Invasive vs Non-Invasive BCI: What Each Can and Cannot Do

By neuraspeak editorial · Updated 2026-10-07 · 3 min read

Invasive brain-computer interfaces place electrodes inside the skull, either on or in the brain or in a blood vessel next to it. So far they are the only BCIs that have restored fast, large-vocabulary communication to people with paralysis. Non-invasive BCIs, such as scalp EEG, MEG and fMRI, need no surgery, but they record weaker and blurrier signals. They support slower, selection-based communication and research-grade decoding, not conversational speech.

What is the difference between invasive and non-invasive BCIs?

The difference is where the sensors sit relative to the neurons they measure, and invasiveness is a spectrum rather than a yes-or-no category. At one end, scalp EEG caps measure electrical activity through the skin and skull. Magnetoencephalography (MEG) measures the magnetic fields that neural activity produces, and functional MRI (fMRI) tracks changes in blood flow.

Inside the skull, there are three main designs. Endovascular devices such as Synchron's Stentrode are threaded through a neck vein into a blood vessel next to the motor cortex. In its first-in-human safety study, a four-patient study published in JAMA Neurology in 2023, the Stentrode had 16 electrodes. Surface electrocorticography (ECoG) arrays rest on the brain without penetrating it. Precision Neuroscience's Layer 7 film carries 1,024 electrodes, and the UCSF speech studies used a 253-electrode grid. Penetrating arrays go into the cortex itself. These include the Utah-style arrays used by the BrainGate consortium and Neuralink's N1, which spreads 1,024 electrodes across 64 flexible threads.

Why do invasive BCIs decode speech so much better?

Invasive BCIs decode speech better because they record from close to the neurons, before bone and tissue blur the signal. A penetrating electrode can detect spikes from single neurons or small groups of neurons. A scalp electrode records the summed activity of a very large population, smeared by the skull. Speaking also makes facial and jaw muscles fire, which creates electrical artifacts that can swamp EEG, so non-invasive systems have trouble separating attempted speech from muscle noise.

The published results reflect this gap. Intracortical and ECoG systems reported in Nature in 2023 decoded attempted speech at 62 and 78 words per minute. In 2024, a UC Davis system reported in the New England Journal of Medicine kept 97.5% word accuracy on a 125,000-word vocabulary over 8.4 months. No non-invasive system has published comparable real-time results for open-vocabulary speech.

What can non-invasive BCIs do today?

Non-invasive BCIs can decode some language-related information in controlled research settings, and EEG supports slower forms of communication. In those forms, the user picks letters or options by attending to flashing or flickering targets on a screen.

Two research results show both what is possible and where the limits are. In 2023, Tang and colleagues at the University of Texas at Austin reported in Nature Neuroscience an fMRI-based decoder. It recovered the gist of speech a person heard or imagined. Successful decoding required the person's cooperation, both to train the decoder and to use it, and fMRI requires a large, stationary scanner. In 2025, Meta researchers described Brain2Qwerty, which decoded sentences from brain activity while 35 healthy volunteers typed. With MEG, the average character error rate was 32%; with EEG, it was 67%. MEG also requires a magnetically shielded room, so neither result is a portable communication aid yet.

Consumer EEG products, such as headphones that estimate focus, measure broad brain states rather than words.

What are the risks of an invasive BCI?

The main risks of an invasive BCI come from surgery and from having hardware in the body for a long time. Opening the skull carries risks of bleeding, infection and seizures. Systems with a connector that passes through the skin leave a permanent opening. Ian Burkhart lived with a penetrating array for about seven and a half years. In 2021, an infection developed where the cable entered his scalp, and the device was removed that August, as MIT Technology Review reported.

Implants can also lose signal over time. In 2024, Neuralink reported that a number of electrode threads had retracted from its first participant's brain, which reduced the data the implant could collect. The company said it compensated with software changes.

Less invasive designs trade signal quality for lower surgical risk. Synchron's six-patient COMMAND early feasibility study met its primary endpoint: no device-related serious adverse events resulting in death or permanent increased disability during a year of follow-up. But the Stentrode's 16 electrodes have not been shown to decode speech directly.

Which approach suits which use?

For people with severe paralysis who have lost the ability to speak, implanted systems are currently the only BCIs with evidence of fast, accurate communication. Today, that access is available only through clinical trials. Non-invasive systems suit people who cannot or do not want to have surgery, short-term use, research, and consumer applications where approximate signals are good enough.

The middle of the spectrum is growing. Endovascular and surface ECoG devices aim to capture more signal than EEG with less tissue damage than penetrating arrays. Better AI models are also getting more out of noisy non-invasive recordings. Still, no non-invasive approach has yet shown open-vocabulary speech decoding at conversational speed, and as of October 2026, no permanently implanted BCI for paralysis has received FDA marketing approval.

The bottom line

Invasive BCIs require surgery but record signals clear enough to decode speech at tens of words per minute with high accuracy. Non-invasive BCIs are safer and easier to access, but in published research they decode language far more slowly and less reliably. Endovascular and surface devices are attempts to find a middle ground.

Questions

Is an endovascular BCI invasive?+

Yes, but less so than open brain surgery. The Stentrode is placed through a vein in the neck, so the skull is not opened. It is still a permanent implant inside a blood vessel next to the brain.

Can EEG headsets read words from the brain?+

Not reliably. In research, EEG-based sentence decoding during typing had an average character error rate of 67%. Consumer EEG devices estimate general states such as focus, not words.

Do invasive BCIs stop working over time?+

Signal quality can drift or decline because of tissue responses or hardware changes, so decoders often need recalibration. Some participants have used implants for years, but long-term data are still limited.

Glossary: EEG vs intracortical recording Brain-computer interface (BCI)Electrocorticography (ECoG)Utah array Endovascular BCI Neural signal drift

Sources

Facts on this page are drawn from the sources listed and were accurate to the best of our knowledge on the update date shown. Company and product names are trademarks of their owners and are used for identification only. Nothing here is medical, investment or legal advice. Companies and individuals can request a correction via the corrections page.