Quantifying Large Language Model Influence in Brain Computer Interface Communication for Amyotrophic Lateral Sclerosis

Large language models are integrated into brain-computer interfaces for communication, but accuracy does not show whether an emitted character depended on neural evidence or on the language-model prior. We retrospectively re-decoded 3,373 P300-speller selections from 47 people with amyotrophic lateral sclerosis, reconstructing a neural posterior and combining it with 25 language priors ranging from 5-grams to 46.7-billion-parameter models. Under held-out, per-source calibration and equal fusion weighting, the prior accounted for a participant-weighted mean 8.6% of posterior displacement (median across selections, 2.9%); the corresponding neural contribution fraction was 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. Results were similar across 21 neural language models. Accuracy alone does not reveal how strongly a fused brain-computer interface decision depends on its language prior.