Most confusion in this field comes from vocabulary. The same phrase covers a 2025 lab demo in healthy volunteers and a device implanted in a person with ALS under an FDA exemption. Once you know seven terms, you can place almost any headline on the ladder: lab demo, early clinical trial, cleared for use, on the market. Here is the list, with real devices and numbers as of late October 2026:
Brain-computer interface (BCI)
Electrode array and channel count
Invasive versus non-invasive
Decoding and word error rate
Investigational device exemption (IDE)
Breakthrough Device designation
510(k) clearance versus PMA
Brain-computer interface and the ladder
A brain-computer interface records neural activity, decodes what the person intends, and turns that into a command for an external device: a cursor, a speech synthesizer, a robotic arm. Some systems also stimulate the brain, which makes them bidirectional. The term says nothing about maturity, and that is the problem.
Take Neuralink. The company told Reuters in January 2026 that it had 21 participants enrolled in trials worldwide, with zero serious device-related adverse events. Both figures are company-reported. The device is investigational. It is not a commercial product, whatever the “high-volume production” talk implies. No permanently implanted BCI has reached FDA premarket approval yet.
So every time a story says a BCI “works,” ask which rung it is on. A paper, a first-in-human implant, a cleared component and an approved system are four different claims. Which of those four do you see most often blurred together in your own feed?
Electrode arrays and channel count
The electrode array is the part that touches (or sits near) the brain. Channel count is how many separate recording sites it has. The oldest human workhorse is the Utah array: 96 electrodes per array, silicon needles up to 1.5 mm long, made by Blackrock Neurotech and implanted in people since 2004. A medRxiv preprint on the BrainGate trials covers 14 participants enrolled between 2004 and 2019, each with one or two arrays.
Newer devices push the numbers up:
Neuralink N1: 1,024 electrodes on 64 threads, inserted into cortex by a surgical robot
Precision Neuroscience Layer 7: 1,024 electrodes on a thin film one-fifth the thickness of a human hair, resting on the cortical surface
Paradromics Connexus: 421 microelectrodes, recording from individual neurons
Synchron Stentrode: 16 electrodes, delivered through a vein
Do not rank these by count alone. A penetrating electrode hears single neurons, a surface film hears populations, and a vein-mounted array hears from a distance. Sixteen well-placed channels can control a cursor. More channels buy bandwidth and finer decoding, but they cost surgical risk and engineering difficulty.
Invasive, non-invasive, and the places between
Invasive usually means electrodes inside the skull. It splits into intracortical (penetrating the cortex, like the Utah array), electrocorticography or ECoG (on the cortical surface, like Precision’s film) and endovascular (inside a blood vessel, like Synchron’s catheter-delivered Stentrode). Non-invasive means sensors outside the head: EEG (scalp electrodes) or MEG (magnetic field sensing in a shielded room).
The trade is signal quality against surgery. Meta’s Brain2Qwerty shows both sides. The first version, from early 2025, used 35 healthy volunteers typing sentences. MEG gave an average character error rate of 32%, while EEG gave 67%. The newer v2 reports 61% average word accuracy from about 22,000 sentences across nine volunteers, each recorded for 10 hours. That is a real improvement over the roughly 8% of earlier non-invasive methods. It is also a lab demonstration in healthy people who physically type, using a scanner that needs a shielded room. Nobody is wearing that at home.
Keep medical and consumer separate here. A research MEG decoder, a Breakthrough-designated EEG headset for people with ALS (Cognixion’s ONE Axon is one), and a pair of focus-tracking EEG headphones are three different products with three different evidence bars.
Decoding and word error rate
Decoding is the software step: a model maps neural features to an intended output. Its quality is reported as words per minute, word error rate (WER) and, critically, vocabulary size. Those three numbers only mean something together.
The 2023 Nature paper from Frank Willett and colleagues at Stanford reported 62 words per minute with a 23.8% word error rate on a 125,000-word vocabulary, in a participant with ALS. That was 3.4 times faster than the prior speech BCI, and the first large-vocabulary demonstration, though natural speech runs around 160 words per minute. A parallel ECoG study using a 253-channel high-density array reached a median 78 words per minute at a median 25% WER.
Later intracortical work improved the error rates. A UC Davis system described in the NEJM paper reached 99.6% accuracy on a 50-word vocabulary on day one, after 30 minutes of attempted-speech training, and 97.5% accuracy beyond eight months, per the medRxiv version. A 99.6% claim on 50 words and a 97.5% claim on a large vocabulary are different achievements. All of these results come from one or a handful of participants, so treat them as strong proof of function, not proof of a product.
IDE, Breakthrough designation, and 510(k) versus PMA
Three regulatory terms get confused constantly, so take them in order.
An investigational device exemption (IDE) lets an unapproved device be used in a clinical study. Many IDE studies are early feasibility studies (EFS): small, safety-first, with efficacy only explored. Synchron’s COMMAND was the first IDE trial of a permanently implanted BCI. In six patients it reported no device-related serious adverse events causing death or permanent increased disability over 12 months. Paradromics received its IDE in November 2025, and its first Connect-One implant took place on June 17, 2026 at the University of Michigan. The first participant has motor neuron disease and will be followed for six years. Every paper, funding round, and trial update we report on ends up summarized and searchable in NeurotechMag Pro if you want to track programs like these across their full timelines.
Breakthrough Device designation is a voluntary FDA program that gives prioritized review and more frequent FDA contact for devices targeting life-threatening or irreversibly debilitating conditions. It is a development accelerator, not an approval. Synchron, Neuralink, Precision, Blackrock and Cognixion hold designations, and CorTec added one in April 2026 for stroke rehabilitation with its Brain Interchange system.
Finally, 510(k) clearance means FDA found a device substantially equivalent to an existing one. PMA (premarket approval) is the full approval route and demands a confirmatory trial showing safety and effectiveness. Precision’s Layer 7 got a 510(k) in April 2025 for up to 30 days of implantation, which is real and useful. But that clears the electrode array, not the fully implantable wireless system still in development. Synchron says its path to a PMA runs through a 2026 confirmatory trial, funded by a $200 million Series D announced November 6, 2025. I found no confirmation that enrollment has begun, so that remains a plan, and the timeline depends on what the trial shows.
If you build in this space, learn to say which rung your device is on before anyone asks. And if you read BCI news weekly, which of these seven terms still trips you up?