2026-08-25 00:00 Papers Foundations & Methods Source language EN

Language Models Changed 4.4% of P300 Speller Choices in Data From 47 ALS Patients

Summary P300 spellers are a common brain-computer interface communication tool for people with ALS, and researchers have begun adding large language models as language priors to raise spelling accuracy. But accuracy alone cannot show whether an output came from the patient's brain signals or the model's prediction. A retrospective study re-decoded 3,373 P300 speller selections from 47 people with ALS, fusing reconstructed neural posteriors with 25 language priors ranging from 5-gram models to a 46.7-billion-parameter model. The prior accounted for a participant-weighted mean of 8.6% of posterior displacement, a neural contribution fraction of 0.914. In 4.4% of selections, the fused system produced the intended character where neural evidence alone would not have. Results were similar across 21 neural language models. The work is a preprint and has not been peer reviewed.

Why it matters This is the first systematic measure of how much LLM priors actually steer P300 BCI output. Neural signals still dominate at 91.4%, but 4.4% of outputs depended on the language model rather than the patient's brain activity, turning the ethical and technical question of whose intent a fused BCI represents into something that can be calculated.

BCIwiki (bciwiki.com) — When a P300 speller BCI fuses a patient's brain signals with a large language model, how much of the final output actually comes from the brain? Researchers at the Department of Neurology, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, USA, retrospectively re-decoded 3,373 P300-speller selections from 47 people with amyotrophic lateral sclerosis, combining a reconstructed neural posterior with 25 language priors ranging from 5-grams to 46.7-billion-parameter models, according to a preprint uploaded to medRxiv on August 25, 2026.

Under held-out, per-source calibration and equal fusion weighting, the language prior accounted for a participant-weighted mean 8.6% of posterior displacement (median across selections, 2.9%), yielding a neural contribution fraction of 0.914 (95% CI, 0.896-0.934). In 4.4% of selections (95% CI, 3.5-5.3%), the fused system emitted the intended character although neural evidence alone would not have selected it — meaning those outputs depended entirely on the language model's prediction. Results were consistent across 21 neural language models tested. The authors concluded that accuracy alone does not reveal how strongly a fused BCI decision depends on its language prior. The work has not been peer reviewed.

Compiled by BCIwiki from public sources

Sources · 1

medrxiv.org 2026-08-25

ALS Decoding Algorithms Non-invasive BCI

Read original ↗

Suggest a correction Revisions · none 中 / EN