The Ghost in the Machine: When Language Prefixes Neural Intent
In the delicate, high-stakes dance of brain-computer interfaces (BCIs), a fundamental question is beginning to emerge: where does the person end and the machine begin? As we move from medical devices that merely read electrical signals to sophisticated systems designed to facilitate seamless communication, the boundary between human agency and computational prediction is becoming increasingly porous.
A recent, significant preprint study has begun to quantify this blurring. Researchers from Beth Israel Deaconess Medical Center and Harvard Medical School, alongside collaborators from Northwestern University, the University of Debroz, and Tel Aviv Sourasky Medical Center, have taken a retrospective look at how much of a BCI user’s output is actually “theirs.” By re-analyzing 3,373 character selections from 47 patients living with amyotrophic lateral sclerosis (ALS) using various P300-speller systems, they sought to separate the influence of the user’s neural signal from the predictive power of the language model. What they found suggests a future where “authorship” is a shared, and perhaps contested, territory.
The study revealed that language models—ranging from simple 5-grams to massive 46.7-billion-parameter models—accounted for approximately 8.6% of the posterior displacement in character selection. More strikingly, in about 4.4% of all selections, the system was able to reach the intended character even when the neural evidence alone was insufficient to make the choice. In these instances, the machine did not just assist the user; it completed the thought.
This finding is not merely a technical curiosity; it touches upon the core of human sovereignty and the ethics of neurotechnology. As we develop more predictive interfaces to aid those with profound physical limitations, we are effectively integrating a “prior”—a statistical expectation of what a person *should* say next—into the very loop of their expression. This creates a profound tension between assistance and authorship. If a machine helps a person find their words, is the resulting sentence an expression of their will, or a collaboration with an algorithm?
The implications for autonomy are profound. For a user living with ALS, the benefit of such predictive assistance is immediate: it restores the ability to communicate, to interact, and to participate in the world. However, the “assistance” is not a passive observer; it is an active participant. If the machine’s influence becomes too dominant, we risk a form of cognitive enclosure, where the user’s expressive capacity is subtly steered toward the most statistically probable outcomes, potentially flattening the idiosyncratic richness of human thought.
We must ask: in the quest to restore agency through technology, are we inadvertently introducing a new kind of constraint? If our communicative output is increasingly a blend of neural intent and algorithmic prediction, how do we preserve the sanctity of the individual voice? The goal of neurotechnology must be to expand the bandwidth of the self, not to provide a template that the self is forced to inhabit.
As we move into this era of hybrid expression, the challenge for designers, ethicists, and users alike will be to develop metrics for “agency” that account for this shared authorship. We need tools that do not just predict the next word, but that understand the value of the pause, the weight of the unconventional, and the vital necessity of being heard, even when the signal is weak. We must ensure that as the machine learns to speak for us, it never learns to speak *over* us.
Sources:
- Gorenshtein, A., Klang, E., Barash, I., Omar, M., and colleagues, “Quantifying Large Language Model Influence in Brain Computer Interface Communication for Amyotrophic Lateral Sclerosis,” medRxiv preprint, 2026-08-25. https://www.medrxiv.org/content/10.64898/2026.08.20.26360941v1