in session 2 during the system decision. Discussion Autonomy in complex civil or military systems (e.g., UAVs, aircraft, automobiles) continues to grow in complexity. This evolution has radically changed the role of system operators. As a result, new issues have arisen, such as the emergence of the OOTL phenomenon and associated performance problems, hampering the supervision of automated systems. The aim of this study was to characterize the influence of trust and vigilance levels, two of the main factors at the origin of the OOTL phenomenon, on the cerebral correlates of supervisory activity of an automated system during an ecological task. To do this, two groups supervised an autonomous UAV avoiding obstacles with perfect or poorer reliability. Supervisory activity was analyzed during the obstacle appearance and system decision periods. For both periods, we first analyzed the effect of trust, which we manipulated through the system’s reliability level during the first two blocks of each session to limit the influence of vigilance or motivation levels. We then evaluated the evolution of brain activity over time during the second experimental session to assess the influence of vigilance level on supervisory activity. The influence of trust on supervisory activity We had postulated that participants supervising a reliable system would become complacent toward the system and disengage from their supervisory activity. In contrast, participants supervising an unreliable system would remain engaged. Contrary to our hypotheses, our results suggest that all participants supervised their system, regardless of its reliability level. Nevertheless, the reliability level of the system and the resulting trust would affect supervisory activity differently, as evidenced by the distinct observed effects on brain activity between the two systems. These results suggest that different cognitive processes are involved depending on the supervised system. Furthermore, the effects of the system’s reliability level on brain activity were observed only during the second session, despite a distinct level of trust between the two systems regardless of the session. Although a difference in trust between the two systems may be apparent as early as the first session, the cognitive processes involved in supervising both systems during this session are likely similar since both groups are learning about the reliability of their respective systems. The results of the performance data illustrate this learning. Integrating the system reliability level effectively in the second session had distinct effects on the neurocognitive processes engaged during supervision. We describe these effects in the following discussion. First, the behavioral data showed that despite the reliable system, participants intervened during the second session given the few errors they made. At the cerebral level, a greater theta activity was observed in frontal regions during the second session for reliable system compared to the unreliable system. This activity increased from the first to the second session. An increase in frontal theta activity indicates the recruitment of cognitive control and, more broadly, the engagement of cognitive effort (Cavanagh et al., 2010). It is typically observed when preparing to process task-relevant information or when a behavioral adjustment is required (Eisma et al., 2021; Kaiser et al., 2022). According to Eisma and collaborators (2021), this theta activity is more important when the information provided is not indicative of the response to be given (compared to when information that leaves no uncertainty). In our task, obstacles appeared during the obstacle appearance period. An increase in theta activity may indicate that participants prepared to process the system decision, which was presented immediately afterwards. Participants then began to integrate information to identify the trajectory that the UAV should follow and validate (or reject) the system’s decision. Increases in theta activity are often been reported during information encoding (Klimesch et al., 1997) or when attentional demands increase (Sauseng et al., 2010). For reliable system (vs. unreliable system), the increase in theta activity observed during the obstacle appearance period was followed by significantly lower alpha activity in the parieto-occipital regions during the system decision period. Reduced alpha activity in these regions is associated with increased attention and the facilitated processing of task-relevant information (Jensen & Mazaheri, 2010). Desynchronization of alpha activity has been observed when attention is directed toward a salient stimulus (Pezetta et al., 2018) or toward the task following an error or an attentional lapse, for example (Van Driel et al., 2012). Despite the absence of errors in the reliable system, these results suggest that participants supervising this system paid closer attention and exert more effort into supervising both the task and the system, even though it was not necessary. Although this increase in engagement contradicts our hypotheses, the positive influence of trust on performance has already been observed, when it is not in excess and when trusting allowed participants to perform (Blais et al., 2019). The novelty of the task for participants may also play a role. Recently, it was shown that new users of an automated system were as engaged as with a manual system despite the automated system’s high reliability (McDonnell et al., 2021). In our study, most participants were students who were unfamiliar with this type of task, and more accustomed to laboratory tasks. Therefore, the novelty possibly increased their engagement. The interpretation of the decrease in alpha activity for the reliable system must be nuanced with regard to the positive correlation we observed for this system between alpha activity and felt control. This correlation indicates that the higher the control felt, the higher the alpha activity, and therefore the less attention they paid to the system’s decision. The lower alpha activity observed for the reliable system, linked to more attention paid to its decisions, would be associated with a lower felt control over the task. The fear of missing a very rare potential error in the highly reliable system could justify this low felt control. Therefore, in the case of a reliable system, it appears that participants re-engaged in supervision before the system’s decision was presented and paid closer attention to it due to its relevance for performance, particularly when the control felt was lower. As expected, the different measures also illustrate an engagement in supervising an unreliable system. However, different brain markers were reflected, suggesting the involvement of cognitive processes other than those observed in participants supervising a reliable system. Specifically, beta activity during the obstacle appearance period and gamma activity during the system decision period were lower for the unreliable system than for the reliable system. These differences between the two groups in the second session were related to a decrease of these activities between the first and second sessions for the unreliable system. The decrease in beta activity for the unreliable system during the obstacle appearance period was mainly observed in a cluster of electrodes located in the motor cortex, and especially in the parieto-occipital regions. According to the literature, this decrease is associated with preparing for voluntary motor movements (Koelewijn et al., 2008). In our context, it could be considered as an improvement in motor preparation. Other studies have also reported a proportional link between reductions in parieto-occipital beta activity and increases in the SoA (Bu-Omer et al., 2021). The SoA refers to the sense of initiating and controlling actions or thoughts to influence events in the outside environment. As expected, supervising an unreliable system that requires trajectory corrections due to numerous system errors would promote this motor preparation and the SoA over the response to be given (higher used control observed). This reduction in beta activity could also be driven by non-motor cognitive processes (Engel & Fries, 2010; Hanslmayr, Staudigl, & Fellner, 2012), such as encoding (Hanslmayr, Spitzer, & Bäuml, 2009). Recently, it has been proposed that the desynchronization of (alpha)/beta activity corresponds to an ”information processing proxy”. This is thought to reflect a decrease in correlated neuronal noise, which would increase the signal-to-noise ratio of a stimulus-evoked neuronal response, facilitating its processing (Hanslmayr et al., 2012; Griffiths et al., 2019).The observation of a reduced beta activity for the unreliable system during the obstacle appearance period could reflect anticipation-oriented supervision with facilitated information gathering to define the optimal trajectory in advance, allowing for the most effective reaction as soon as the system’s decision appears. This hypothesis would be support by a reduced error rate and level of control for the unreliable system. And this process would be particularly reinforced in session 2 compared to session 1, given the reduction observed in beta activity between session 1 and session 2 particularly for the unreliable system. The decrease in beta activity observed during the obstacle appearance period for unreliable system was followed by a decrease in gamma activity during the system decision period. This decrease was located in a large occipital, frontal and right temporal cluster. Depending on its topography and the gamma frequency band analyzed, several functional roles have been attributed to gamma activity. Notably, it is involved in allocation of attention and working memory (Fries et al., 2001; Jensen, Kaiser & Lachaux, 2007; Roux et al., 2012), promoting information acquisition. Frontal or parietal gamma activities have also been associated with maintaining information from multiple sensory modalities (Kaiser et al., 2009; Haegens et al., 2010; Roux et al., 2012). Posterior gamma activity in sensory areas is believed to play a role in integrating sensory information (Friese et al., 2016; Ni et al., 2016). It may be a central mechanism for the spatiotemporally binding sensory information to generate an unified perceptual representation and sensory awareness (Ghiani et al., 2021; Engel & Singer, 2001), which is necessary for the correct identification of the direction to follow in our task context. Based on this literature and by considering the reduced beta activity (reflecting better motor preparation and information integration) during the obstacle appearance period for the unreliable system, the reduced gamma activity observed in a large cluster of the right occipital, frontal, and temporal regions during the system’s decision period would reflect a reduction in attention paid to the unreliable system’s decision and in processing of information during this period. Due to the unreliable system’s numerous errors and reduced trust, participants would primarily base their choice to validate or reject the system’s decision on their previous analysis of the situation rather than on their trust in the system’s decision. For the unreliable system, most of the information would thus be gathered and integrated when the obstacles appear. The effectiveness of this prior information gathering would enable participants to identify the appropriate motor trajectory, as illustrated by the reduced beta activity. This would facilitate the validation of the system’s decision, as illustrated by the behavioral data, the level of control used, and reduce attention and information gathering during the system’s decision period (reflected by the reduced gamma activity). These results demonstrate that the effect of trust, and particularly complacency, cannot be reduced to a simple engagement/disengagement dynamic. Instead, it reflects differences in supervision strategies based on the system’s reliability and, consequently, the relevance of the information it provides. Influence of vigilance level on supervision activity The second objective of this study was to assess how the level of vigilance influenced the correlates of supervisory activity depending on the participants’ level of trust. Only the supervision of the unreliable system has been affected by time on task. We observed a decrease with time on task in theta and alpha activities during the system’s decision period, as well as lower alpha activity during the obstacle appearance period. A decrease in theta activity over time during the supervision activity may indicate a mental fatigue (Arnau et al., 2021). This mental fatigue could be explained by the task requiring constant supervision of an unreliable system, as well as by participants becoming sleepier as they spend more time on the task. Nevertheless, a decrease in alpha activity over time would indicate increased vigilance and attention to the task. Taken together, these results suggest that this reduction in alpha activity may indicate an attempt to maintain attention on the supervisory task despite the presence of mental fatigue (Chuang et al., 2018). The effect of time on task could also reflect a learning process, whereby performance improves over time for the unreliable system without the control also increasing. Enhanced attention (as shown by decreased alpha activity during the obstacle appearance and system decision periods) coupled with a learning effect over time could reduce cognitive engagement (as shown by decreased theta activity, Arnau et al., 2021) during the system decision period. Implication of the sense of agency on supervisory activity Our findings confirm the existence of an inverse relationship between sense of control and mobilized resources. Currently, the literature debates how effort influences the sense of agency. Some studies suggest that physical or cognitive effort enhances it (Demanet et al., 2013; Minohara et al., 2016; Sidarus & Haggard, 2016), while others highlight the positive impact of fluency and ease in executing or selecting actions (Chambon, Sidarus, & Haggard, 2014). Our data reveal an opposing dynamic between felt control and used control: felt control decreases over time, while used control increases. This trend suggest that participants experience a stronger SoA (control felt) when resources (control used) are depleted. These observations support the hypothesis of a negative relationship between effort and the SoA. The observed increase in felt control during the second session for all participants could be explained by improved system predictability. Several studies have indeed shown that increased predictability enhances the control felt, particularly if it improves performance (Synofzik et al., 2008; Wen, Kuroki, & Asama, 2019; Ueda, Nakashima, & Kumada, 2021; Vantrepotte et al., 2023). During the second session, participants had developed task expertise and integrated the system’s reliability level. This allowed both groups to more accurately anticipate the UAV’s trajectory. When the UAV was reliable, performance was highly predictable because the system made no errors. Participants therefore knew they could rely on this reliability to enhance their own performance. With a less reliable UAV, their expertise enabled them to more easily detect potential trajectory errors. This expertise was reflected in the reduction of used control and improved performance over time, as well as the adoption of distinct supervision strategies, as evidenced by the impact of trust on neural correlates. Participants who supervised a reliable system showed increased engagement just before and during the system’s decision. In contrast, those who supervised an unreliable system appeared to pay less attention to its decisions. Thus, increased expertise improved their supervision efficiency and strengthened their sense of control. Finally, we observed that the decrease in alpha activity for the reliable system was positively correlated with felt control. This correlation suggests that the greater the sense of control participants had, the greater their alpha activity, and thus the less attention they paid to the system’s decision. One possible explanation is that participants who relied more on the automated system’s choice (indicated by a significant decrease in alpha activity) integrated less information during obstacle exploration, particularly during the conflict detection phase. This could result in a lower sense of control compared to participants who anticipated the correct decision and compared their choice to that of the system, thereby increasing the predictability of the outcome. Increased attention to the system’s choice may therefore reflect a search for control due to a lack of upstream supervision of the system’s decision (Wen & Haggard, 2018). However, we did not observe any correlation between theta activity in the identified cluster during the conflict detection phase and the sense of agency. The link between a sense of control and supervisory activity was not observed when the system was unreliable. A SoA is generated by integrating several cues, which can be internal (e.g., effort, beliefs and expectations) or external (e.g., perceived information and rewards). These cues are integrated and weighted to generate a SoA (Moore and Fletcher, 2012). When the system was unreliable, the control felt likely stemmed not only from the system’s performance but also from the participants’ own actions and performance. An essential difference between the two groups was the obligation to monitor the unreliable system to correct it. Therefore, when intervention is necessary, the sense of control is not necessarily linked to the supervisory activity. Limits The main limitation of this study is the lack of evidence for the OOTL phenomenon, as defined in the literature, particularly for reliable systems. Despite the high level of trust, the results showed maintenance of supervision markers and an absence of degradation of these markers over time. This is potentially linked to a high level of involvement of the participants. Future work addressing the evidence of the OOTL phenomenon using a more dynamic approach would be interesting. However, our findings of distinct supervision patterns between the two levels of system reliability in our study raise questions about how the concept of OOTL should be characterized in this type of more ecological task. The second limitation is based on the small number of control measures used to objectively assess the participants’ state and supervisory activity independently of brain measurements. To maximize the probability of the OOTL phenomenon occurring, this study limited the motor requirements (no performance measurement on each trial; participants only pressed in case of trajectory correction) and subjective measures (e.g., control felt and used at the block level, not trial level). Using implicit performance measurements without the active participation of the participants in future studies would improve the interpretation of the results and confirm our conclusions. The absence of subjective measures of felt and used control at each trial prevented us from linking them precisely to measured brain activity. However, it is possible that the sense of control remained consistent from one trial to the next during the second session, during which we observed a correlation between alpha activity and felt control. Participants had integrated that the system’s reliability level and could therefore perfectly predict the outcome of each trial. In future work, considering other cerebral measures from EEG signal, such as brain connectivity, could help improve our understanding of the cognitive processes engaged in our study. For instance, a study by Huang and collaborators (2021) showed that a lack of trust in a system is associated with increased effective connectivity between regions of the default network and the prefrontal cortex, as well as a more complex brain network than when trust is high. In this study, participants performed a subtask of the Air-Force Multi-Attribute Task Battery (AF-MATB; Miller et al., 2010) that involved monitoring several gauges. According to the authors, the observed connectivity modulations would reflect an increase in cognitive load induced by low trust (Huang et al., 2021). Given the variability inherent in dynamic ecological tasks such as ours of the cognitive processes involved and their temporality, using other measurement tools could be relevant for better characterizing and dissociating these processes. In our study, participants’ visual exploration paths and the information they extracted from different parts of the screen (e.g., relative position of obstacles and the UAV’s trajectory) were essential to their decision-making. Using EEG and eye-tracking together would have improved our understanding of the cognitive processes and strategies involved, as well as their temporality based on oculomotor activity (e.g., Guérin-Dugué et al., 2018). Conclusion This study aimed to analyze how trust and vigilance affect brain activity when supervising an automated system during a realistic task. The effect of reliability (and associated trust) levels was observed across all frequency bands. The results showed that participants continued to supervise the automated system regardless of its reliability, but in different ways. However, supervision of a reliable system was influenced by the SoA. This confirms the importance of this factor in the OOTL phenomenon and calls into question the necessity of promoting a high level of control in reliable systems in all cases. This supervisory activity changed over time only when the system was unreliable. These results improve our understanding of how the factors behind the OOTL phenomenon influence our supervision of an automated systems, and provide insights that could help detect its occurrence in real time. Acknowledgements We would like to thank Marin Le Guillou for his work on the development of the obstacle avoidance task. Data Availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Funding information This work has been partially supported by MIAI@Grenoble Alpes, (ANR-19-P3IA-0003) and Carnot Cognition Institute. Conflict of interest The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Author Contributions Mike Salomone: conceptualization, methodology, software, investigation, data curation, formal analysis, writing – original draft, review and editing. Bruno Berberian: conceptualization, methodology, writing – review and editing. Aurélie Campagne: conceptualization, methodology, formal analysis, funding acquisition, project administration, supervision, writing – original draft, review and editing. Reprints Requests for reprints should be sent to Mike Salomone ( [email protected] ). References 1. Åkerstedt, T., & Gillberg, M. (1990). Subjective and objective sleepiness in the active individual. International Journal of Neuroscience , 52 (1‑2), 29‑37. https://doi.org/10.3109/00207459008994241 Arnau, S., Brümmer, T., Liegel, N., & Wascher, E. (2021). Inverse effects of time-on-task in task-related and task-unrelated theta activity. Psychophysiology , 58 (6), e13805. https://doi.org/10.1111/psyp.13805 Berberian, B. (2019). Man-Machine teaming: A problem of Agency. IFAC-PapersOnLine , 51 (34), 118‑123. https://doi.org/10.1016/j.ifacol.2019.01.049 Berberian, B., Gouraud, J., Somon, B., Sahai, A., & Le Goff, K. (2017). My brain is out of the loop: A neuroergonomic approach of OOTL phenomenon. In D. D. Schmorrow & C. M. Fidopiastis (Éds.), Augmented Cognition. Neurocognition and Machine Learning (Vol. 10284, p. 3‑18). Springer International Publishing. https://doi.org/10.1007/978-3-319-58628-1_1 Berberian, B., Sarrazin, J.-C., Blaye, P. L., & Haggard, P. (2012). Automation technology and sense of control: A window on human agency. PLOS One , 7 (3), e34075. https://doi.org/10.1371/journal.pone.0034075 Blais, C., Ellis, D. M., Wingert, K. M., Cohen, A. B., & Brewer, G. A. (2019). Alpha suppression over parietal electrode sites predicts decisions to trust. Social Neuroscience , 14 (2), 226‑235. https://doi.org/10.1080/17470919.2018.1433717 Borghini, G., Astolfi, L., Vecchiato, G., Mattia, D., & Babiloni, F. (2014). Measuring neurophysiological signals in aircraft pilots and car drivers for the assessment of mental workload, fatigue and drowsiness. Neuroscience & Biobehavioral Reviews , 44 , 58‑75. https://doi.org/10.1016/j.neubiorev.2012.10.003 Bu-Omer, H. M., Gofuku, A., Sato, K., & Miyakoshi, M. (2021). Parieto-occipital alpha and low-beta EEG power reflect sense of agency. Brain Sciences, 11 (6), 743. Cavanagh, J. F., Frank, M. J., Klein, T. J., & Allen, J. J. B. (2010). Frontal theta links prediction errors to behavioral adaptation in reinforcement learning. NeuroImage , 49 (4), 3198‑3209. https://doi.org/10.1016/j.neuroimage.2009.11.080 Chambon, V., Sidarus, N., & Haggard, P. (2014). From action intentions to action effects: How does the sense of agency come about? Frontiers in Human Neuroscience , 8 . https://doi.org/10.3389/fnhum.2014.00320 Chuang, C.-H., Cao, Z., King, J.-T., Wu, B.-S., Wang, Y.-K., & Lin, C.-T. (2018). Brain electrodynamic and hemodynamic signatures against fatigue during driving. Frontiers in Neuroscience , 12 , 181. https://doi.org/10.3389/fnins.2018.00181 Demanet, J., Muhle-Karbe, P. S., Lynn, M. T., Blotenberg, I., & Brass, M. (2013). Power to the will: How exerting physical effort boosts the sense of agency. Cognition, 129 (3), 574-578. https://doi.org/10.1016/j.cognition.2013.08.020 Donner, T. H., Siegel, M., Fries, P., & Engel, A. K. (2009). Buildup of choice-predictive activity in human motor cortex during perceptual decision making. Current Biology , 19 (18), 1581‑1585. https://doi.org/10.1016/j.cub.2009.07.066 Dzindolet, M. T., Pierce, L. G., Beck, H. P., & Dawe, L. A. (2002). The perceived utility of human and automated aids in a visual detection task. Human factors , 44 (1), 79-94. Dzindolet, M. T., Peterson, S. A., Pomranky, R. A., Pierce, L. G., & Beck, H. P. (2003). The role of trust in automation reliance. International Journal of Human-Computer Studies , 58 (6), 697‑718. https://doi.org/10.1016/S1071-5819(03)00038-7 Eisma, J., Rawls, E., Long, S., Mach, R., & Lamm, C. (2021). Frontal midline theta differentiates separate cognitive control strategies while still generalizing the need for cognitive control. Scientific Reports , 11 (1), Article 1. https://doi.org/10.1038/s41598-021-94162-z Eitam, B., Kennedy, P. M., & Tory Higgins, E. (2013). Motivation from control. Experimental Brain Research , 229 (3), 475‑484. https://doi.org/10.1007/s00221-012-3370-7 Elkin, L. A., Kay, M., Higgins, J. J., & Wobbrock, J. O. (2021, October). An aligned rank transform procedure for multifactor contrast tests. In The 34th annual ACM symposium on user interface software and technology (pp. 754-768). Endsley, M. R., & Kiris, E. O. (1995). The out-of-the-loop performance problem and level of control in automation. Human Factors , 37 (2), 381‑394. https://doi.org/10.1518/001872095779064555 Engel, A. K., & Fries, P. (2010). Beta-band oscillations—Signaling the status quo? Current Opinion in Neurobiology , 20 (2), 156‑165. https://doi.org/10.1016/j.conb.2010.02.015 Engel, A. K., & Singer, W. (2001). Temporal binding and the neural correlates of sensory awareness. Trends in cognitive sciences , 5 (1), 16-25. Fries, P., Reynolds, J. H., Rorie, A. E., & Desimone, R. (2001). Modulation of oscillatory neuronal synchronization by selective visual attention. Science , 291 (5508), 1560‑1563. https://doi.org/10.1126/science.1055465 Friese, U., Daume, J., Göschl, F., König, P., Wang, P., & Engel, A. K. (2016). Oscillatory brain activity during multisensory attention reflects activation, disinhibition, and cognitive control. Scientific Reports , 6 (1), Article 1. https://doi.org/10.1038/srep32775 Gallagher, S. (2000). Philosophical conceptions of the self: Implications for cognitive science. Trends in Cognitive Sciences , 4 (1), 14‑21. https://doi.org/10.1016/S1364-6613(99)01417-5 Ghiani, A., Maniglia, M., Battaglini, L., Melcher, D., & Ronconi, L. (2021). Binding mechanisms in visual perception and their link with neural oscillations: a review of evidence from tACS. Frontiers in Psychology , 12, 643677. https://doi.org/10.3389/fpsyg.2021.643677 Gladwin, T. E., Lindsen, J. P., & de Jong, R. (2006). Pre-stimulus EEG effects related to response speed, task switching and upcoming response hand. Biological Psychology , 72 (1), 15‑34. https://doi.org/10.1016/j.biopsycho.2005.05.005 Grandchamp, R., & Delorme, A. (2011). Single-trial normalization for event-related spectral decomposition reduces sensitivity to noisy trials. Frontiers in Psychology , 2 . https://doi.org/10.3389/fpsyg.2011.00236 Guérin-Dugué, A., Roy, R. N., Kristensen, E., Rivet, B., Vercueil, L., & Tcherkassof, A. (2018). Temporal dynamics of natural static emotional facial expressions decoding: A study using event- and eye fixation-related potentials. Frontiers in Psychology , 9 , 1190. https://doi.org/10.3389/fpsyg.2018.01190 Haegens, S., Osipova, D., Oostenveld, R., & Jensen, O. (2009). Somatosensory working memory performance in humans depends on both engagement and disengagement of regions in a distributed network. Human Brain Mapping , 31 (1), 26‑35. https://doi.org/10.1002/hbm.20842 Hanslmayr, S., Spitzer, B., & Bauml, K.-H. (2009). Brain oscillations dissociate between semantic and nonsemantic encoding of episodic memories. Cerebral Cortex , 19 (7), 1631‑1640. https://doi.org/10.1093/cercor/bhn197 Hanslmayr, S., Staudigl, T., & Fellner, M.-C. (2012). Oscillatory power decreases and long-term memory: The information via desynchronization hypothesis. Frontiers in Human Neuroscience , 6 . https://www.frontiersin.org/articles/10.3389/fnhum.2012.00074 Hoff, K. A., & Bashir, M. (2015). Trust in automation: Integrating empirical evidence on factors that influence trust. Human Factors , 57 (3), 407‑434. https://doi.org/10.1177/0018720814547570 Hoogeveen, S., Schjoedt, U., & van Elk, M. (2018). Did I do that? Expectancy effects of brain stimulation on error-related negativity and sense of agency. Journal of Cognitive Neuroscience , 30 (11), 1720‑1733. https://doi.org/10.1162/jocn_a_01297 Hu, X., Xu, Z., Li, Y., & Mai, X. (2018). The impact of trust decision-making on outcome processing: Evidence from brain potentials and neural oscillations. Neuropsychologia , 119 , 136‑144. https://doi.org/10.1016/j.neuropsychologia.2018.07.036 Huang, J., Choo, S., Pugh, Z. H., & Nam, C. S. (2021). Evaluating effective connectivity of trust in human–automation interaction: A dynamic causal modeling (DCM) study. Human Factors , 0018720820987443. https://doi.org/10.1177/0018720820987443 Jensen, O., Kaiser, J., & Lachaux, J.-P. (2007). Human gamma-frequency oscillations associated with attention and memory. Trends in Neurosciences , 30 (7), 317‑324. https://doi.org/10.1016/j.tins.2007.05.001 Jensen, O., & Mazaheri, A. (2010). Shaping functional architecture by oscillatory alpha activity: Gating by inhibition. Frontiers in Human Neuroscience , 4 , 186. https://doi.org/10.3389/fnhum.2010.00186 Jung, E.-S., Dong, S.-Y., & Lee, S.-Y. (2019). Neural Correlates of Variations in Human Trust in Human-like Machines during Non-reciprocal Interactions. Scientific Reports , 9 (1), 9975. https://doi.org/10.1038/s41598-019-46098-8 Kaiser, J., Iliopoulos, P., Steinmassl, K., & Schütz-Bosbach, S. (2022). Preparing for success: Neural frontal theta and posterior alpha dynamics during action preparation predict flexible resolution of cognitive conflicts. Journal of Cognitive Neuroscience , 34 (6), 1070‑1089. https://doi.org/10.1162/jocn_a_01846 Kaiser, J., Lutzenberger, W., Decker, C., Wibral, M., & Rahm, B. (2009). Task- and performance-related modulation of domain-specific auditory short-term memory representations in the gamma-band. NeuroImage , 46 (4), 1127‑1136. https://doi.org/10.1016/j.neuroimage.2009.03.011 Klimesch, W., Doppelmayr, M., Schimke, H., & Ripper, B. (1997). Theta synchronization and alpha desynchronization in a memory task. Psychophysiology , 34 (2), 169‑176. https://doi.org/10.1111/j.1469-8986.1997.tb02128.x Koelewijn, T., van Schie, H. T., Bekkering, H., Oostenveld, R., & Jensen, O. (2008). Motor-cortical beta oscillations are modulated by correctness of observed action. NeuroImage , 40 (2), 767‑775. https://doi.org/10.1016/j.neuroimage.2007.12.018 Kohn, S. C., De Visser, E. J., Wiese, E., Lee, Y.-C., & Shaw, T. H. (2021). Measurement of trust in automation: A narrative review and reference guide. Frontiers in Psychology , 12 , 604977. https://doi.org/10.3389/fpsyg.2021.604977 Lal, S. K. L., & Craig, A. (2002). Driver fatigue: Electroencephalography and psychological assessment. Psychophysiology , 39 (3), 313‑321. https://doi.org/10.1017/S0048577201393095 Lee, J. D. (2008). Review of a pivotal human factors article: “Humans and automation: Use, misuse, disuse, abuse”. Human Factors: The Journal of the Human Factors and Ergonomics Society , 50 (3), 404‑410. https://doi.org/10.1518/001872008X288547 Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors: The Journal of the Human Factors and Ergonomics Society , 46 (1), 50‑80. https://doi.org/10.1518/hfes.46.1.50_30392 Li, P., Han, C., Lei, Y., Holroyd, C. B., & Li, H. (2011). Responsibility modulates neural mechanisms of outcome processing: An ERP study: Modulation of outcome processing by responsibility. Psychophysiology , 48 (8), 1129‑1133. https://doi.org/10.1111/j.1469-8986.2011.01182.x Lopes, S., Chalil Madathil, K., Bertrand, J., Brady, C., Li, D., & McNeese, N. (2022). Effect of mental fatigue on trust and workload with ai-enabled infrastructure visual inspection systems. Proceedings of the Human Factors and Ergonomics Society Annual Meeting , 66 (1), 285‑286. https://doi.org/10.1177/1071181322661256 Mackworth, N. H. (1948). The breakdown of vigilance during prolonged visual search. Quarterly Journal of Experimental Psychology , 1 (1), 6‑21. https://doi.org/10.1080/17470214808416738 Maille, N., Salomone, M., Desantis, A., Le Goff, K., Sciabica, J.-F., Bressolle, M.-C., & Berberian, B. (2022). Physiological markers of vigilance variation in a supervisory task . 13th International Conference on Applied Human Factors and Ergonomics (AHFE 2022). https://doi.org/10.54941/ahfe1001819 Maris, E., & Oostenveld, R. (2007). Nonparametric statistical testing of EEG- and MEG-data. Journal of Neuroscience Methods , 164 (1), 177‑190. https://doi.org/10.1016/j.jneumeth.2007.03.024 Masci, P., & Muñoz, C. A. (2020). A graphical toolkit for the validation of requirements for detect and avoid systems. In W. Ahrendt & H. Wehrheim (Éds.), Tests and Proofs (Vol. 12165, p. 155‑166). Springer International Publishing. https://doi.org/10.1007/978-3-030-50995-8_9 McDonnell, A. S., Simmons, T. G., Erickson, G. G., Lohani, M., Cooper, J. M., & Strayer, D. L. (2021). This Is Your Brain on Autopilot: Neural Indices of Driver Workload and Engagement During Partial Vehicle Automation. Human Factors: The Journal of the Human Factors and Ergonomics Society , 001872082110390. https://doi.org/10.1177/00187208211039091 Merat, N., Seppelt, B., Louw, T., Engström, J., Lee, J. D., Johansson, E., Green, C. A., Katazaki, S., Monk, C., Itoh, M., McGehee, D., Sunda, T., Unoura, K., Victor, T., Schieben, A., & Keinath, A. (2019). The “Out-of-the-Loop” concept in automated driving: Proposed definition, measures and implications. Cognition, Technology & Work , 21 (1), 87‑98. https://doi.org/10.1007/s10111-018-0525-8 Merritt, S. M., Ako-Brew, A., Bryant, W. J., Staley, A., McKenna, M., Leone, A., & Shirase, L. (2019). Automation-induced complacency potential: Development and validation of a new scale. Frontiers in Psychology , 10 , 225. https://doi.org/10.3389/fpsyg.2019.00225 Miller, J. (2010). The U.S. Air force-developed adaptation of the multi-attribute task battery for the assessment of human operator workload and strategic behavior. Consortium research and fellows program arlington VA. https://apps.dtic.mil/sti/citations/ADA537547 Minohara, R., Wen, W., Hamasaki, S., Maeda, T., Kato, M., Yamakawa, H., … & Asama, H. (2016). Strength of intentional effort enhances the sense of agency. Frontiers in psychology, 7 , 1165. https://doi.org/10.3389/fpsyg.2016.01165 Moore, J. W., & Fletcher, P. C. (2012). Sense of agency in health and disease: A review of cue integration approaches. Consciousness and Cognition , 21 (1), 59‑68. https://doi.org/10.1016/j.concog.2011.08.010 Neubauer, C., Matthews, G., Langheim, L., & Saxby, D. (2012). Fatigue and voluntary utilization of automation in simulated driving. Human Factors: The Journal of the Human Factors and Ergonomics Society , 54 (5), 734‑746. https://doi.org/10.1177/0018720811423261 Ni, J., Wunderle, T., Lewis, C. M., Desimone, R., Diester, I., & Fries, P. (2016). Gamma-rhythmic gain modulation. Neuron , 92 (1), 240‑251. https://doi.org/10.1016/j.neuron.2016.09.003 Norman, D. A., Broadbent, D. E., Baddeley, A. D., & Reason, J. (1990). The ‘problem’ with automation: Inappropriate feedback and interaction, not ‘over-automation’. Philosophical Transactions of the Royal Society of London. B, Biological Sciences , 327 (1241), 585‑593. https://doi.org/10.1098/rstb.1990.0101 Oh, S., Seong, Y., Yi, S., & Park, S. (2020). Neurological measurement of human trust in automation using electroencephalogram. International Journal of Fuzzy Logic and Intelligent Systems , 20 (4), 261‑271. https://doi.org/10.5391/IJFIS.2020.20.4.261 Oostenveld, R., Fries, P., Maris, E., & Schoffelen, J.-M. (2010). FieldTrip: Open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data. Computational Intelligence and Neuroscience , 2011 , e156869. https://doi.org/10.1155/2011/156869 Oostenveld, R., & Praamstra, P. (2001). The five percent electrode system for high-resolution EEG and ERP measurements. Clinical Neurophysiology , 112 (4), 713‑719. https://doi.org/10.1016/S1388-2457(00)00527-7 Parasuraman, R., Molloy, R., & Singh, I. L. (1993). Performance consequences of automation-induced « complacency ». The International Journal of Aviation Psychology , 3 (1), 1‑23. https://doi.org/10.1207/s15327108ijap0301_1 Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors , 39 (2), 230‑253. https://doi.org/10.1518/001872097778543886 Parasuraman, R., Warm, J. S., & Dember, W. N. (1987). Vigilance: Taxonomy and utility. In L. S. Mark, J. S. Warm, & R. L. Huston (Éds.), Ergonomics and Human Factors (p. 11‑32). Springer New York. https://doi.org/10.1007/978-1-4612-4756-2_2 Pezzetta, R., Nicolardi, V., Tidoni, E., & Aglioti, S. M. (2018). Error, rather than its probability, elicits specific electrocortical signatures: A combined EEG-immersive virtual reality study of action observation. Journal of Neurophysiology , 120 (3), 1107‑1118. https://doi.org/10.1152/jn.00130.2018 Pfurtscheller, G., Neuper, C., Brunner, C., & da Silva, F. L. (2005). Beta rebound after different types of motor imagery in man. Neuroscience Letters , 378 (3), 156‑159. https://doi.org/10.1016/j.neulet.2004.12.034 Potts, C. A., & Carlson, R. A. (2019). Control used and control felt: Two sides of the agency coin. Attention, Perception, & Psychophysics , 81 (7), 2304‑2319. https://doi.org/10.3758/s13414-019-01771-y Rice, S. (2009). Examining single- and multiple-process theories of trust in automation. The Journal of General Psychology , 136 (3), 303‑322. https://doi.org/10.3200/GENP.136.3.303-322 Roux, F., Wibral, M., Mohr, H. M., Singer, W., & Uhlhaas, P. J. (2012). Gamma-band activity in human prefrontal cortex codes for the number of relevant items maintained in working memory. Journal of Neuroscience , 32 (36), 12411‑12420. https://doi.org/10.1523/JNEUROSCI.0421-12.2012 Sahaï, A., Desantis, A., Grynszpan, O., Pacherie, E., & Berberian, B. (2019). Action co-representation and the sense of agency during a joint Simon task: Comparing human and machine co-agents. Consciousness and Cognition , 67 , 44‑55. https://doi.org/10.1016/j.concog.2018.11.008 Sarter, N. B., Woods, D. D., & Billings, C. E. (1997). Automation surprises. In Handbook of Human Factors & Ergonomics, (Vol. 2, p. 1926‑1943). Wiley. Sauseng, P., Griesmayr, B., Freunberger, R., & Klimesch, W. (2010). Control mechanisms in working memory: A possible function of EEG theta oscillations. Neuroscience and Biobehavioral Reviews , 34 (7), 1015‑1022. https://doi.org/10.1016/j.neubiorev.2009.12.006 Schoeller, F., Miller, M., Salomon, R., & Friston, K. J. (2021). Trust as extended control: human-machine interactions as active inference. Frontiers in Systems Neuroscience , 15 , 669810. https://doi.org/10.3389/fnsys.2021.669810 Seemann, A. (2009). Joint agency: Intersubjectivity, sense of control, and the feeling of trust. Inquiry , 52 (5), 500‑515. https://doi.org/10.1080/00201740903302634 Sidarus, N., & Haggard, P. (2016). Difficult action decisions reduce the sense of agency: A study using the Eriksen flanker task. Acta psychologica, 166 , 1-11. https://doi.org/10.1016/j.actpsy.2016.03.003 Sidarus, N., Vuorre, M., & Haggard, P. (2017). How action selection influences the sense of agency: An ERP study. NeuroImage , 150 , 1‑13. https://doi.org/10.1016/j.neuroimage.2017.02.015 Somon, B., Campagne, A., Delorme, A., & Berberian, B. (2022). Brain mechanisms of automated conflict avoidance simulator supervision. Psychophysiology , e14171. https://doi.org/10.1111/psyp.14171 Synofzik, M., Vosgerau, G., & Newen, A. (2008). Beyond the comparator model: A multifactorial two-step account of agency. Consciousness and Cognition , 17 (1), 219‑239. https://doi.org/10.1016/j.concog.2007.03.010 Ueda, S., Nakashima, R., & Kumada, T. (2021). Influence of levels of automation on the sense of agency during continuous action. Scientific Reports , 11 (1), 2436. https://doi.org/10.1038/s41598-021-82036-3 van Driel, J., Ridderinkhof, K Richard, & Cohen, Michael X. (2012). Not all errors are alike. The Journal of Neuroscience , 32 (47), 16795‑16806. https://doi.org/10.1523/jneurosci.0802-12.2012 Vantrepotte, Q., Chambon, V., & Berberian, B. (2023). The reliability of assistance systems modulates the sense of control and acceptability of human operators. Scientific Reports , 13 (1), 14410. https://doi.org/10.1038/s41598-023-41253-8 Vereschak, O., Bailly, G., & Caramiaux, B. (2021). How to evaluate trust in ai-assisted decision making? A survey of empirical methodologies. Proceedings of the ACM on Human-Computer Interaction , 5 (CSCW2), 1‑39. https://doi.org/10.1145/3476068 Wahn, B., Schmitz, L., Gerster, F. N., & Weiss, M. (2023). Offloading under cognitive load: Humans are willing to offload parts of an attentionally demanding task to an algorithm. PLOS ONE , 18 (5), e0286102. https://doi.org/10.1371/journal.pone.0286102 Wascher, E., Getzmann, S., & Karthaus, M. (2016). Driver state examination—Treading new paths. Accident Analysis & Prevention , 91 , 157‑165. https://doi.org/10.1016/j.aap.2016.02.029 Wascher, E., Rasch, B., Sänger, J., Hoffmann, S., Schneider, D., Rinkenauer, G., Heuer, H., & Gutberlet, I. (2014). Frontal theta activity reflects distinct aspects of mental fatigue. Biological Psychology , 96 , 57‑65. https://doi.org/10.1016/j.biopsycho.2013.11.010 Wen, W., & Haggard, P. (2018). Control changes the way we look at the world. Journal of Cognitive Neuroscience , 30 (4), 603‑619. https://doi.org/10.1162/jocn_a_01226 Wen, W., Kuroki, Y., & Asama, H. (2019). The sense of agency in driving automation. Frontiers in Psychology , 10 . https://www.frontiersin.org/articles/10.3389/fpsyg.2019.02691 Wen, W., Yamashita, A., & Asama, H. (2015). The sense of agency during continuous action: Performance is more important than action-feedback association. PLOS ONE , 10 (4), e0125226. https://doi.org/10.1371/journal.pone.0125226 Wohleber, R. W., Matthews, G., Lin, J., Szalma, J. L., Calhoun, G. L., Funke, G. J., Chiu, C.-Y. P., & Ruff, H. A. (2019). Vigilance and automation dependence in operation of multiple unmanned aerial systems (UAS): A simulation study. Human Factors: The Journal of the Human Factors and Ergonomics Society , 61 (3), 488‑505. https://doi.org/10.1177/0018720818799468 Wobbrock, J. O., Findlater, L., Gergle, D., & Higgins, J. J. (2011, May). The aligned rank transform for nonparametric factorial analyses using only anova procedures. In Proceedings of the SIGCHI conference on human factors in computing systems (pp. 143-146). Crossref Google Scholar Information & Authors Information Version history V1 Version 1 24 December 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Authors Affiliations Salomone. M. 0009-0005-5418-9632 [email protected] Universite Grenoble Alpes View all articles by this author Bruno Berberian 0000-0002-3908-4358 ONERA Salon-de-Provence View all articles by this author Aurélie Campagne Universite Grenoble Alpes View all articles by this author Metrics & Citations Metrics Article Usage 159 views 74 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Salomone. M., Bruno Berberian, Aurélie Campagne. Neural correlates of the out-of-the-loop phenomenon during the supervision of an automated system in aeronautic context. Authorea . 24 December 2025. DOI: https://doi.org/10.22541/au.176656330.08996904/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.176656330.08996904/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a0066419d9c452ad',t:'MTc3OTU2Mzc0NA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();