A Scientist and Engineer Explain Everything Elon Musk’s Neuralink Can (and Can’t) Do
Co-authored by Chris Chiang
In late August, more than 150,000 people tuned into a webcast to watch a live demonstration of the latest tech from brain-computer interface company Neuralink. The secretive startup, founded by entrepreneur Elon Musk, plans to use a tiny brain implant to merge humans with A.I.
Viewers were treated to one of the most fascinating and bizarre tech demos in recent memory, complete with neurosurgery robots, a live feed of spiking neurons in a living brain, and the delightful spectacle of a billionaire in a bespoke sport coat speaking furtively to a Tamworth pig while attempting to coax her onstage (“Come on, Gertrude, we have snacks!”).
Musk is a gifted showman, known for stunts like launching his personal Tesla into space and selling functioning flamethrowers to promote a tunnel-building enterprise. His theatrics, though, can make it hard to tell when he’s discussing a real technology that’s here today or an aspirational vision that’s still decades off. Neuralink is also an extremely complex company. It merges medical devices, brain surgery, robotics, neuroscience, and machine learning. That makes understanding its tech even more of a challenge.
I have a degree in cognitive science (neuroscience and linguistics) from Johns Hopkins University and a decade of experience in A.I. My friend and co-author Chris Chiang has a degree in biomedical engineering (also from Hopkins) and has spent the last eight years in the medical device field. Before Musk had even finished speaking, we connected and started delving into what Neuralink’s demo actually shows. What parts of its tech are medically useful? What changes and sacrifices has the company made since it last shared details of its device in 2019? When will we get functioning telepathy? And what does this all mean for the future of computer-connected brains?
His theatrics can make it hard to tell when he’s discussing a real technology that’s here today or an aspirational vision that’s still decades off.
The core of Neuralink’s tech is the Link, a coin-sized implant that Musk describes in the demo as “a Fitbit for your brain, with wires.” The Link charges through electromagnetic induction like a phone you charge wirelessly and is about the thickness of the human skull (8 mm). Neuralink would install the Link with a custom-built surgical robot, which operates a bit like a sewing machine. The robot would drill a hole in a person’s skull, and the Link would be inserted, filling the hole like a wine cork in a bottle. The skin of the person’s scalp would be closed over the device, making it invisible. As Musk quips in the demo, he could have a Link installed already, and no one would know.
On its underside, the Link sports more than 1,000 tiny electrodes organized into “threads,” each about 5 microns wide — 1/20 the thickness of a human hair. The robot would plunge these threads into the cortical surface of the brain, allowing the device to both read electrical impulses from the cortex and “write” signals to the brain. Musk describes the installation process, which is shown briefly in a video during the demo, as “not too gruesome.” The robot rapidly inserts electrode after electrode into a living brain, dodging blood vessels to minimize bleeding.
Entrusting a robot to repeatedly stab your brain with a sewing needle might seem creepy, but Neuralink needs a robot surgeon because of the tiny size of the Link’s electrode threads, the massive number of minuscule blood vessels that must be avoided, and the large number of electrodes that need to be inserted. Neuralink’s robot has steadier “hands” than even the best human surgeons. During surgery, skilled micro-surgeons’ hands move about 50–100 microns side to side. When you’re installing electrodes that are five microns wide, even those tiny movements are still far too much.
The robot is also necessary in order to minimize brain damage during implantation. Hitting blood vessels while installing the Link would not only cause nearby neurons to wither but also reduce the device’s signal quality over time, and perhaps even significantly compromise the blood-brain barrier. Avoiding this is especially important since disruption of the blood-brain barrier underlies many neurodegenerative conditions, such as Alzheimer’s, Parkinson’s, and multiple sclerosis.
The Brain That Remade Itself
Doctors removed one-sixth of this boy’s brain — and what was left did something incredible
The installation surgery, Neuralink claims, would take about an hour and could be performed on an outpatient basis without general anesthesia. The company also claims that the procedure is fully reversible (users would have a hole in their skull, though, if they had the device taken out). Musk hopes users would upgrade their Links every few years, just as we upgrade our iPhones today.
Already, this raises some compelling possibilities — as well as some seriously concerning risks and challenges. Compared to existing neural implants (yes, these do already exist), the Link is an impressive piece of kit. The Utah Array — a competing system used for neural implants (which Musk mentions in the demo) — is generally connected to a bulky box on the skull and requires extensive external hardware and cables. It also has to be installed by a skilled human surgeon (the surgical manual for the device runs to 41 pages).
Neural implants like the Utah Array also have major risks, including at least a 1% risk of stroke during the installation surgery, and up to a 5% risk of infections after installation, which are usually severe enough to require temporary removal of at least part of the implant. Because the Link is completely embedded under the skin and transmits data wirelessly, there are fewer ways that bacteria can use the device as a conduit to slowly encroach upon the brain.
Electrodes in existing devices like the Utah Array are also rigid and bulky, and they must usually be applied in a specific, small area of the brain, with uniform patterns of coverage. The Link, in contrast, can use its threads to spread electrodes through multiple adjacent regions, potentially accessing more of the brain at once. The Utah Array also has far fewer electrodes — generally 256, versus 1,000+ in the Link. Like the number of pixels in a camera’s sensor, more electrodes mean a higher resolution of data and potentially more insight into the brain’s activities. Notably, the Link also offers much better resolution than noninvasive brain-reading tech like scalp EEGs, which sit atop the head rather than in the brain itself.
Neuralink’s robot has steadier “hands” than even the best human surgeons.
Neuralink’s ambitions for the initial version of the Link, though, have changed dramatically since the company’s 2019 demo. Originally, Musk had planned for the Link’s electrodes to extend deep into the brain, giving it access to regions like the limbic system (which regulates emotion and plays a role in addiction and anxiety) and its sub-component, the hippocampus (which mediates memory formation, among other functions). Now, as the demo reveals, the team is limiting itself to interfacing with the cortical surface, a much more accessible region on the outer surface of the brain.
That means many of the treatments and capabilities that Musk promises in the demo are still years off. Without access to deep brain regions, Neuralink would have a harder time addressing memory loss, depression, anxiety, insomnia, addiction, or many strokes, much less the “downloading” of new skills, which Musk also promises. Matrix-style downloading of skills into the brain, for example, would likely require accessing the basal ganglia, which is involved in forming procedural memories. That structure is buried deep within the brain — far below the cortical surface, and likely out of reach of the current generation Link.
Still, there’s an awful lot one can do on the cortical surface. The primary somatosensory cortex, for example, is located on the parietal lobe of the brain, and would likely be accessible to Neuralink’s device. It handles the “inputs” that allow for physical sensation, including proprioception (the sense of body position in space), pain, touch, and temperature. Likewise, the primary motor cortex would also likely be accessible. It controls the “outputs” required to move the limbs, trunk, eyes, and other body parts. Both regions are physically mapped to areas of the body, so knowing where to place electrodes would be relatively straightforward.
Connecting the Link to these two regions would provide some powerful capabilities. For one, Neuralink could potentially use the device to create a brain-connected artificial limb for amputees. Sensors on a prosthetic limb could be connected to the Link wirelessly to send their signals into the primary somatosensory cortex, giving the artificial limb realistic feeling and sensation. The Link could then read from the primary motor cortex and translate the region’s signals into instructions for moving the prosthetic arm.
How to Give A.I. a Pinch of Consciousness
A.I. researchers are turning to neuroscience to build smarter, more powerful neural networks
The end result would be an artificial limb a patient could control using their brain, and that would restore a sense of touch, which is crucial for fine motor skills. Indeed, this has already been demonstrated using other brain implants. A section of Musk’s demo hints at this direction. In the demo, a Link-implanted pig walks on a treadmill, and a graph on a screen shows the predicted position of her legs based on signals read from her brain — which line up well (though not perfectly) with the way her legs are actually positioned. The Link’s small size, much higher resolution, and wireless capabilities would likely result in much better fine motor control for patients, more realistic sensation, and fewer external boxes and wires to wear than with existing devices.
A similar technique could be used to bypass damaged nerves in the spine, restoring feeling and movement to people with paraplegia. Neuralink has already hinted that treating paralysis conditions would be its first human use of the technology. It’s likely this possibility, as well as the device’s potential impact, that led the U.S. Food and Drug Administration (FDA) to grant the Link a “Breakthrough Device” designation, which gives Neuralink priority in FDA reviews, as well as additional ways of interacting with the FDA to receive feedback more rapidly.
The Link could also likely interface with other brain areas on the cortical surface, such as the primary auditory cortex and primary visual cortex. These regions contain physical maps of the auditory spectrum and visual field, respectively. They’re more complex than sensory and motor regions of the brain, so it’s unlikely Neuralink will be superimposing images onto our visual field or playing music into our heads anytime soon. But even the current version of the Link could likely stimulate these areas in a blunt way, creating a blinking dot, audible tone, or flash of color that only you could perceive. This could be used to notify you of some event — like an incoming text on your phone, or an upcoming obstacle on the road in front of your Tesla.
With a slightly deeper reach into the brain, the Link could even potentially stimulate the primary gustatory cortex, the region that receives input related to taste. As a user ate broccoli or another healthy, relatively bland food, the Link could stimulate sweetness neurons in their gustatory cortex, giving the impression that they were enjoying a delicious dessert. This seems silly at first, but augmenting taste is already a proven therapy for cancer patients, who often struggle to eat enough during chemotherapy.
To go beyond interacting solely with movement and sensation — and to begin to perform functions like altering memories or controlling mood — the Neuralink team would need to do more than bluntly prod specific brain areas and read basic signals. They would need to develop a deep understanding of how the brain operates so that the Link could interface with it properly.
Without access to deep brain regions, Neuralink would have a harder time addressing memory loss, depression, anxiety, insomnia, addiction, or many strokes, much less the “downloading” of new skills, which Musk also promises.
Neuralink would likely achieve this by using machine learning to associate patterns of brain activity with things happening in the real world. For example, Neuralink’s software could learn that a particular firing pattern of neurons corresponds to a user hearing the sound of a specific frequency. This might allow the Link to turn a user’s ears and primary auditory cortex into the equivalent of a hidden microphone, recording everything the user heard by interpreting the frequencies of sounds reaching the brain.
By accessing brain regions related to speech planning (such as Broca’s area), the device could also one day theoretically eavesdrop on words a user is thinking but not actually saying. These could be sent to another person’s phone — or even directly into their brain if they also had a Link. This would allow Neuralink to deliver the telepathy features that Musk promises.
Today’s machine learning, though, is not omnipotent. It generally depends on access to copious amounts of input data, which is then correlated with specific outputs. But for many medical conditions, it’s unclear what this input data and output would look like. What signal, for example, corresponds to a user feeling depressed? Or, what signal might cancel out a drug craving?
For complex conditions and mental states like these, it’s hard to imagine how Neuralink would even identify the right input signals to look for, much less train a computer to generate effective output signals. Machine learning primarily focuses on finding (and copying) patterns. But something as complex as mental illness is unlikely to have obvious signal patterns for a machine learning system to copy. Trying to treat mental illness without a clear sense of how it manifests in the brain could have unintended consequences, or might risk making conditions worse.
When you combine technologies like robotics, brain augmentation, machine learning, and neurosurgery, a whole range of other things can go wrong, too. Operating on the brain is always risky, even with a high-tech robot. And even if Neuralink can insert Links successfully, that’s no guarantee that the device will last long-term.
One key concern for the longevity of an implanted device is how hermetic its packaging is — how well it keeps water out. Your body is 60% water and the instant an implantable device is installed, water starts to slowly work its way in. Once enough molecules of water infiltrate a device, condensation can happen and the internal electronics can corrode.
Making the Link even slightly bigger would help it to last longer in the body. Like a sinking boat, the larger the device, the less a small leak would matter. Increasing the diameter of the Link by as little as 1 millimeter might buy users an additional four to six years of useful life, without the need to drill a substantially larger hole in the skull. Device longevity will likely also be tied to the durability and coatings of its electrodes, and whether users’ bodies mount a foreign body response and reject it.
Beyond physical and medical risks, the data privacy and security risks of connecting a computer directly to your brain are substantial and terrifying. In many ways, these risks mirror the risks we face with existing technologies — pervasive surveillance, influence through advertising, risks of hacking, and blackmail — but amplified (to paraphrase Musk) by orders of magnitude.
If an attacker could gain read access to a future Link device with electrodes in regions like Broca’s area, malicious mind-reading and secret-stealing could be possible. The attacker could potentially “hear” a user’s thoughts as they were thinking them, even if they didn’t express them out loud. This could lead to information asymmetry in negotiations (the attacker knowing more than another party), enable the device to act as a hidden polygraph, or let the attacker conjure someone’s darkest secrets for blackmail.
If the attacker gained write access to the brain via a compromised Link, the possibilities are even scarier.
The attacker could produce phantom sensations, unintended movements, or cognitive effects at the worst of times, like when you’re driving a car. A few simple impulses to your primary motor cortex could cause you to suddenly swerve left while on the highway — likely with deadly consequences. An attacker could also potentially cause a user to throw themselves into traffic — or merely headfirst at the sidewalk.
Can Our Brains Be Used Against Us in a Court of Law?
Neuroscience in the courtroom could jeopardize our privacy and civil rights
Interestingly, though, this is likely one of the least concerning security risks of the Link. Existing implantable devices (like pacemakers and insulin pumps) are historically notorious for lax security. Yet there have been no notable cases of hackers using these devices to harm users. The reason likely comes down to economics. Malware today is primarily used for monetary gain. It’s much more profitable for hackers to ransom individuals or even large institutions for millions of dollars than to conduct secret assassinations. Even if a hacker gained write access to a device like a Link, it’s much more likely that they’d use it to swindle you, exploit you, or try to sell you things than to crash your car.
An attacker might use a future version of the device, for example, to induce emotions, will, or “gut feelings” during critical times when an individual is making important decisions. For example, a con artist could offer you the “deal of a lifetime” while flooding your brain with oxytocin, giving you an unexplainable sense of trust and goodwill, and making you more likely to accept their con.
Such attacks would essentially erase the concept of consent. By manipulating a user’s feelings and basic emotions — likely without the user even knowing this was happening — an attacker could convince them to do nearly anything, from agreeing to a bad deal, to accepting sexual advances, or engaging in other risky behaviors they would normally decline.
The Link could also be a boon for advertisers. The neural equivalent of adware is unlikely to manifest as blatant, obnoxious pop-ups in your visual field. Instead, advertisers could use the Link to stimulate dopamine-releasing areas of your brain’s ventral tegmental area when you interact with certain brands, giving you artificial positive feelings toward their products.
Advertisers could also take a more subtle approach, stimulating your brain to naturally develop a positive association with a certain smell, and then releasing that smell in a brick and mortar store. Advertisers already use scent marketing to influence your purchases — adding in a Pavlovian, neural component would only make these techniques stronger. Marketers could use the device to gain all kinds of insights into your preferences and desires — again, extending a process that’s already well underway.
An attacker with time on their hands could also implement a long-term assault on your brain, leveraging neural plasticity. The brain is not a static organ — its neurons are constantly modifying their connections and patterns of firing based on your environment. An attacker could, for example, use the Link to slowly rotate your vision by a few degrees a year. Experiments have shown that the brain adapts to these kinds of changes — if the attacker moved slowly enough, you’d have no problem functioning in daily life, and likely wouldn’t notice the change was taking place.
Yet if the attacker undid the cumulative rotation suddenly (if you failed to make a ransom payment, for example), this would leave your vision distorted and dysfunctional until your brain readapted to normal inputs. Removing the Link wouldn’t help, since your brain would already have adjusted to the malware’s distorted input at a neural level. Most people would probably pay a few hundred dollars in ransom to avoid spending weeks or months seeing the world distorted or upside down.
An attacker could also make it harder to detect changes in the body, such as shifts in hormone release or the homeostasis mechanisms that control energy metabolism. If you failed to pay the attacker’s ransom — or removed your Link in an attempt to escape them — these changes could leave you extremely tired, unnaturally hungry, or with a condition similar to diabetes until your body readjusted.
Neuralink could use many tactics to protect its users from these threats. Just as neural threats mirror existing security threats and privacy issues, the steps needed to protect users from neural hacking mirror traditional cybersecurity methods. Obvious steps might involve using robust authentication methods, implementing safe hardware limits on the voltage and currents the device uses, and limiting the regions of the brain where the electrodes are placed.
Additionally, Neuralink could limit physical access to the device (intentionally building very short range or directional communication radios), placing physical “on-off” switches on the device, or locking the range of device output patterns to a limited set of safe “on-board functions” (mirroring a technique used to protect today’s SQL databases). Machine learning could protect users’ ability to give consent by detecting worrying patterns of neural signals or behaviors, much like the techniques used to detect fraudulent bank transactions today.
To begin implanting the Link in real human brains, Neuralink must meet a variety of requirements from the FDA (short-term biological safety likely being the primary concern), allowing it to move ahead with human trials. This could happen within a year. Later, the company would need to address these security and longevity concerns in order to get the FDA’s blessing for premarket approval, which is necessary to sell the device commercially.
Musk is clear from the start of the demo, though, that his presentation is not meant to sell Links. Instead, it’s primarily intended to recruit the programmers, engineers, scientists, marketers, lawyers, and animal handlers needed to continue developing the Link and move into human trials (full disclosure: Chiang has applied to positions at Neuralink in the past, and may do so again in the future). Cyborg pigs are an interesting demo of early brain-reading techniques. But they’re an even better way to get tech employees (and investors) excited about your company.
Musk’s theatrics can be off-putting to those who are used to the slow, calculated march of traditional science. MIT’s Technology Review, for example, calls Musk’s demo “neuroscience theater” and says that Neuralink “made promises that will be hard to keep.”
Such issues haven’t stopped Musk in the past. As Technology Review points out, one cannot yet actually buy a Link. But the same was largely true of the Tesla Model 3 when Musk announced it in 2016 and started taking $1,000 deposits for the car. In the first week, Tesla received over $325,000,000 for a product it was woefully unprepared to actually build and deliver, surprising even Musk.
The subsequent roll-out of the Model 3 required building manufacturing tents in parking lots, asking customers to tolerate major delays, and creating a work environment that was reportedly “hellish” enough to drive many employees and managers away from the company.
But despite the fact that the Model 3 didn’t technically exist when Tesla launched it, the company now ships up to 50,000 of the cars per month, and the Model 3 has helped Tesla become the most valuable car company in the world. The lack of an actual product to sell won’t necessarily stop Neuralink from following a similar pattern of growth. If the FDA weren’t around, Musk would almost certainly be taking preorders for the Link already.
The human brain is far more complex than any car. But even a Link device that promised incremental improvements over existing neural implants for treating serious conditions like paralysis would be a formidable force with Musk’s promotional muscle (and cash) behind it — provided the company can address the thorny privacy and security challenges of brain-computer interfaces. And if the device proved beneficial and safe in treating other, more pervasive conditions (like anxiety, depression, or addiction), the market for the Link could be vast.
More than 150,000 people currently have neural implants installed, so the idea of the Link finding a large audience of users who aren’t pigs isn’t outlandish. Telepathy is likely years off. Hopefully, the same goes for neural adware and ransomware. But a day when brain implants like the Link help amputees walk again — or let you summon your Tesla with your mind — may be here sooner than you’d think.