Brain-Machine Interfaces in 2026

This is one of the most thrilling moments in the history of neuroscience, and one of the easiest to misread. Read a headline this week and you will think paralyzed patients are typing with their minds at full speed, that Neuralink is about to ship a consumer product, and that we are months away from downloading skills into our brains. Read the actual peer-reviewed papers and you find something more honest, and in many ways more remarkable: a small number of patients with serious paralysis are doing things they could not do before, the engineering challenges are still enormous, and the curve is bending fast because of AI.

This guide is for the reader who keeps seeing brain-machine interface stories and wants the real picture. No hype, no doom. We will cover what a BMI actually is, who is building them in 2026, what patients are actually doing with them, where the science still gets stuck, and the role AI is now playing in turning electrical noise into language and motion. By the end you will be able to read any BMI news story with calibrated skepticism and real excitement.

What is a brain-machine interface, in plain English?

A brain-machine interface, sometimes called a brain-computer interface or BCI, is a device that records electrical activity from neurons inside or near the brain and translates that activity into a signal a computer can act on. The neurons fire. Sensors pick up the firing. Software decodes the pattern. The computer moves a cursor, types a letter, drives a robotic arm, or in the research lab, speaks a word the patient intended to say.

That is the entire idea. The difficulty hides inside every word of the sentence. Neurons are tiny. The electrical signals are weak. The brain is a salty wet environment that destroys most electronics within months. The patterns that correspond to intended movement are different in every patient and change over time even within one patient. And the computer side of the stack has to be fast and reliable enough that the patient does not get a delay between thought and action.

BMIs come in three broad flavors. Invasive systems put electrodes directly into brain tissue. They get the cleanest signal but require open-skull surgery and have the worst long-term durability. Semi-invasive systems sit on the surface of the brain or inside a blood vessel that runs through it, trading some signal quality for far less surgical risk. Non-invasive systems read through the skull with EEG or fNIRS sensors, which is safe and easy to set up but recovers a much noisier and slower signal. The whole field is a tradeoff between signal quality and surgical risk, and every major company has picked a different point on that curve.

Where are brain-machine interfaces in 2026?

The field has more momentum in 2026 than at any point in its sixty-year history. There are now eight people walking around with a Neuralink implant. Ten people in Synchron trials. Dozens of BrainGate participants with Blackrock arrays going back to 2004. A handful of newer companies are running first-in-human studies, and several research labs are publishing speech-decoder papers that would have been science fiction five years ago. The reason this is happening now is not one breakthrough, it is three converging trends: better electrodes, vastly more compute, and the same AI revolution making everything else feel different.

The players you should know about:

• Neuralink. Elon Musk’s company, founded 2016. Full-skull implant. Robotic surgical insertion of more than a thousand electrodes. First human patient, Noland Arbaugh, implanted January 2024. Eight patients in the PRIME trial as of early 2026.
• Synchron. Founded 2012, based in New York. The stentrode, an implant the size of a paperclip that is fed through the jugular vein into a blood vessel above the motor cortex. No open-skull surgery. Ten patients across the SWITCH (US) and COMMAND (Australia) trials.
• Blackrock Neurotech. The grandparent of the field. Their Utah Array has been used by BrainGate consortium researchers since 2004. Most of the published “patient with paralysis types with thought” research papers in the field used Blackrock hardware.
• Paradromics. Austin-based. Building a higher-channel-count implant designed to read from many more neurons simultaneously. First human implant in 2024.
• Precision Neuroscience. Founded by a former Neuralink co-founder, Ben Rapoport. Their Layer 7 implant sits on the surface of the cortex without penetrating it. FDA-cleared for short-term use in 2025.
• Onward Medical. Spinal stimulation rather than cortical recording. The company that helped a paralyzed patient, Gert-Jan Oskam, walk again in 2023 using a digital bridge between brain and spinal cord.

Three other companies worth watching: Kernel (consumer-grade fNIRS), Forest Neurotech (focused ultrasound BMIs that promise penetration without electrodes), and Inbrain Neuroelectronics (graphene electrodes from Spain).

What is Noland Arbaugh actually doing with a Neuralink?

Noland Arbaugh, the first Neuralink human patient, was paralyzed below the shoulders in a 2016 diving accident. He received his implant on January 28, 2024. Within weeks he was using thought alone to move a cursor, play chess online, post on X, and, in his own words, set a new world record for cursor speed in a brain-machine interface. The footage of him beating Civilization VI with his mind is the most-shared BMI clip in the field’s history.

What is also true is that within the first month after implantation, many of the electrode threads retracted from the brain tissue. Neuralink’s blog post in May 2024 was honest about this, and his usable channel count dropped sharply. The engineering team responded by re-tuning the decoding algorithms to do more with fewer channels, and his performance recovered to roughly the original speed. The story is a clean example of what BMIs in 2026 are really like: real, impressive function alongside hardware problems that the next generation of devices has to solve.

Patients two through eight have followed with the device generation that came after Arbaugh’s, with the electrode-threading procedure refined to reduce retraction risk. Several have ALS rather than spinal-cord injury. The PRIME trial is designed to learn from each one, and the implant itself is still iterating between patients in a way that would be unusual for a stabilized medical device but is appropriate for an early human study.

How does Synchron’s stentrode work without skull surgery?

This is the bit of the field most beginners find genuinely surprising. Synchron’s implant is fed up through the patient’s jugular vein and parked inside a blood vessel that runs over the motor cortex. From inside that vessel it picks up the same kind of electrical signals an implanted electrode would, just from the other side of a thin layer of tissue. The surgery looks more like a cardiac catheterization than a craniotomy. There is no open-skull procedure. Recovery is days, not weeks.

The tradeoff is signal quality. A stentrode parked in a vein records a noisier and less spatially specific signal than electrodes pushed directly into cortex. What Synchron has bet, and what its first ten patients are testing, is whether the engineering and decoding software can close the gap enough to deliver real function with a fraction of the surgical risk.

The published results so far show patients sending text messages, controlling smart-home devices, and shopping online using thought alone. The bandwidth is lower than what Neuralink’s higher-channel-count system delivers, but the gap is narrower than the surgical-risk gap. If the consumer BMI of the 2040s ends up being a vessel-routed implant rather than a skull-cracked one, Synchron is the company that proved it could be done.

Who else is in the race, and why do they matter?

The field is much wider than Neuralink and Synchron, even though those two get most of the press. The other names worth knowing:

Blackrock Neurotech. If you have read a “patient uses thought to type” research paper from a major university in the last twenty years, the implant was almost certainly a Blackrock Utah Array. The BrainGate consortium, which includes Brown, Stanford, Harvard, Case Western and a half-dozen other institutions, has run continuous research with Blackrock hardware since 2004. Blackrock is the closest thing the field has to a quiet incumbent, and most of the foundational science that newer companies are building on came out of papers using their arrays.

Paradromics. Founded by Matt Angle, an alumnus of the Howard Hughes Medical Institute and Janelia Research Campus. The bet is higher channel count: their device aims to record from many more neurons in parallel than current systems. Their first human implant happened in 2024 at the University of Michigan. Their argument is that the bandwidth needed for actually-useful BMI applications, like restoring fluent speech or fine motor control, will require ten to a hundred times more channels than the field has today.

Precision Neuroscience. Co-founded by Ben Rapoport, who left Neuralink in 2018. Their Layer 7 implant is a thin film that drapes over the cortex without piercing it. The bet is that surface recording is good enough for many applications and dramatically safer than penetrating electrodes. They received FDA clearance in 2025 for short-term use during brain surgery, and human studies for longer-term use are progressing.

Onward Medical. A Swiss company doing something different. Onward does not record from the brain at all. Instead it pairs a small surface array on the motor cortex with an implanted spinal stimulator and uses the brain reading to drive coordinated spinal stimulation. The result is the digital bridge that helped Gert-Jan Oskam, a man paralyzed for over a decade, walk again. That work was published in Nature in 2023 and is probably the most emotionally powerful BMI footage of the decade.

Kernel, Forest Neurotech, Inbrain. These are the wild cards. Kernel builds a wearable fNIRS helmet that reads blood-oxygen changes through the skull. Forest is exploring focused-ultrasound stimulation. Inbrain is building graphene-based electrodes from a lab in Barcelona. Each represents a path the field might take if the dominant cortical-electrode approach hits a ceiling.

What can brain-machine interfaces actually do today versus the hype?

This is where most beginners get misled, and it is worth being precise.

What patients with implants reliably do in 2026: move a cursor on a screen by imagining hand movement, click on targets at speeds comparable to slow typing on a phone, control prosthetic limbs to grasp objects, send text messages, play games, browse the web, and in two well-known cases at UC San Francisco and Stanford, generate speech at 62 and 78 words per minute respectively from intended speech. Normal speech is about 160 words per minute, so this is real but not yet fluent.

What no implanted patient does in 2026: read or write at full conversational speed, transmit complex thoughts directly between brains, control devices without focused effort, or have any function that resembles the science-fiction version. The speech-decoder papers from Willett et al at Stanford and Metzger et al at UC San Francisco, both published in Nature in August 2023, remain the field’s high-water mark for bandwidth and are still in research-lab settings rather than at-home daily use.

What research-only systems with no implant have done: something almost stranger. A paper from Alexander Huth’s lab at UT Austin, published in Nature Neuroscience in 2023, showed an fMRI-based semantic decoder that could reconstruct the rough meaning of stories a person was listening to with about 74 percent accuracy. No implant. No surgery. But the accuracy is gist-level rather than word-perfect, the equipment is room-sized, and the system has to be trained on each individual subject for hours before it works. It is the most impressive non-invasive BMI result published, and also a useful reminder that science fiction is still science fiction.

What is still hard, and why?

The four chronic engineering problems the field has not solved:

• Implant longevity. Brain tissue is hostile to implanted electronics. Scar tissue forms around electrodes within weeks. Signals degrade over months. Most implants in research settings work for two to five years before signal quality drops to unusable. Solving this is probably worth a Nobel Prize.
• Surgery risk. Open-skull procedures carry real complication rates: infection, bleeding, stroke. Even the most refined neurosurgical practice puts a ceiling on how many people will accept a BMI for non-life-threatening reasons. This is why Synchron’s vessel-routed and Precision’s surface implants matter so much: they slide the cost-benefit curve.
• Signal drift. Neurons move around. The cells that fired for “imagine moving my hand right” on day one may not be the same cells doing that on day ninety. Decoders have to either retrain continuously or use AI techniques that adapt to drift. This is one of the parts of the stack where modern machine learning is making the biggest contribution.
• Decoding language. Turning electrical patterns into words is harder than turning them into cursor movement. The motor cortex has well-mapped regions that fire predictably for specific intended movements. Speech and language live in more distributed and individual-variable circuits. The Willett and Metzger results are evidence that it is possible, but generalizing across patients without re-training is a major open problem.

How does AI fit into all of this?

This is where the field is moving fastest. A BMI is, fundamentally, a signal processing problem: noisy electrical patterns at one end, intended action at the other end, and a decoder in the middle. For decades the decoders were classical machine learning models, often hand-crafted feature extractors paired with relatively simple classifiers. They worked, but each new patient required hours of calibration and the systems were brittle when neurons drifted.

The neural networks that drove the deep-learning revolution turned out to be useful for the literal neurons in the brain too. Modern BMI decoders are mostly recurrent and transformer-based architectures that learn the mapping from electrode signals to intended action with the same kind of flexibility that a large language model uses to predict the next token. The Willett 62-WPM paper used a recurrent neural network. The Metzger paper used a deep neural language model on top of the neural decoder, effectively adding the same kind of language prior that helps you understand a friend speaking through a bad phone connection.

The bigger near-term effect is calibration. Foundation-model-style approaches let a decoder pre-trained on data from many patients adapt to a new patient in minutes instead of hours. That is the difference between a research curiosity and a clinically useful product, and it is happening in published work right now. Expect every major BMI company in 2026 and beyond to have a serious AI team, and to use the same general techniques the rest of the AI world is using.

When does a consumer brain-machine interface arrive?

Honest answer: not in the 2020s, and probably not in the early 2030s either.

The reason has nothing to do with software and everything to do with the implant durability problem above. A consumer product needs to work reliably for ten years without intervention. Current research devices struggle past five. A consumer product also needs a surgical risk profile that healthy people will accept, which probably rules out open-skull approaches forever and puts a lot of pressure on vessel-routed, surface, and ultimately on whatever non-invasive technique might eventually deliver useful bandwidth.

The realistic medical timeline runs ahead of the consumer one. Expect FDA-approved BMI products for paralysis, ALS, locked-in syndrome, and certain forms of treatment-resistant depression in the late 2020s and 2030s. Restoring function to people who have lost it has a different risk-benefit math than enhancing function in people who do not need it.

Beyond that, the honest answer is that nobody knows. If someone solves the durability problem with a new electrode material, or if focused-ultrasound systems mature enough to give wearable bandwidth, the field could move much faster than the current curve suggests. If those problems remain stubborn, BMIs may stay a powerful medical tool for the people who need them most and never become the consumer device the press writes about. Both outcomes are within the range of what we honestly know in 2026.

Is this safe?

For the medical patients involved today, yes within the bounds you would expect for any neurosurgery. The first generations of these implants are being tested in patients who have severe paralysis or terminal neurodegenerative disease, where the benefit clearly outweighs the surgical risk. FDA approval requires demonstrating that.

The harder questions arrive later. What does long-term electrode-tissue interaction look like over decades? What happens when an implanted patient’s company is acquired, goes bankrupt, or simply stops supporting the device? Who owns the data a BMI generates? Could a hostile actor remotely access the device? These are not science-fiction questions, they are policy and engineering questions the field needs to answer well before consumer BMIs are anywhere close to real. Some of the most thoughtful work is being done at the Neuroethics Society and at the IEEE Brain Initiative.

Frequently asked questions

Is a brain-machine interface the same as a brain-computer interface?

Yes. The terms are used interchangeably. BMI is more common in motor-prosthetics literature, BCI is more common in cognitive-neuroscience literature, but both refer to the same broad category of device.

Can a BMI read my thoughts?

Not in the way the question usually means. Current BMIs read patterns that correspond to specific intended actions, like imagining hand movement or trying to speak. They do not pick up your private inner monologue while you are thinking about lunch. The semantic-decoder fMRI work at UT Austin is the closest research-lab demonstration of something more general, and it is gist-level on consenting participants who spent hours training the system to their individual brain.

Will I be able to type with my mind?

For a healthy person in the 2020s, almost certainly no. For a person with paralysis whose typing is currently slow or impossible, yes, and they are doing it already. The 62 to 78 WPM speech-decoder rates from the Stanford and UCSF labs are real, and they are higher than what most people type on a phone.

Are BMIs dangerous?

Implantation is neurosurgery and carries the risks of neurosurgery. The implants themselves, once in place, have not shown dangerous side effects in published trials, but the long-term track record is short. The patients who have BMIs in 2026 accepted real surgical risk because the alternative was severe paralysis or terminal illness.

How does AI make BMIs better?

Mostly by improving the decoder, the software that turns raw electrode signals into intended action. Modern AI techniques handle signal drift better than older approaches, transfer learning between patients better, and use language-model-style priors to clean up noisy speech decoding. AI is also helping the surgical-planning and electrode-placement side, though that is earlier-stage.

Where should I go to keep up with this field?

Primary sources: the Neuralink, Synchron, Blackrock, Paradromics, and Precision Neuroscience official blogs. Research literature: the Nature family of journals and the Journal of Neural Engineering. For thoughtful long-form coverage, MIT Technology Review and STAT News both do credible BMI reporting.

The Beginners in AI position on brain-machine interfaces

This is genuinely one of the most exciting frontiers in technology, and it is also a place where the temptation to write hype is enormous. We are pro-technology and we are excited by what these devices already do. A man who was paralyzed for years is playing chess with his mind. A man who could not speak for a decade is generating words at near-conversational speed. The video footage from the Onward digital bridge of Gert-Jan Oskam walking is one of the things that should make any thinking person hopeful about the medical century we are entering.

We are also pro-human first, and that means the same thing here it means everywhere on this site. The brains being interfaced with are people, with families and identities and a future they deserve to control. The right question to ask of every BMI development is whether it expands the patient’s agency or constrains it. Restoring speech to someone who lost it: agency. Letting a patient browse the web with their mind: agency. The thing nobody should want, and nobody respectable in the field is asking for, is a future where these devices become a way to harvest attention or hijack thinking. The medical use case is the right one to lead with, and probably for a long time.

If you are a curious beginner reading this on a Wednesday morning, the takeaway is: read the primary sources, watch the patient interviews, hold the excitement and the skepticism at the same time, and remember that the most important thing about a brain-machine interface is the brain on the other side of it.

Sources

• Willett FR et al. “A high-performance speech neuroprosthesis.” Nature, August 2023. nature.com/articles/s41586-023-06443-4
• Metzger SL et al. “A high-performance neuroprosthesis for speech decoding and avatar control.” Nature, August 2023. nature.com/articles/s41586-023-06443-3
• Tang J, LeBel A, Jain S, Huth AG. “Semantic reconstruction of continuous language from non-invasive brain recordings.” Nature Neuroscience, 2023. nature.com/articles/s41593-023-01304-9
• Lorach H et al. “Walking naturally after spinal cord injury using a brain-spine interface.” Nature, May 2023. nature.com/articles/s41586-023-06094-5
• Neuralink PRIME Study Progress Update, May 2024. neuralink.com/blog/prime-study-progress-update
• Synchron official site and SWITCH/COMMAND trial updates. synchron.com
• Blackrock Neurotech and BrainGate consortium. blackrockneurotech.com
• Paradromics official site. paradromics.com
• Precision Neuroscience FDA clearance and Layer 7 overview. precisionneuro.io
• Onward Medical brain-spine digital bridge. onwd.com

You might also like

• The Brain and Neuroscience Tech Frontier, our master hub for everything brain-related
• Why writing by hand still beats typing for learning, the 2024 Norwegian EEG study, in plain English
• The Biotech Frontier, what CRISPR, synthetic embryos, and de-extinction mean for the next decade
• The Robotics and Drones Frontier, with Figure, Tesla Optimus, and the humanoid race
• The AI Glossary, plain-English definitions for every AI term we use