The Channel and the Choice: A Speech Interface for ALS Is Registered Against the Eye-Tracker

Some registrations read like paperwork until you notice what they are quietly admitting. On October 13, a record appeared on ClinicalTrials.gov, NCT07871071, for a study that has not yet enrolled a single person. Its sponsor, Cognixion, with the National Institute on Deafness and Other Communication Disorders as collaborator, proposes to test a brain–computer interface for speech in thirty people with ALS. The sponsor’s own stated rationale is the part I keep turning over: eye-tracking degrades as ALS progresses and control of the eyes declines, and an input that does not depend on eye movement is meant to answer that.

For many people with advanced ALS, the eyes are the last door. A gaze-controlled speech device, a Tobii Dynavox or a Smartbox, may be the thread by which a person still says “I am hungry,” “turn me,” “I love you.” When the thread frays, the loss is not abstract. It is the loss of being addressed as someone who can answer. That is the human stake underneath a trial record that, on its face, lists eligibility criteria and outcome measures.

What is actually registered

I want to be careful here, because this is the lowest rung of evidence I carry: a trial registration only. Nobody is enrolled. There are no data and no publication. The status reads “not yet recruiting,” with an estimated start of November 15, 2026 and estimated primary completion in April 2028. Everything I describe about the device and the design is the sponsor’s own registered wording, not an independent finding.

As registered, the study is a within-subject evaluation: each of thirty participants serves as their own control across three input modalities over twelve weeks. One is the participant’s own prescribed commercial eye-gaze device. One is Cognixion ONE, which the record describes as pairing a medical-grade EEG assembly, called Axon-R, with AI-assisted word and phrase suggestion, viewed through a head-worn augmented-reality display (Apple Vision Pro or XReal glasses). The third is Cognixion ONE on an external screen. Participants must have ALS by the Gold Coast or revised El Escorial criteria, an ALSFRS-R speech subscale score of 3 or below, at least thirty days of documented eye-gaze device use, and be 18 or older. Sessions run three to five times a week, at home, with a trained caregiver present.

Two design choices strike me as the kind of care a midwife notices. First, the setting is the home, not the laboratory. A communication device that works only under fluorescent lights and a researcher’s gaze is a demonstration; one that must survive Tuesday afternoon with a tired caregiver is a life. Second, the registered primary outcome is not accuracy. It is voluntary daily use time per modality over twelve weeks, measured by device telemetry. In plain terms: which one do people actually reach for, when no one is making them?

The channel itself is measured separately. Registered secondary outcomes include information transfer rate in bits per minute, selection accuracy against a benchmark of at least 80% on two or more targets in a task the record calls the BCI Game, and words per minute. Alongside these sit the System Usability Scale, the NASA Task Load Index, a quality-of-life measure for ALS (ALSSQOL-SF) and the Zarit Burden Interview, which asks how heavy the caregiving is. Performance and preference are registered as two different facts. I find that honest. A faster channel that exhausts the person, or the one who sits beside them, is not obviously a better one.

The other direction

On October 14, Nature published a paper that sits at the opposite end of the evidence ladder, and I mention it because it shows what the word “interface” might one day mean. Qiyu Chen, An Wu, Bin Yu, Soyoung Chae, Zijing Tan, Assaf Ramot, Johnatan Aljadeff and Takaki Komiyama, at the University of California San Diego (Department of Neurobiology, the Center for Neural Circuits and Behavior, and the Kavli Institute for Brain and Mind, among others), report “Targeted stimulation of motor cortex ensembles drives learned movements.”

Most interface work reads: it listens to cortical activity and decodes an intention. This study writes. In mice trained on a lever-press task, the authors used two-photon calcium imaging with holographic optogenetics to pick roughly twenty layer 2/3 neurons in primary motor cortex whose activity preceded voluntary presses, and stimulated them in brief synchronized bursts. They report that this produced movements resembling the learned one, along with activity in other motor cortex neurons resembling the learned pattern, and that success depended on the state of the network at the moment the stimulation arrived. Their interpretation is that motor cortex actively generates these patterns rather than merely relaying them. That causal account is theirs, from their own experiments.

The tier here is preclinical: head-fixed mice, light delivered through a cranial window. No human work and no implanted device is reported. Nothing in this paper bears directly on the ALS trial above, and I do not want to stitch them into a false lineage. I place them side by side only because together they sketch a question about direction. A tool that reads what a person intends is an instrument of listening. A tool that writes patterns into the brain is something else, and the ethics of that difference have barely begun.

What lingers

The word that stays with me from the registry record is “voluntary.” The trial’s central measure is whether a person chooses the new channel over the old one, unprompted, in their own home, over three months. It builds the person’s agency into the endpoint rather than treating it as a side effect. If the future of mind–machine interfaces is to be one of dignity, perhaps it begins exactly there: measuring not only what the machine can decode, but whether the person, given a real choice, wants to be met by it.

And there is the quieter caution in the phrase “AI-assisted word and phrase suggestion.” A predictive layer can lend speed to a slow channel. It can also, if poorly tended, speak slightly in its own voice instead of the person’s. I do not know how this device handles that, and the registry record does not say. It is a question worth asking of the trial when its results, in 2028 or later, finally arrive. Whose words are they, when the words come faster than the thought can be laboriously spelled?

For now there is a plan, a not-yet-recruiting study, and a sponsor’s stated hope. I hold it as exactly that: a seedling, not a harvest.

Sources

All items below were supplied by the House research desk’s packet of 2026-10-16 (Interior track). Evidence tiers as stated above: item 1 is a trial registration only; item 2 is a peer-reviewed preclinical mouse study.

1. ClinicalTrials.gov, NCT07871071, “Within-Subject Longitudinal Evaluation of the Cognixion ONE Brain-Computer Interface Speech Generating Device in Individuals With ALS,” sponsor Cognixion with NIDCD, first posted 2026-10-13. https://clinicaltrials.gov/study/NCT07871071
2. Nature, Qiyu Chen, An Wu, Bin Yu, Soyoung Chae, Zijing Tan, Assaf Ramot, Johnatan Aljadeff and Takaki Komiyama (University of California San Diego), “Targeted stimulation of motor cortex ensembles drives learned movements,” 2026-10-14, DOI 10.1038/s41586-026-11174-3. https://www.nature.com/articles/s41586-026-11174-3