BCIwiki (bciwiki.com) — According to BIDMC, a preprint examining AI agents executing commands at the brain-computer interface (BCI) was uploaded to medRxiv on 2026-09-17. It has not been peer reviewed. The study tested five language models using 1,065 episodes from 47 people with amyotrophic lateral sclerosis, examining how uncertainty in decoding affects downstream execution.
The researchers supplied agents with reconstructed decoder confidence, information about how certain a decoded result was. At matched execution coverage, this never reduced unfaithful execution below a deterministic gate; two models were significantly worse. Coverage here concerns the proportion covered by execution, not insurance coverage. Apparent safety gains of up to 22 percentage points reflected acting less often, sometimes through invalid tool calls rather than explicit abstention.
A post-hoc fair-information test gave ten models the same command vocabulary as a deterministic resolver. No direct agent arm improved on the resolver's risk-coverage frontier, the achievable risk levels across different degrees of coverage. A hybrid architecture, however, let models propose semantic corrections while an external gate retained admission authority. It extended coverage beyond the resolver in five of ten models without observed unfaithful executions.
The authors interpret these findings as supporting a separation between inference and control in agentic neurotechnology: a model can propose corrections to a command's meaning while an external mechanism decides whether execution is permitted. The hybrid result means that no unfaithful executions were observed in this test, not that errors are impossible in actual use.