The Timing Dependencies of Trust: Speed, Accuracy, and cBCI Neuro-Decoupling in Human-AI Teams Christopher Baker [email protected] School of Electronics, Electrical Engineering and Computer Science Queen's University Belfast Stephen Hinton [email protected] School of Psychology Liverpool John Moores University Akashdeep Nijjar [email protected] School of Computer Science and Electronic Engineering University of Essex Riccardo Poli [email protected] School of Computer Science and Electronic Engineering University of Essex Caterina Cinel [email protected] School of Computer Science and Electronic Engineering University of Essex Tom Reed [email protected] Defence Science Technology Laboratory Stephen Fairclough [email protected] School of Psychology Liverpool John Moores University Collaborative Brain-Computer Interface (cBCI), Human-AI Teaming, Team Decision-Making, Neuroergonomics, AI Deception, Error Resilience, Cognitive Workload, Virtual Reality (VR), Machine Learning, Event-Related Potentials (ERPs)
Abstract The speed and accuracy of an artificial teammate fundamentally alter the failure states of Human-AI integration. While high-speed AI interventions risk inducing reflexive "blind compliance", delayed
interventions can induce ambiguous "cognitive conflict". This study investigates how the fundamental characteristics of an in-task AI assistant (Fast/Less-Accurate (FLA-AI) versus Slow/Accurate (SA-AI)—impact the synergy of Collaborative Brain-Computer Interface (cBCI) teams in a Virtual Reality drone task. Seventeen operators completed continuous search tasks under high cognitive workload while their spatial covariance was mapped using a 2D Adaptive Riemannian Oracle. The results mathematically demonstrate that AI timing dictates the mechanism of team failure. Fast AI induced instant, blind compliance; human accuracy under deception collapsed to 50.2%, and pure behavioural teams (N=8) failed to scale beyond 74.1%. In contrast, Slow AI induced delayed cognitive conflict; humans hesitated (61.1% accuracy), but N=8 behavioural teams eventually recovered to 100.0%. Crucially, the Riemannian Oracle mathematically adapted to these states: it heavily restricted temporal windows (<0.8s) to intercept fast reflexive compliance, while widening windows (>1.2s) to capture delayed cognitive conflict. Integrating these isolated veridical signals via Hybrid Fusion successfully rescued the Fast AI team (+7.6% at N=8) and significantly accelerated the recovery of smaller Slow AI teams (+6.9% at N=4). These findings prove that cBCI synergy is heavily contingent on the temporal dynamics of trust, providing a critical framework for designing dynamically gated Human-AI systems.
Introduction Modern human-AI teaming holds immense promise for enhancing operational effectiveness across a range of important decision domains, from everyday recommendations to critical workplace predictions in fields like medicine, law, or financial services [1,2]. This has led to a drive to improve collaboration by developing better mental models of AI behaviour [3] and calibrating human trust in the AI system [4–6]. However, this promise is tempered by the fact that humans often struggle to rely on AI appropriately, leading to sub-optimal team performance [7]. This introduces novel vulnerabilities, particularly when a systematically flawed or deceptive AI provides guidance that induces correlated errors across multiple human operators [1,4]. Under conditions of high cognitive demand [8,9], this risk is amplified. The statistical advantage of a group [10,11], often termed the "wisdom of crowds", is predicated on the independence of its members' judgments. A flawed AI, acting as a common source of biased information, can systematically violate this independence [12]. This can invert the statistical advantage of a team, transforming the group into a mechanism for catastrophic, widespread failure [13,14]. Conventional team aggregation methods are designed to harness this collective intelligence by condensing individual inputs into a single team output [15,16]. Simple strategies like majority voting treat all members equally, while more sophisticated approaches apply weightings to individual inputs based on metrics like past performance or self-reported confidence [17,18]. Both approaches, however, are particularly susceptible to the failure mode introduced by a deceptive AI because they rely on the implicit assumption that the behavioural outputs of team members, their decisions and their confidence, are reliable indicators of ground truth. This assumption is fundamentally challenged by contemporary models of metacognition (see [19] for review), which frame subjective confidence not as a direct readout of decision evidence, but as a separate, "second-order" computation that makes an inference about the likely quality of a decision [20,21]. Empirical work supports this theoretical separation, demonstrating that the sensory evidence contributing to a first-order decision can be dissociated from the information that supports a subsequent metacognitive judgment [21]. This confidence-accuracy relationship is especially fragile under conditions of high cognitive demand or information overload, where flawed metacognitive assessments can lead to paradoxically sub-optimal performance [22]. A misleading AI can therefore systematically corrupt this distinct metacognitive channel, creating a state where operators become confidently incorrect. The existence of distinct neural processes for performance monitoring, which can even give rise to "early error sensations" before a motor response is executed [23], further highlights that
the decision and its evaluation are separable processes and thus independently vulnerable to bias. This invalidation of behavioural confidence as a reliable signal creates a critical gap in current safeguards, as the aggregation of corrupted behavioural reports can transform the team into a mechanism for widespread failure. As a potential safeguard, the collaborative Brain-Computer Interface (cBCI) offers access to an insulated channel of evidence. A cBCI can be designed to bypass the potentially corrupted conscious cognitive process by accessing implicit neural activity generated before a final behavioural judgment is formed. This pre-decisional neural signal, it is argued, can remain insulated from the cognitive biases that affect the overt response, providing a more robust source of information under deception [24,25]. The neuroscientific plausibility for this approach is grounded in the brain's generation of reliable neural markers that differentiate between effortful, deliberative processing and efficient, automatic states [26]. When individuals are confronted with conflicting information that requires active problem-solving, a specific neural signature reliably emerges from frontal midline brain regions: an increase in theta-band (~4–8 Hz) power. This signal is considered a general marker for the need to engage cognitive control across many different types of cognitive challenges [27], reflecting a genuine neural oscillation rather than a simple evoked response [28], and is believed to originate from a dedicated neural microcircuit for conflict detection [29]. This same conflict-monitoring system also gives rise to the brain's automatic and rapid responses to errors and negative feedback, known respectively as the Error-Related Negativity (ERN) and Feedback-Related Negativity (FRN) [30]. Crucially, the conscious feeling of having made an error can even arise before a physical response is executed, supporting the existence of a pre-decisional signal of internal conflict [23]. Following this initial conflict detection, a more sustained and effortful re-evaluation of evidence is indexed by a late positive potential (LPP), an event-related potential associated with successful cognitive reappraisal and conscious deliberation [31]. Conversely, when tasks can be performed automatically and without conflict, the brain enters a more efficient 'autopilot' state. This state is characterized by its own distinct set of 'honest signals'. One key signature is the modulation of posterior alpha rhythms (8-13 Hz), where increased power is a well-established marker of reduced visual engagement and the active inhibition of irrelevant sensory information, a process known as attentional gating [32,33]. A second, key signature is the modulation of sensorimotor beta power (13-30 Hz). Beta oscillations are characteristically suppressed during active movement but are prominent during steady-state motor control, reflecting the maintenance of the current motor 'status quo' [34]. More than just a signal of motor idling, this beta activity has been shown to index the brain's confidence in its internal models; higher beta power reflects a higher confidence in the current motor plan and a reduced need for adaptation [35]. From a computational perspective, this corresponds to a lower weighting of sensory prediction errors, effectively promoting stability [36] and response certainty [36,37]. However, extracting these pre-decisional signals in continuous, ecologically valid 3D Virtual Reality environments presents a significant neurophysiological challenge. Because operators visually detect asynchronous targets before a formal system prompt appears, traditional time-domain Event-Related Potentials (ERPs) and static frequency bands are often washed out by latency jitter and temporal smearing. To overcome this, the current study deploys a state-of-the-art topological spatial covariance mapping pipeline utilizing Riemannian geometry. By mapping the full multi-dimensional spatial state of the visual and executive cortices into a Euclidean Tangent Space, the system can mathematically decouple the veridical, preserved sensory representation of the environment from the temporally noisy and cognitively corrupted behavioural response. It is important to clarify that this system operates as a Passive BCI (pBCI). Unlike Active BCIs, which require voluntary control of neural signals for communication, Passive BCIs monitor implicit cognitive states to augment performance without diverting the user's attention. Our approach builds specifically on the foundational collaborative BCI (cBCI) literature established by Poli, Cinel, and Valeriani [15,16], extending their methods into high-fidelity Virtual Reality.
The objective of the current study is therefore twofold: first, to mathematically evaluate how the timing of an artificial assistant (Fast/Less-Accurate vs. Slow/Accurate) alters the failure states of human operators under deception; and second, to determine if a purely Neuro-Decoupled Team (NDT) can adapt to these distinct temporal windows. To achieve this, we compare the performance of Hybrid cBCI teams against traditional behaviour-based teams within a high-workload drone surveillance task. We hypothesised that while fast AI would induce reflexive "blind compliance" and slow AI would induce delayed "cognitive conflict," the Riemannian Oracle would successfully isolate the veridical pre-decisional neural states in both conditions, mathematically rescuing the team from systematic failure.
Results To evaluate the impact of AI characteristics on neuro-behavioural decoupling, team performance was simulated under two distinct conditions: a Fast/Less-Accurate AI (FLA-AI) providing instant feedback, and a Slow/Accurate AI (SA-AI) providing delayed feedback. Baseline Operations: AI Correct During baseline operations, when both the FLA-AI and SA-AI provided veridical guidance, individual operators maintained high accuracy (averaging 87.1% and 90.3%, respectively). Consequently, pure behavioural teams scaled efficiently, reaching near-perfect accuracy (99.0% and 100.0% at N=8) and demonstrating baseline human-AI competence before adversarial conditions were introduced.
Fast AI (FLA-AI): Blind Compliance and Reflexive Hijack When the instant FLA-AI provided deceptive guidance under high workload, it induced severe automation bias. The immediacy of the false cue preempted human deliberation, causing average individual accuracy to collapse to 50.2%. Because this instant feedback induced blind compliance, the resulting errors were highly correlated. Consequently, pure behavioural human teams (RT Weighted) failed to adequately scale, plateauing at 74.15% for an eight-person group (N=8). To rescue the team, the 2D Adaptive Riemannian Oracle evaluated the spatial covariance of the operators. Demonstrating the temporal dependency of the lie, the Oracle overwhelmingly selected highly restrictive Reaction Time upper-bounds (e.g., 0.8s) to amputate corrupted data. By isolating the veridical sensory reflex before the blind compliance took hold, the BCI maintained a 70.1% accuracy. When integrated into the collective, the Hybrid Fusion team achieved 81.77% (N=8), generating a statistically significant rescue of +7.62% (p < 0.0001) over the pure behavioural team.
Figure 1. Team decision accuracy under Slow AI Deception. While behavioural teams eventually resolve the cognitive conflict at N=8, Hybrid Fusion significantly accelerates recovery at N=2 and N=4.
Slow AI (SA-AI): Cognitive Conflict and Synergistic Acceleration The failure mode fundamentally shifted when the AI feedback was delayed (SA-AI). The delay allowed human operators to form an initial perceptual baseline before the AI intervened. When the delayed AI provided deceptive guidance, it caused measurable cognitive conflict rather than instant compliance. Average individual accuracy dropped to 61.1%, but operators were not completely hijacked. Because the errors were less correlated, pure behavioural teams successfully recovered over time, scaling to 100.0% at N=8. The Riemannian Oracle adapted its mathematical boundaries to this new psychological state. Instead of restrictive 0.8s bounds, the Oracle dynamically widened its temporal parameters (e.g., 1.2s, 1.5s, Unlimited) to capture the delayed cognitive conflict, achieving an individual BCI accuracy of 75.8%. While behavioural scaling eventually solved the SA-AI deception at N=8, the Hybrid Fusion team provided a massive acceleration in decision superiority for smaller, operationally realistic team sizes. For a four-person team (N=4), the pure behavioural team achieved 85.85%, while the Hybrid Fusion team surged to 92.79% (a highly significant rescue delta of +6.94%, p < 0.0001).
Figure 2. Team decision accuracy under Fast AI Deception. Instant deceptive feedback induces systemic blind compliance (capping behavioural teams at 74.1%), requiring Hybrid Fusion to bypass the hijack and rescue the team.
Discussion The primary finding of this study is that the fundamental characteristics of an artificial assistant, specifically its speed of intervention, dictate the psychological mechanism of team failure and the requisite neurophysiological response. Under Fast AI (FLA-AI), the immediacy of the deceptive cue triggered reflexive "blind compliance". Operators effectively bypassed their own visual processing, executing motor responses perfectly aligned with the AI's lie. To achieve a neuro-decoupled rescue, the Riemannian Oracle mathematically tightened its temporal parameters, capturing the brief sensory reflex before executive compliance solidified. Conversely, the Slow AI (SA-AI) allowed humans to establish a veridical visual baseline. The subsequent deceptive cue induced prolonged "cognitive conflict" and hesitation. The Oracle adaptively widened its temporal bounds to map this sustained spatial disconnect, proving that cBCI parameters must be as dynamic as the psychological states they monitor. These findings mathematically demonstrate that AI timing dictates the mechanism of team failure, and that Hybrid Fusion can successfully rescue teams from instant compliance while significantly accelerating recovery under delayed conflict.
Limitations and Future Work Several limitations must be acknowledged in the context of the results presented. The primary limitation is that team performance was evaluated via offline simulation. However, we argue that this approach possesses ecological validity for the specific domain of remote operations. In modern high-pressure environments, teams are often distributed, ad-hoc, or asynchronous (e.g., swarm operators), contributing to a collective output without physical proximity or direct social interaction. Our simulation accurately models this increasingly critical form of distributed human-machine teaming. The sample size (N=17) may limit the generalisability of individual difference analyses. However, this sample size is consistent with neuroergonomic protocols and is sufficient for the within-subject comparisons employed here. Furthermore, the team-level conclusions are drawn from over 11.7 million simulated decision points, providing a high degree of statistical confidence. The next crucial step is to implement this strategy in a real-time, closed-loop collaborative BCI to study live team interactions and validate the synergistic gains. Secondly, our deception paradigm was standardised. A valuable follow-up study could directly test the BCI's resilience against different AI feedback mechanisms within the high-workload state, for example, by contrasting an AI with consistent errors against one with intermittent or varied performance. Furthermore, it must be acknowledged that in a real-world drone scenario, operators would likely have full control of the vehicle. This active piloting introduces additional cognitive demands and agency which may mitigate or increase the effect of AI errors in taxing scenarios, an interaction that warrants future investigation. Finally, while the Signal Detection metrics mathematically prove the preservation of veridical sensory states under deception, Riemannian Tangent Space mappings are inherently high-dimensional. Future work should seek to project these tangent space coefficients back into sensor space to generate topographical maps, directly visualising the specific visual-motor spatial disconnect driving this classification.
Conclusion This study demonstrates that the timing of AI feedback fundamentally alters human failure states under cognitive load, ranging from rapid blind compliance to delayed cognitive conflict. By deploying a 2D Adaptive Riemannian Oracle, we proved that implicit BCI strategies can dynamically adapt to these shifting temporal windows, capturing the veridical sensory state before corrupted executive responses solidify. The results mathematically confirm that while behavioural aggregation fails under instant deception, integrating spatial covariance via Hybrid Fusion provides a robust, synergistic safeguard. This work provides a critical neuroergonomic framework for designing dynamically gated Human-AI systems, where the implicit wisdom of the brain can be harnessed to protect teams against adversarial or malfunctioning artificial assistants.
Methods Participants Seventeen participants were included in data analyses ([10 Female]; Mean age ± SD = [24.2 ± 5.04). An initial cohort of 20 individuals was recruited; however, data from three participants were excluded prior to the main analysis due to criteria established in preliminary quality control checks, including insufficient data quality after EEG preprocessing, or significant deviations in trial sequence alignment during task execution due to technical issues and the need for alignment of trials across all combinations of team group size. All included participants reported normal or corrected-to-normal vision and no history of
neurological disorders or particular susceptibility to VR-induced motion sickness. Prior to the experiment, participants provided written informed consent and completed the Barratt Impulsiveness Scale (BIS-11) [40] and the Balloon Analogue Risk Task (BART) prior to commencing the study. Participants received monetary compensation for their participation. The experimental protocol received favourable opinion by UK MoDREC, App No: 2309/MODREC/24 Ref: RQ0000037929 and all procedures were conducted in accordance with the ethical standards outlined in the Declaration of Helsinki.
Study Design The study employed a within-subject repeated measures design to evaluate the effectiveness of a collaborative Brain-Computer Interface (cBCI) system designed to enhance group decision-making in a dynamic virtual reality (VR) environment. To induce a necessary baseline of perceptual ambiguity, the virtual environment was strictly fixed to a Degraded Visual Environment (High Workload) across all trials. The primary within-subject factor manipulated was the timing and accuracy characteristic of the in-task AI assistant, presented at two levels across distinct experimental blocks: Fast/Less-Accurate AI (FLA-AI) versus Slow/Accurate AI (SA-AI). To isolate the veridical anticipatory cognitive state and prevent the neurophysiological signal from being corrupted by motor-execution artefacts, Electroencephalographic (EEG) data were analysed time-locked exclusively to the imperative decision cue ('ReticleOn'). Team performance was assessed through offline simulations, evaluating decision accuracy for simulated groups of varying sizes (2, 4, 6, and 8 members) under different decision aggregation algorithms.
Fig. 3 Timeline of single trial within task
VR Drone Task and Procedure Participants were seated in the laboratory while they viewed the virtual environment via a Varjo Aero HMD. The simulation, developed in Unreal Engine 5 and rendered on a high-performance PC, depicted the viewpoint of a quadcopter type drone flying over a simulated landscape designed to be semi realistic i.e.textured realistically for a temperate climate and populated with sparse foliage but tailored to not interfere with stimuli acquisition (See Fig. 11). Task stimuli consisted of 3D models appearing on the ground: Non-Targets (See Fig. 12) and Targets (See Fig. 13). Participants completed six 8-minute blocks, with each block containing 50 trials (30 Non-Targets, 20 Targets). Stimuli were evenly spaced 300m apart, but varied in lateral placement (-30m, -15m, 0m, +15m, +30m)] and rotation (0°, 90°, 180°, 270°).
Fig. 4. Example of a Non-target stimulus
Fig. 5. Example of a Target stimulus
Each trial followed a consistent sequence. Initially, the drone flew to a set virtual height (25m) and advanced across the landscape. As stimuli came into view a targeting reticle appeared over the stimuli ('ReticleOn'). Crucially, the colour of this reticle immediately indicated an in-task AI's assessment of the stimulus: the reticle appeared blue if the AI classified the stimulus as a Non-Target (see Fig. 14), and red if the AI classified it as a Target (see Fig. 15). Participants were informed that this AI-indicated reticle colour was intended to assist their decision-making but that the AI was not perfect and they should make their own final judgment. The AI-indicated reticle remained on-screen for a total of 2500ms. Participants were tasked with discriminating the stimulus type and responding as quickly and accurately as possible by pressing the designated joystick button for Target or Non-Target. Participants were instructed to respond whilst the reticle was on-screen. To provide a response deadline warning, 2000ms after the 'ReticleOn' event, the reticle would change to white for 500ms and then disappear. (See Fig. 16). Any responses made after the reticle disappeared were not counted, and those trials were classed as misses. Following this primary response ('ButtonPress'), a prompt appeared, requiring participants to rate their confidence in the preceding decision via tilting the joystick to select a value on a 0-100 scale (0 = Not Confident, 100 = Very Confident) (See Fig. 17).
Fig. 6. Reticle appears blue indicating non-target feedback from AI (lighting increased for illustration).
Fig. 7. Reticle appears red indicating target feedback from AI (lighting increased for illustration).
Fig. 9. Reticle response warning (lighting increased for illustration).
Fig. 10. User interface element for subjective confidence rating (lighting increased for illustration).
To ensure the visual discrimination process was sufficiently taxing to elicit reliance on the AI assistant, the environmental cognitive workload was fixed across all experimental blocks. The virtual light level was reduced by 50% and the solar angle decreased by 9 degrees, simulating a highly degraded night condition. Within this fixed degraded environment, the characteristics of the in-task AI assistant were manipulated across different blocks of trials. Under the Fast/Less-Accurate (FLA-AI) condition, the AI provided instantaneous feedback upon stimulus detection (0ms delay) but operated with a programmed reliability of 78%. Under the Slow/Accurate (SA-AI) condition, the AI provided highly reliable feedback (90% accuracy) but with a programmed processing latency (e.g., 900 to 1600ms delay) before the reticle changed colour. This manipulation was designed to test the temporal boundaries of human-AI trust, contrasting instant reflexive compliance against delayed cognitive conflict. Participants completed three blocks under each AI condition.
Fig. 11. View from the simulated drone in as participants view it during task lighting reduced
Data Acquisition Continuous EEG data were recorded using a 32-channel LiveAmp system (Brain Products GmbH). Electrodes were arranged according to the international 10-20 configuration using an electrode cap (actiCAP snap electrode cap). Average referencing was utilised, the actiCAP utilised a dedicated ground electrode point at the front and centre of the head, between Fp1 and Fp2. Electrode impedances were kept below 30KΩ throughout the recording. The EEG signal was recorded at a sampling rate of 500Hz. behavioural data, including response button presses (for RT calculation) and subsequent confidence ratings, were logged via joystick actions which generated tLabStreamingLayer [41] (LSL) markers. Event markers corresponding to critical task events (e.g., ReticleOn, StimulusOn, ButtonPress) were generated by the Unreal Engine environment using the LSL UE5 plugin and transmitted via LSL. Both EEG and event marker streams were simultaneously recorded and synchronised using LabRecorder and OpenSignals software, resulting in .xdf files for each session.
Procedure Upon arrival at the laboratory, participants were briefed on the study's objectives and procedures. After providing written informed consent, they completed the BART and BIS-11 questionnaires. Participants were then fitted with the EEG cap and the VR HMD, see Fig. 19. Electrode impedances were checked and adjusted to be below. After baselining, participants undertook a training task of two blocks (one for each workload condition) then the VR drone task. They received instructions on the target/non-target discrimination, the response mapping (button presses), and the confidence rating scale. The main experiment consisted of six blocks, divided equally between the Low Workload and High Workload conditions (three blocks per condition). To ensure that trial-level data were directly comparable across all participants for the exhaustive combinatorial team simulations, a fixed, alternating block order was used for all participants: [Low, High, Low, High, Low, High]. This design deliberately standardises any potential time-on-task influences, such as practice or fatigue effects, ensuring they are applied identically to all participants. By treating block order as a constant, it is removed as a confounding variable from the core analysis, which compares the relative performance of different team aggregation methods. This approach was therefore a necessary
methodological choice to guarantee the trial-by-trial data integrity required for the simulation. The entire experimental session, including setup, task execution, and debriefing, lasted approximately 150 minutes per participant. Upon completion, participants were debriefed and received their monetary compensation.
Fig. 12. Participant partaking in the study
EEG Signal Processing EEG data were processed offline using MNE-Python [42] and custom Python scripts. The processing pipeline included: 1. Loading and Filtering: Data from each XDF file were loaded. A band-pass filter (0.1–30 Hz, FIR) and a 50 Hz notch filter were applied to the data. 2. Trimming: The continuous recording was trimmed to the duration of the experimental task using the first and last LSL markers. 3. Artifact Rejection (ICA): Because fully automated Independent Component Analysis (ICA) algorithms can aggressively drop components and mathematically flatten the spatial rank of the covariance matrices required for Riemannian geometry, a topological preservation strategy was enforced. ICA was computed on concatenated session data, and components reflecting ocular artefacts and slow-wave mechanical VR headset strain were manually identified by visual inspection and projected out, strictly preserving all cortical components. 4. Epoching: The cleaned continuous data were segmented into epochs relative the primary imperative cue ('ReticleOn' epochs, -200 ms to +800 ms). To dynamically heal sensor dropouts caused by participant movement in VR without dropping trials, the Autoreject algorithm was deployed, utilising dynamic Spherical Spline Interpolation to reconstruct transiently noisy channels on a per-trial basis. 5. Baseline Correction: Epochs were baseline-corrected using the pre-event interval: -200 ms to 0 ms.
BCI Feature Extraction and Classification To overcome the limitations of standard time-domain classifiers caused by latency jitter in active VR tasks, a state-of-the-art Riemannian geometry pipeline was engineered. For each participant, classification was executed using the following architecture: Topological Covariance Mapping: To mitigate contamination from facial musculature and ocular artefacts inherent to active VR, a topographical mask was applied, isolating 16 central, parietal, and occipital channels encompassing the relevant visual and motor cortices. For each trial, the spatial covariance matrix of these 16 channels was calculated using the Oracle Approximating Shrinkage (OAS) estimator to ensure robust conditioning. These matrices reside on a Symmetric Positive Definite manifold. The Riemannian Fréchet Mean was calculated to establish a stable geometric centre of the participant's baseline neural state. The matrices were then projected into a Euclidean Tangent Space to generate a high-dimensional spatial feature vector for linear classification. The Zero-Weight Quarantine: To prevent the classifier from learning corrupted conscious responses ('motor poisoning'), 100% of the epochs were passed into the Tangent Space mapping to calculate the stable spatial geometry. However, during the Logistic Regression boundary fitting, a dynamic sample weight array was applied, assigning a strict 0.0 mathematical weight to incorrect or missed trials. Thus, the spatial geometry remained stable, but the decision boundary learned exclusively from veridical ground-truth. The 2D Adaptive Oracle (Phase x Reaction Time): To combat non-stationary alpha power fatigue drift, the pipeline dynamically split the experimental session into chronological phases (Early, Mid, Late). Within each phase, the Oracle swept through strict reaction time upper bounds (e.g., 0.8s, 1.0s, 1.2s, 1.5s) to amputate delayed 'blank stare' trials that pushed the cognitive cascade outside the 800ms epoch window. The Oracle adaptively selected the exact chronological phase and physiological speed where the operator's spatial covariance geometry was optimally separated. Prediction and Confidence Score: L2 regularisation (C=0.1) was enforced within the Logistic Regression solver to prevent overfitting the Tangent Space vector. The classifier's decision function output (the signed distance from the separating hyperplane) was extracted as the continuous BCI Confidence score. This absolute score was normalised to a 0 to 1 scale via min-max scaling to serve as the standardised weight for subsequent team simulations.
Team Simulation Procedure and Aggregation Methods Performance of simulated teams was evaluated offline. To ensure a comprehensive and robust analysis, an exhaustive combinatorial approach was employed. For each of the approximately 150 valid experimental trials within each workload condition (Low and High), team performance was simulated independently for every possible unique combination of participants drawn from the final N=17 cohort for each specified team size (m = 2, 4, 6, and 8). This meant that for any single trial, decisions were simulated for: ● all 136 unique two-person teams (the number of distinct combinations of 2 participants from 17), ● all 2,380 unique four-person teams (the number of distinct combinations of 4 participants from 17),
● all 12,376 unique six-person teams (the number of distinct combinations of 6 participants from 17), ● all 24,310 unique eight-person teams (the number of distinct combinations of 8 participants from 17). Given approximately 150 trials per workload condition, and two workload conditions (Low and High, totaling ~300 trials available for analysis after individual trial validation), the total number of simulated team decisions generated was substantial: ● For two-person teams, this involved 136 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 40,800 simulated decisions. ● For four-person teams, this involved 2,380 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 714,000 simulated decisions. ● For six-person teams, this involved 12,376 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 3,712,800 simulated decisions. ● For eight-person teams, this involved 24,310 unique team combinations, each assessed over approximately 300 trials, resulting in approximately 7,293,000 simulated decisions. In total, this comprehensive simulation strategy yielded over 11.7 million individual team decision points for analysis across all team sizes and trials. For these per-trial, per-team-composition simulations, data were grouped by a unique trial identifier (Trial_Number) ensuring that for each simulated team decision, only data (e.g., individual response, RT, ML confidence) from participants who had experienced that identical trial were included when constituting that specific team instance. Team decisions were aggregated using several methods (summarized in Table 1 and detailed below), which include various human-only, BCI-only, and mixed human-BCI approaches. For the team simulations, the BCI confidence scores used were derived strictly from the 'held-out' test predictions generated during the cross-validation procedure, ensuring the simulation utilised data unseen during the classifier training phase: 1. BCI Confidence Weighted: Riemannian-predicted labels (Target/Non-Target) from each team member were weighted by their corresponding normalised BCI confidence (a 0-1 scale derived from the absolute decision function output). The team decision favoured the predicted label type with the higher sum of confidence weights. Ties were resolved by favouring Target classification. 2. RT + BCI Confidence Mixed: This method integrated human response time and BCI confidence. For each individual, the evidence supporting a Target or Non-Target decision was calculated by giving equal 0.5 weighting to their RT-weighted human response and their BCI confidence-weighted BCI prediction. For example, the evidence for a particular choice was determined as half the human-derived score for that choice plus half the BCI-derived score for that choice. The team decision favoured the response type with the higher total summed evidence across members. Ties were resolved by favouring Target classification. 3. Subjective Confidence + BCI Confidence Mixed: This method integrated human subjective confidence and BCI confidence. For each individual, the evidence supporting a Target or Non-Target decision was calculated by giving equal 0.5 weighting to their subjective confidence-weighted human response and their BCI confidence-weighted BCI prediction. For example, the evidence for a particular choice was determined as half the human-derived score for that choice plus half the BCI-derived score for that choice. The team decision favoured the response type with the higher total summed evidence across members. Ties were resolved by favouring Target classification.
4. RT + Subjective Confidence + BCI Confidence Mixed: This comprehensive method integrated human RT, human subjective confidence, and BCI confidence. For each individual, a "human component" was derived from the average of their normalised RT and normalised subjective confidence. This human component and the individual's "BCI component" (normalised BCI confidence) were then each given a 0.5 weighting. These two weighted components (one reflecting the human response and one reflecting the BCI prediction) were summed to determine the individual's contribution to the team's evidence for a Target or Non-Target decision. The team decision favoured the response type with the higher total summed evidence. Ties were resolved by favouring Target classification. Individual Performance Baselines: Team performance for these aggregation methods was also compared against baseline accuracies derived from the average accuracy of the best, worst, and average human performer within each specific simulated group combination on each trial. The main outcome measure reported is the mean team accuracy, averaged across all unique team combinations and valid trials for each group size and decision method. Unless otherwise specified for a particular method (e.g., Majority Human), ties in accumulated evidence for weighted methods were resolved by favouring Target classification. Table 1: Calculation Summary of Key Team Decision Methods Method Label in Plot
Core Logic Description
Key Information Used per Team Member (Trial-Level)
BCI Confidence Weighted
BCI-predicted labels are weighted by normalised BCI confidence. Team decision by sum of weights. Ties favour Target.
BCI Predicted Label, Normalised BCI Confidence
RT + BCI Confidence Mixed
For each member, evidence for a decision (Target/Non-Target) is 0.5 * (Human Score from RT) + 0.5 * (BCI Score from BCI Confidence). Team decision by sum of total evidence. Ties favour Target.
Human Response, Normalised RT, BCI Predicted Label, Normalised BCI Confidence
Subjective Confidence + BCI Confidence Mixed
For each member, evidence for a decision (Target/Non-Target) is 0.5 * (Human Score from Subjective Confidence) + 0.5 * (BCI Score from BCI Confidence). Team decision by sum of total evidence. Ties favour Target.
Human Response, Normalised Subjective Confidence, BCI Predicted Label, Normalised BCI Confidence
RT + Subjective Confidence + BCI Confidence Mixed
For each member, a "human component" score is the average of Normalised RT and Normalised Subjective Confidence. Evidence for a decision (Target/Non-Target) is 0.5 * (Human Component Score) + 0.5 * (BCI Score from BCI Confidence). Team decision by sum of total evidence. Ties favour Target.
Human Response, Normalised RT, Normalised Subjective Confidence, BCI Predicted Label, Normalised BCI Confidence
Statistical Analysis Individual behavioural data (accuracy, RT, confidence) were analysed to assess the impact of the Workload manipulation (Low vs. High). Depending on data distributions, paired t-tests or Wilcoxon signed-rank tests were used for continuous variables (RT, confidence), while Chi-square tests were used for accuracy (comparing counts of correct/incorrect decisions). For simulated team performance, mean accuracies for the proposed cBCI weighting method(s) were compared against baseline methods (Majority, Best/Average Individual) for each group size using paired t-tests or Wilcoxon tests. Corrections for multiple comparisons (e.g., Bonferroni) were applied where appropriate. Statistical significance was defined at an alpha level of p < 0.05. All statistical analyses were performed using Python [43] and its scientific computing libraries, primarily SciPy, for significance testing. Data processing and manipulation were conducted using Pandas [44] and NumPy [45]. Visualisations were generated with Matplotlib[46] and Seaborn [47]. For simulated team performance, mean accuracies for the proposed cBCI weighting method(s) were compared against baseline methods using paired t-tests or Wilcoxon tests. While mixed-effects models are often utilised for repeated measures, the exhaustive combinatorial nature of this simulation (testing every possible team composition on the exact same set of trials) meant that the primary source of variance being tested was the algorithmic difference between Aggregation Method A and Aggregation Method B. Therefore, paired tests were deemed the most appropriate and transparent statistical tool to directly compare the performance of different aggregation algorithms on identical team sets.
Hypotheses Based on the premise that pre-decisional neural activity provides an insulated channel of evidence against external deception, we formulated two primary hypotheses regarding the temporal dynamics of trust: 1. The Temporal Dynamics of Failure: We hypothesised that the speed of the AI assistant would fundamentally alter the team's failure state under deception. Specifically, we predicted that Fast AI (FLA-AI) would induce immediate, blind compliance (characterised by highly correlated errors and behavioural team collapse), whereas Slow AI (SA-AI) would induce delayed cognitive conflict, allowing behavioural teams to eventually recover via error cancellation. 2. Adaptive Topological Preservation: We hypothesised that mapping the holistic spatial covariance of the visual and motor cortices via a 2D Adaptive Riemannian manifold would be necessary to decouple the veridical sensory state from the hijacked response. Crucially, we predicted the Oracle would mathematically adapt to the AI's speed, selecting highly restrictive temporal windows to intercept fast compliance, and wider windows to capture delayed conflict, successfully generating a synergistic rescue across both conditions.
Declarations Ethics approval and consent to participate The experimental protocol received favourable opinion by the UK Ministry of Defence Research Ethics Committee (MoDREC), Application Number: 2309/MODREC/24, Reference: RQ0000037929. All
procedures were conducted in accordance with the ethical standards outlined in the Declaration of Helsinki. Written informed consent was obtained from all individual participants included in the study.
Data Availability Statement The datasets generated and/or analysed during the current study are not publicly available due to restrictions imposed by the funding body (Defence Science and Technology Laboratory - Dstl). However, data are available from the corresponding author (CB) on reasonable request and subject to a data sharing agreement, if appropriate and in accordance with Dstl policy.
Competing interests The authors declare that they have no competing interests.
Funding This research was funded by the Defence Science and Technology Laboratory (DSTL) via RQ0000037929.The funders contributed to the conceptualisation of the broader project aims. The funders did not have a direct role in the specific design of this study, data collection, detailed analysis, interpretation of data from this specific study, or in the writing of this manuscript beyond the contributions of the DSTL-affiliated co-author (T.R.) as described in the Author Contributions section.
Authors' contributions C.B.: Conceptualisation, Methodology, Software, Validation, Formal Analysis, Investigation, Data Curation, Writing – Original Draft, Writing – Review & Editing. S.H.: Conceptualisation, Methodology, Formal Analysis, Investigation, Writing – Review & Editing. S.F.: Conceptualisation, Methodology, Software, Formal Analysis, Investigation, Writing – Review & Editing. A.N.: Conceptualisation, Methodology, Writing – Review & Editing. R.P.: Conceptualisation, Methodology, Writing – Review & Editing. C.C.: Conceptualisation, Methodology, Writing – Review & Editing. T.R.: Writing – Review & Editing.
References 1. Erlei, A., Sharma, A. & Gadiraju, U. Understanding Choice Independence and Error Types in Human-AI Collaboration. in Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems 1–19 (Association for Computing Machinery, New York, NY, USA, 2024). doi:10.1145/3613904.3641946. 2. Goldfarb, A. & Lindsay, J. R. Prediction and Judgment: Why Artificial Intelligence Increases the Importance of Humans in War. Int. Secur. 46, 7–50 (2022). 3. Cabrera, Á. A., Perer, A. & Hong, J. I. Improving Human-AI Collaboration With Descriptions of AI Behavior. Proc ACM Hum-Comput Interact 7, 136:1-136:21 (2023).
4. Li, J. et al. Understanding the Effects of Miscalibrated AI Confidence on User Trust, Reliance, and Decision Efficacy. Preprint at https://doi.org/10.48550/arXiv.2402.07632 (2025). 5. Roeder, L. et al. A Quantum Model of Trust Calibration in Human–AI Interactions. Entropy 25, 1362 (2023). 6. Canonico, L. B., Flathmann, C. & McNeese, N. Collectively Intelligent Teams: Integrating Team Cognition, Collective Intelligence, and AI for Future Teaming. Proc. Hum. Factors Ergon. Soc. Annu. Meet. 63, 1466–1470 (2019). 7. He, G., Buijsman, S. & Gadiraju, U. How Stated Accuracy of an AI System and Analogies to Explain Accuracy Affect Human Reliance on the System. Proc ACM Hum-Comput Interact 7, 276:1-276:29 (2023). 8. Koopman, C. & Zammit-Mangion, D. Artificial Intelligence for Real-Time Tolerance to Critical Flight Data Errors in Large Aircraft. J. Aerosp. Inf. Syst. 21, 726–734 (2024). 9. Possibilities and Challenges for Artificial Intelligence in Military Applications. 10. GALTON, F. Vox Populi. Nature 75, 450–451 (1907). 11. Kerr, N. L. & Tindale, R. S. Group Performance and Decision Making. Annu. Rev. Psychol. 55, 623–655 (2004). 12. Orzechowski, K. P., Sienkiewicz, J., Fronczak, A. & Fronczak, P. When the crowd gets it wrong – the limits of collective wisdom in machine learning. Sci. Rep. 15, 22139 (2025). 13. Jayles, B. et al. The impact of incorrect social information on collective wisdom in human groups. J. R. Soc. Interface 17, 20200496 (2020). 14. Mavrodiev, P. & Schweitzer, F. Enhanced or distorted wisdom of crowds? An agent-based model of opinion formation under social influence. Swarm Intell. 15, 31–46 (2021). 15. Poli, R., Valeriani, D. & Cinel, C. Collaborative Brain-Computer Interface for Aiding Decision-Making. PLoS ONE 9, e102693 (2014). 16. Bhattacharyya, S., Valeriani, D., Cinel, C., Citi, L. & Poli, R. Collaborative Brain-Computer Interfaces to Enhance Group Decisions in an Outpost Surveillance Task. in 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) 3099–3102 (IEEE, Berlin, Germany, 2019). doi:10.1109/EMBC.2019.8856309.
17. Ma, S. et al. “Are You Really Sure?” Understanding the Effects of Human Self-Confidence Calibration in AI-Assisted Decision Making. in Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems 1–20 (Association for Computing Machinery, New York, NY, USA, 2024). doi:10.1145/3613904.3642671. 18. Arshad, S. Z., Zhou, J., Bridon, C., Chen, F. & Wang, Y. Investigating User Confidence for Uncertainty Presentation in Predictive Decision Making. in Proceedings of the Annual Meeting of the Australian Special Interest Group for Computer Human Interaction 352–360 (Association for Computing Machinery, New York, NY, USA, 2015). doi:10.1145/2838739.2838753. 19. Fleming, S. M. & Lau, H. C. How to measure metacognition. Front. Hum. Neurosci. 8, (2014). 20. Wokke, M. E., Cleeremans, A. & Ridderinkhof, K. R. Sure I’m Sure: Prefrontal Oscillations Support Metacognitive Monitoring of Decision Making. J. Neurosci. 37, 781–789 (2017). 21. Wokke, M. E., Achoui, D. & Cleeremans, A. Action information contributes to metacognitive decision-making. Sci. Rep. 10, 3632 (2020). 22. Murayama, K., Blake, A. B., Kerr, T. & Castel, A. D. When enough is not enough: Information overload and metacognitive decisions to stop studying information. J. Exp. Psychol. Learn. Mem. Cogn. 42, 914–924 (2016). 23. Di Gregorio, F., Maier, M. E. & Steinhauser, M. Are errors detected before they occur? Early error sensations revealed by metacognitive judgments on the timing of error awareness. Conscious. Cogn. 77, 102857 (2020). 24. Sadras, N., Sani, O. G., Ahmadipour, P. & Shanechi, M. M. Post-stimulus encoding of decision confidence in EEG: toward a brain–computer interface for decision making. J. Neural Eng. 20, 056012 (2023). 25. Schnuerch, R., Schmuck, J. & Gibbons, H. Cortical oscillations and event-related brain potentials during the preparation and execution of deceptive behavior. Psychophysiology 61, e14695 (2024). 26. Yeung, N. & Summerfield, C. Metacognition in human decision-making: confidence and error monitoring. Philos. Trans. R. Soc. Lond. B. Biol. Sci. 367, 1310–1321 (2012).
27. Eisma, J., Rawls, E., Long, S., Mach, R. & Lamm, C. Frontal midline theta differentiates separate cognitive control strategies while still generalizing the need for cognitive control. Sci. Rep. 11, 14641 (2021). 28. Cohen, M. X. & Donner, T. H. Midfrontal conflict-related theta-band power reflects neural oscillations that predict behavior. J. Neurophysiol. 110, 2752–2763 (2013). 29. Cohen, M. X. A neural microcircuit for cognitive conflict detection and signaling. Trends Neurosci. 37, 480–490 (2014). 30. Potts, G. F., Martin, L. E., Kamp, S.-M. & Donchin, E. Neural response to action and reward prediction errors: Comparing the error-related negativity to behavioral errors and the feedback-related negativity to reward prediction violations. Psychophysiology 48, 218–228 (2011). 31. Cao, D., Li, Y. & Niznikiewicz, M. A. Neural characteristics of cognitive reappraisal success and failure: An ERP study. Brain Behav. 10, e01584 (2020). 32. McDonnell, A. S. et al. This Is Your Brain on Autopilot: Neural Indices of Driver Workload and Engagement During Partial Vehicle Automation. Hum. Factors 65, 1435–1450 (2023). 33. Choe, J., Coffman, B. A., Bergstedt, D. T., Ziegler, M. D. & Phillips, M. E. Transcranial Direct Current Stimulation Modulates Neuronal Activity and Learning in Pilot Training. Front. Hum. Neurosci. 10, (2016). 34. Kilavik, B. E., Zaepffel, M., Brovelli, A., MacKay, W. A. & Riehle, A. The ups and downs of beta oscillations in sensorimotor cortex. Exp. Neurol. 245, 15–26 (2013). 35. Tan, H., Wade, C. & Brown, P. Post-Movement Beta Activity in Sensorimotor Cortex Indexes Confidence in the Estimations from Internal Models. J. Neurosci. 36, 1516–1528 (2016). 36. Parker, A. et al. Role of uncertainty in sensorimotor control. Philos. Trans. R. Soc. Lond. B. Biol. Sci. 357, 1137–1145 (2002). 37. Palmer, C. E., Auksztulewicz, R., Ondobaka, S. & Kilner, J. M. Sensorimotor beta power reflects the precision-weighting afforded to sensory prediction errors. NeuroImage 200, 59–71 (2019). 38. Plainis, S. & Murray, I. J. Reaction times as an index of visual conspicuity when driving at night. Ophthalmic Physiol. Opt. J. Br. Coll. Ophthalmic Opt. Optom. 22, 409–415 (2002). 39. Wickens, C. D. Multiple resources and mental workload. Hum. Factors 50, 449–455 (2008).
40. Barratt, E. S. Impulsiveness Defined Within a Systems Model of Personality. in Advances in Personality Assessment (Routledge, 1985). 41. Kothe, C. et al. The Lab Streaming Layer for Synchronized Multimodal Recording. bioRxiv 2024.02.13.580071 (2024) doi:10.1101/2024.02.13.580071. 42. Gramfort, A. et al. MEG and EEG data analysis with MNE-python. Front. Neurosci. 7, 267 (2013). 43. van Rossum, G. & de Boer, J. Interactively testing remote servers using the Python programming language. CWI Q. 4, 283–304 (1991). 44. McKinney, W. Data Structures for Statistical Computing in Python. in 56–61 (Austin, Texas, 2010). doi:10.25080/Majora-92bf1922-00a. 45. Harris, C. R. et al. Array programming with NumPy. Nature 585, 357–362 (2020). 46. Hunter, J. D. Matplotlib: A 2D Graphics Environment. Comput. Sci. Eng. 9, 90–95 (2007). 47. Waskom, M. L. seaborn: statistical data visualization. J. Open Source Softw. 6, 3021 (2021).