Casey Harrell said close to two million words in nearly two years. His lips barely moved.

Harrell is 47 and has ALS. His arms and legs are weak, his speech too slurred for strangers to follow. In 2023, UC Davis neurosurgeon David Brandman placed four microelectrode arrays in the left precentral gyrus of his brain, the strip of cortex that coordinates the muscles used for speaking. Over the following two years Harrell clocked more than 3,800 hours on the resulting system at home, with no researcher in the room, producing over 183,000 sentences at an average of 56 words per minute. That is well under half the speed of normal conversation, but fast enough to actually hold one. Brandman, neuroscientist Sergey Stavisky and their colleagues published the results in Nature Medicine in June 2026.

One participant. One brain. Worth holding onto as the numbers below get thrown around.

What the decoding actually does

Nothing here reads thoughts. The electrodes sit above the part of the brain that issues commands to the tongue, jaw and larynx, and they listen in on those commands as they are sent. A trained model turns the firing patterns into phonemes, the phonemes into words, the words into synthesised speech tuned to sound like Harrell before the illness. He attempts to speak. The muscles ignore him. The computer does not.

Accuracy has climbed steeply. An earlier version of the same system, described by Nicholas Card and colleagues in a paper published in the New England Journal of Medicine in 2024, held 97.5 per cent accuracy over 8.4 months with a 125,000-word vocabulary, running at about 32 words per minute. The 2026 follow-up pushed word accuracy above 99 per cent in controlled testing and nearly doubled the speed, while shifting the whole thing out of the lab and into a living room.

How many people have actually received one

Very few, which is the part the headlines tend to skip. A review by K. Michelle Patrick-Krueger, Ian Burkhart and Jose Contreras-Vidal in Nature Reviews Bioengineering counted every identified clinical trial of implanted brain-computer interfaces from the first chronic implantation in 1998 through to December 2023 and found 67 participants across 21 research groups worldwide. Twenty-five years. Sixty-seven people. No device approved for sale by any regulator.

Neuralink has moved fastest on volume. Reuters reported in January 2026 that the company had 21 participants enrolled in trials worldwide, two years after its first human implant, with no serious device-related adverse events on the record. Participants have driven cursors, played games, and more recently steered powered wheelchairs. All of it remains investigational.

Two routes into the skull

There is a live disagreement in the field about how far in you need to go. Neuralink threads flexible electrodes directly into the motor cortex using a surgical robot, which buys signal quality at the cost of open brain surgery. Synchron takes the plumbing route: its Stentrode is a mesh device pushed up through the jugular vein and parked in a vessel sitting against the motor cortex, reading through the vessel wall.

Lower resolution, far lower barrier.

In its COMMAND feasibility study, Synchron implanted six people with severe upper-limb paralysis and reported no device-related serious adverse events over twelve months, with a median deployment time of twenty minutes. It is the same route neurointerventional radiologists already use routinely to treat strokes. Whether sixteen electrodes reading through a blood vessel can ever match the 1,024 packed directly into tissue is unresolved, and both companies have obvious commercial reasons for the answer they prefer.

The bit nobody puts in the launch video

Ian Burkhart was 19 when a diving accident left him paralysed below the elbows. Four years later, in 2014, researchers at Ohio State implanted an array in his motor cortex, and he became the first person to move his own hand again using signals decoded from his brain. Then in 2021 the device came out. Funding for the trial had run dry after seven years, and his scalp had developed a persistent infection around the connector that protruded from his skull.

Burkhart told MIT Technology Review that he had regained a function and then lost it a second time, which he described as losing some sense of himself. Speaking at Georgia Tech, he said he would still take the risk again, and he has since founded a coalition to put the experience of trial participants in front of designers. Others fared worse. The same outlet documented a woman whose implant was removed against her wishes after the company behind it ran short of money. Ethicists, including Marcello Ienca at the Technical University of Munich, have argued that this kind of explantation may breach human rights.

Nobody signs a consent form imagining the hardware will outlast the company.

What has to happen next

Regulatory approval is the gate everything else waits behind, and no permanently implanted motor or speech interface has passed through it. Synchron has been working towards a pivotal trial that could support the first premarket approval filing for such a device, backed by a $200 million Series D round announced in November 2025. Neuralink’s own pivotal trial design remains unsettled. Even on optimistic timelines, a doctor writing a prescription for one of these things is years away rather than months.

Meanwhile the demonstrations keep landing, and each one is a person rather than a benchmark. Harrell used his system to remind his daughter what his voice sounded like, since she had been too young to remember. That is the thing being built here, underneath the electrode counts and the funding rounds and the promises about telepathy. A man telling his kid what he sounds like.

The engineering has largely stopped being the hard part. Working out who keeps the lights on for a voice that now runs on somebody else’s server is the question the field has barely started to answer.