Stabilization-Responsiveness Trade-offs in Continuous Shared-Control for Invasive Brain-Computer Interfaces
Intracortical brain-computer interfaces (iBCIs) can enable people with paralysis to control assistive devices, but reliable operation in dynamic environments remains limited by fluctuations in decoded neural commands. Here we develop a confidence-modulated AI-brain shared-control framework in which an artificial intelligence copilot adaptively integrates the decoded neural commands with a probabilistic temporal prior to stabilize execution while preserving user intent. In two macaques performing closed-loop virtual navigation tasks in complex environments, shared-control nearly eliminated execution-level failures, including obstacle collisions and target overshoot, while maintaining the directional structure of neural commands. Abrupt target changes revealed a boundary condition: temporal stabilization transiently delays responsiveness when recent history was no longer predictive. Offline replay showed that resetting the temporal prior eliminated this lag and restored performance, demonstrating that the impairment was algorithmic rather than a failure of neural decoding. These results provide a mechanistic characterization of confidence-modulated AI-brain shared control for continuous intracortical BCI navigation and identify design principles for safer and more reliable neuroprosthetic control in dynamic environments.