Human-in-the-Loop Governance for Brain-Computer Interface-Enabled Assistive Technologies: A Framework for Safe and Responsible Human-AI Co-Design

Brain-computer interfaces (BCIs) are reshaping assistive technologies, enabling AI-driven prosthetics, neuro-controlled mobility systems, and communication aids for people with disabilities. As these systems mature, AI agents interfaced with the human brain increasingly participate in decision-making and actuation, producing a human-AI co-design in which actions emerge from the combined influence of human volition and autonomous machine behaviour. This convergence blurs the boundary between human intention and machine agency, and harm arising from an AI-mediated action cannot be attributed straightforwardly to either the user or the system. This paper proposes a governance framework for BCI-enabled assistive technologies grounded in the human-in-the-loop (HITL) paradigm and the irreducibility of human consciousness. Drawing on an iceberg model of cognition spanning observable behaviours, brain signals, subjective experiences, feelings, and fundamental experiences, we argue that core dimensions of the mind resist reduction to data, logic, or computation, which makes sustained human oversight necessary rather than optional. We operationalise this argument through a Markov Decision Process (MDP) sandbox that fuses neural, physiological, behavioural, and contextual data to train and audit AI agents under continuous human supervision, embedding ethical evaluation, policy feedback, and diversity, equity, and inclusion safeguards. To move the proposal towards practice, we specify an evaluation pathway for the sandbox in realistic assistive scenarios such as neuro-controlled prosthetic driving and powered wheelchair navigation, with agreement between AI predictions and human responses, override and intervention rates, oversight latency, and audit traceability as candidate measures. We also set out limitations and implementation challenges, including the variability of neural signals, the burden that continuous supervision places on users and clinicians, and the absence of experimental validation at this stage. Case studies and controlled sandbox trials are proposed as the next step, and the framework is aligned with regulatory expectations such as the EU AI Act and international AI standards.