AI agents at the brain-computer interface: separating inference from control
- 1BIDMC
Abstract
In medicine, AI agents are moving from generating text to executing actions, making uncertainty from upstream decoders a control problem. We studied this at the brain-computer interface using 1,065 episodes from 47 people with amyotrophic lateral sclerosis and five language models. Prompting agents with reconstructed decoder confidence never reduced unfaithful execution below a deterministic gate at matched coverage; two models were significantly worse. 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 resolvers risk-coverage frontier, but a hybrid architecture in which models proposed semantic corrections and an external gate retained admission authority extended coverage beyond the resolver in five of ten models without observed unfaithful executions. These results separate inference from control in agentic neurotechnology.
Related
- Brain‐computer interface training for motor recovery after stroke
- Encoding and decoding of acupuncture-evoked electrical signals and their value in acupuncture–brain-computer interface-integrated closed-loop regulation
- Effects of combined non-invasive brain stimulation and brain-computer interface on upper extremity motor function after stroke: a systematic review and meta-analysis
- The application of brain-computer interface on cognitive function and activities of daily living in stroke patients: a systematic review and meta-analysis
- A Retrospective Real-World Assessment of Motor Status After Stroke Using a Home-Based Brain-Computer Interface
BibTeX
@misc{gorenshtein2026agents,
title = {{AI agents at the brain-computer interface: separating inference from control}},
author = {A. Gorenshtein and M. Omar and E. L. Jia and Y. Adiniaev and O. Daniel and J. Kruskal and M. Ahmed and O. R. Brook and E. Klang and Y. Barash},
howpublished = {medRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.13.26362955},
url = {https://www.medrxiv.org/content/10.64898/2026.09.13.26362955v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - AI agents at the brain-computer interface: separating inference from control
AU - A. Gorenshtein
AU - M. Omar
AU - E. L. Jia
AU - Y. Adiniaev
AU - O. Daniel
AU - J. Kruskal
AU - M. Ahmed
AU - O. R. Brook
AU - E. Klang
AU - Y. Barash
PY - 2026
DA - 2026/09/13
PB - medRxiv
DO - 10.64898/2026.09.13.26362955
UR - https://www.medrxiv.org/content/10.64898/2026.09.13.26362955v1
AB - In medicine, AI agents are moving from generating text to executing actions, making uncertainty from upstream decoders a control problem. We studied this at the brain-computer interface using 1,065 episodes from 47 people with amyotrophic lateral sclerosis and five language models. Prompting agents with reconstructed decoder confidence never reduced unfaithful execution below a deterministic gate at matched coverage; two models were significantly worse. 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 resolvers risk-coverage frontier, but a hybrid architecture in which models proposed semantic corrections and an external gate retained admission authority extended coverage beyond the resolver in five of ten models without observed unfaithful executions. These results separate inference from control in agentic neurotechnology.
ER -