Human-AI Decision Making Costs in Synthetic Teams (arxiv.org)

Researchers tested an implicit collaborative brain–computer interface (cBCI) as a safeguard against misleading AI in a virtual-reality drone surveillance task. Teams of human operators faced high cognitive workload and systematically deceptive AI cues; traditional behavior-based aggregation (e.g., majority vote) crashed—team accuracy fell to 44% under deception. In contrast, a Neuro-Decoupled Team (NDT) that aggregated pre-response EEG-derived confidence scores (an implicit BCI signal decoupled from overt behavior) maintained 98% accuracy, a statistically significant synergistic improvement over the best individual (p < .001). Technically, the benefit arises from neuro–behavioral decoupling: the BCI’s pre-decision neural signatures remained predictive when subjective confidence and behavior were biased by the AI. The system learned to interpret context-dependent neural markers—using signatures of efficient “autopilot” processing in easy cases and indicators of effortful deliberation when cognitive conflict emerged—thereby resisting AI-induced error. Implications: neural signals can provide an orthogonal, resilient cue for team aggregation, suggesting new designs for human–AI teaming in high-stakes settings (drones, healthcare, defense). The work points to promising robustness gains but also raises practical questions about EEG reliability, privacy, and deployment in real-world operations.