Neuro-Decoupled Team (NDT) Decision-Making

• Neuro-Decoupled Teaming is a decision-making framework that uses pre-decisional neural signals from EEG to aggregate team confidence while excluding traditional behavioral reports.
• The system employs a collaborative BCI pipeline with rigorous EEG preprocessing, feature extraction, and SVM classifiers to generate reliable neural confidence scores under varying cognitive loads.
• Empirical results in a VR drone surveillance task show that NDT outperforms traditional aggregators, achieving 98.35% accuracy and effectively mitigating AI-induced biases.

A Neuro-Decoupled Team (NDT) is a team decision-making architecture that excludes overt behavioural reports such as choices, reaction times, and subjective confidence in favor of a purely neural aggregation mechanism. The NDT utilizes pre-decisional brain–computer interface (BCI) confidence signals derived from single-trial electroencephalography (EEG), enabling team decisions that are insulated from systematic behavioural biases—particularly those induced by deceptive or erroneous AI cues. This approach is positioned in direct contrast with traditional behaviour-based aggregators such as majority vote or behaviourally-weighted voting schemes, which are highly susceptible to “anti-wisdom of crowds” effects under correlated error conditions (Baker et al., 24 Nov 2025).

1. Principles of Neuro-Decoupled Teaming

Neuro-Decoupled Teaming (NDT) is defined by the exclusive use of a neural confidence signal, $C_{BCI}(t)$, acquired prior to the subject’s behavioral commitment. All behavioural information—discrete choices, reaction time (RT), and subjective confidence—is discarded. The rationale is that neural signals, extracted before the influence of a deceptive AI can propagate to overt behavior, remain insulated from externally induced biases that can collapse behavioural fidelity across team members.

By contrast, classical aggregation approaches—majority vote, RT-weighted, or subjective confidence-weighted voting—implicitly assume each individual response represents an unbiased sample of the ground truth. When AI feedback is both influential and systematically incorrect, these aggregators not only fail, but can amplify error prevalence due to correlated bias. NDT’s decision rule replaces all behavioural weights with neural weights, operationalized as confidence scores produced by BCI classifiers prior to the individual's decision (Baker et al., 24 Nov 2025).

2. Collaborative BCI Architecture and Analytical Pipeline

The NDT architecture is implemented using a collaborative brain-computer interface (cBCI) pipeline designed for precise, temporally resolved neural decoding.

EEG Acquisition is performed using a 32-channel LiveAmp (10–20 montage, 500 Hz sampling, <30 kΩ impedances). Signals are continuously recorded during task performance, synchronized via LabStreamingLayer event markers (e.g., ReticleOn, ButtonPress).

Preprocessing includes:

• Band-pass filtering (0.1–30 Hz, FIR) with 50 Hz notch,
• Concatenation across sessions for independent-component analysis (ICA, FastICA) and artefact removal (ocular/muscle),
• Epoching around task events (–200 ms to +800 ms for ReticleOn; –500 ms to +300 ms for ButtonPress),
• Baseline correction and exclusion of trials with outlier SVM scores or RT > 2.6 s.

Feature extraction is performed per channel for each ReticleOn epoch:

• Time-domain: mean amplitude, maximum amplitude, variance,
• Frequency-domain: Welch PSD in Theta (4–8 Hz), Alpha (8–13 Hz), and Beta (13–30 Hz) bands.

Classifier design employs subject-specific SVMs (RBF/linear kernel), with:

• Feature selection (SelectKBest by mutual_info_classif, top 5 features per subject),
• Class balancing (random undersampling),
• Hyperparameter tuning ($C\in[0.1,100]$, gamma = {“scale”,“auto”}) via 5-fold cross-validation.

The SVM decision function $f(x_t)$ (signed distance to hyperplane) is min–max normalized across all held-out trials:

$$C_{BCI}(t) = \frac{\bigl|f(x_{t})\bigr| - \min_{u}\bigl|f(x_{u})\bigr|}{\max_{u}\bigl|f(x_{u})\bigr| - \min_{u}\bigl|f(x_{u})\bigr|} \in [0,1]$$

The NDT team decision $\hat y_{NDT}$ is given by the class $$c \in \{\mathrm{Target}, \mathrm{Non\text{-}Target}\}$$ maximizing the sum of BCI confidences across team members:

$$\hat y_{NDT} = \underset{c \in \{\mathrm{T}, \mathrm{NT}\}}{\arg\max}\; \sum_{i=1}^N \mathbf{1}_{\{c^{(i)}=c\}} C_{BCI}^{(i)}$$

Ties are resolved in favor of “Target.”

3. Empirical Evaluation and Quantitative Results

NDT was evaluated in a VR drone surveillance paradigm (N=17 participants) involving first-person quadcopter flights in Unreal Engine. Each trial required classification of a 3D object (at 300 m) following display of an AI-generated reticle cue (blue=Non-Target, red=Target) for 2.5 s, a button-press decision, and a self-rated confidence report (0–100, joystick tilt).

Workload manipulation involved low (bright midday) and high (50% reduced illumination, 9° solar angle, “night”) conditions. AI deception was implemented by making the AI cue systematically incorrect on 100% of high workload trials, inducing a worst-case correlated bias.

Aggregated performance under deception (N=8 simulated teams, ~7.3 million decisions) demonstrates clear superiority for neural-only NDT aggregation:

Aggregation Strategy	Accuracy (%)
Majority Vote	44.41
Subjective Confidence-Weighted	51.09
Average Individual	84.43
Best Individual	92.17
NDT (SVM Confidence-Weighted)	98.35

Synergy gain over the best individual ($G = 6.18$ percentage points). Statistical comparisons (t-tests) confirm superiority:

• NDT vs. Majority: $t=598.75, p<.001$,
• NDT vs. Subjective Confidence: $t=518.30, p<.001$ (Baker et al., 24 Nov 2025).

4. Mechanisms of Neuro-Behavioral Decoupling

Under deception, the fidelity of behaviour-based signals collapses. The Pearson correlation between subjective confidence and correctness is weak but significant ($r = 0.194, p<.001$), indicating that self-assessed confidence loses utility as a correctness indicator. By contrast, the BCI-derived neural confidence signal remains statistically independent of the corrupted metacognitive judgments ($r = -0.017, p = .380$).

Direct statistical comparison (Steiger’s Z-test: $Z = 7.80, p<.001$) establishes the neural confidence signal as a superior predictor under intentional AI-led bias. Feature weighting adapts to context:

• Under low workload (“Autopilot”/Rhythmic Stability), the SVM prioritizes: low variance at left prefrontal (Var_F9), increased posterior alpha power, and sensorimotor beta modulations.
• Under high workload (“Effort & Instability”), the classifier emphasizes: large positive deflection at F9 and FC6 (LPP, deliberation), theta power at CP2/CP6 (conflict monitoring), and motor-cortex variance (Var_C4, response selection uncertainty).

Event-related potentials (ERPs) time-locked to ReticleOn show rapid negative deflection (~350 ms) under low workload (automatic processing), and a large sustained late positive potential (LPP) under high workload (extended cognitive control), confirming context-dependent neural strategies (Baker et al., 24 Nov 2025).

5. Broader Significance and Comparative Analysis

Behaviour-based team strategies are vulnerable to “anti-wisdom of crowds” failure, wherein correlated AI-induced biases lead to systemic error amplification. NDT’s implicit BCI channel provides “early dissent” at a neural level, flagging conflict before overt team errors manifest. Relevant domains include remote piloting, intelligence analysis, cyber-defense, and medical decision teams.

Potential misconceptions may arise regarding the sufficiency of behavioural aggregation in high-stakes human–AI teaming; empirical evidence demonstrates catastrophic failure modes under systematic AI deception, with majority vote accuracy dropping near chance levels and failing to surpass individual baselines. NDT not only preserves but augments team performance, leveraging neural adaptation to defend against error propagation (Baker et al., 24 Nov 2025).

6. Limitations and Current Research Trajectories

The reported NDT implementation is based on offline simulations rather than real-time closed-loop validation. AI deception was standardized (100% wrong) rather than variable as might occur in operational AI deployments. Task structure constrained operators to passive observation; introducing active control could alter cognitive signatures and team-level neural dynamics. Classifiers were trained and evaluated in a participant-specific fashion, leaving generalizability to unseen users untested.

Future research directions include:

• Closed-loop cBCI experiments for live validation of NDT synergies.
• Systematic testing of resilience to temporal patterns of AI error (intermittent vs. persistent).
• Addition of active piloting to probe how agency modulates neural conflict markers.
• Exploration of F9-based biomarkers (maximal amplitude) as real-time triggers for neural dissent or alerting.
• Development and validation of subject-independent classifiers and earlier predictors (e.g., pre-stimulus neural signatures) (Baker et al., 24 Nov 2025).

7. Summary and Outlook

The Neuro-Decoupled Team operationalizes a pre-decisional, adaptive neural aggregation strategy—shifting from features indicative of efficient “autopilot” processing to conflict-monitoring when deception or cognitive demand increase. This system provides a robust defence against correlated AI errors that nullify conventional behaviour-based team aggregation. NDT establishes pre-response BCI confidence as a reliable, scalable signal for team decision-making in high-stakes, deception-prone environments, serving as a foundation for resilient human–AI collaboration (Baker et al., 24 Nov 2025).