Skip to main content

Elon Musk's dreams of merging AI and brains are likely to remain just that — for at least a decade

Elon Musk's dreams of merging AI and brains are likely to remain just that — for at least a decade


It will probably work in the end, just not that soon

Share this story

Asa Mathat for Recode

Elon Musk wants to fix disability, replace language as we know it, and use brain implants to usher us into a telepathic world. And he wants the first part of this to be done in the next four years. All of this could theoretically work, experts say. But probably not on the timeline that Musk has set.

Yesterday, the SpaceX and Tesla CEO gave more details about NeuraLink Corp, his venture to merge the brain with artificial intelligence, in a Wait But Why explainer. In four years, Musk hopes to have a brain-machine interface that helps people with disabilities, like strokes or brain lesions, regain bodily function. In 8-to-10 years, he wants people without disabilities to be able to use it too. The possibilities he raises are exciting: play piano with your thoughts, stimulate the brain to help with mood disorders, or get the enjoyment of eating an entire tube of Pringles without actually wrecking your body with sodium and fat.

This could be incredibly exciting for therapeutics

One of the ultimate goals is to communicate mind-to-mind with none of this pesky talking, because talking leads to miscommunication. Right now, we have a thought and say it in a language we know. Someone else hears the language, and then understands it. With the brain-machine interface, says Musk, you could just send thoughts directly. “This actually is feasible because we know that when you think about words, even without moving your mouth, there’s brain activity in patterns that change and that we can tap into,” says Chad Bouton, vice president of advanced engineering and technology at the Feinstein Institute of Medical Research.

University of Chicago neuroscientist Sliman Bensmaia says he’s excited that Musk is working on brain-machine interfaces. But that doesn’t mean that all this will happen when Musk suggests, which isn’t surprising given Musk’s history of missing deadlines. Bensmaia says the four year timeline seems “basically impossible.” Decades would be reasonable, and ten years would be very, very ambitious. “When you’re talking about things like curing stroke or at least resolving the symptoms of lesions, you’re giving people hope that this will happen and now the countdown has begun,” he says. “They’ll be like, ‘in four years our problem will be resolved,’ and chances are that won’t be the case and there’ll be a lot of disappointed people.”

Even if the tech itself is figured out in four years, there will be human testing and regulatory hurdles to pass. (The good news, adds Bensmaia, is that if you can make these interfaces work for disability, then you can almost instantaneously use it to do other cool things.)

It’s nice that Musk is getting involved, but brilliant minds have already been working on brain-machine interfaces, and millions of dollars have already been spent. The first sensor was implanted in the brain of a paralyzed patient in 1998. Since then, about a dozen people have received similar implants. Groups at the University of Pittsburgh and Brown University are working on brain implants for medical purposes. A government-funded DARPA project called REPAIR also studied this exact issue. Facebook just announced that it is working on a way to let you type with your brain — or rather, type more quickly, since Stanford University researchers already created a program that lets paralyzed patients type eight words a minute with their brain. Facebook wants to bring that number up to 100. There’s also Kernel, another brain-interface startup that focuses on medical applications.

“Non-invasive” brain interfaces are impossible

And though it doesn’t hurt that Musk is very wealthy, it’s not like the rest have been begging for money. Bensmaia estimates that $100 million to $200 million of funding has been poured into revolutionizing prosthetics in the past 15 years.

Despite all that money and brilliance, we’re not particularly close to the future Musk is talking about. Here are some of the challenges.

We don’t know enough about the brain.

Your brain is composed of a lot of individual cells, called neurons, which have to communicate with each other. So they send signals through chemical or electrical impulses. A brain implant consists of an electrode or a computer chip that senses some of these signals in order to communicate with a computer. (Some implants, like Medtronic’s device for deep brain stimulation, send their own electrical impulses to help regulate the brain, but that’s not quite the same thing; it doesn’t receive information, only sends shocks.)

Even if we developed the perfect interface that anyone could use, it wouldn’t be very useful yet because we don’t understand enough about how the brain works, Bouton says. There’s been a lot of progress over the past century, but “as a neuroscientist, I can say with confidence that there’s not a lot of brain that we understand at that level of detail,” says Bensmaia.

NeuraLink scientist Flip Sabes agrees that we don’t know enough about the brain. “If it were a prerequisite to understand the brain in order to interact with the brain in a substantive way, we’d have trouble,” he said in the Wait But Why post. “But it’s possible to decode all of those things in the brain without truly understanding the dynamics of the computation in the brain.” He thinks the team can apply machine-learning techniques to solve this problem. In this scenario, a volunteer would think a series of thoughts — for example, about their dog. There’s a pattern of neural activity in the brain that represents the dog, and it’s active when they’re thinking about it. The machine-learning algorithm would record that activity, you tell it that that pattern pertains to “dog” and it would learn to make that connection.

When it comes to language, though, it’s almost impossible to gather enough data to train the algorithm, says University of Toronto neuroscientist Blake Richards. This approach could be useful for learning to move a prosthetic arm: it’s one very specific goal, there’s only one thing you want to do with it, and the person receives immediate feedback about whether or not they’re successfully moving the arm.

But there are infinite thoughts that we could have and we would need enormous amounts of data. “Even if you took a few unlucky volunteers who spend two years of their lives thinking different thoughts for you to decode with your machine learning algorithm, there’s no guarantee that different people encode information in the same way, so you’d have to go through this entire process with every person that gets the implants,” Richards adds.

We don’t know how to record enough neurons.

The brain has roughly 80 billion neurons. Brain-machine interfaces need to record the activity of these neurons to send and receive signals, and the more the better. NeuraLink’s team is hoping for one million simultaneously recorded neurons. Keep in mind that right now, usually one electrode can record one neuron.

State-of-the-art technology currently has about 100 electrodes, according to Bensmaia. There’s a big gap between 100 and one million, or even 100,000. “If they said they were shooting for 10,000 in four to eight years, that’s impressive, but I think they have a chance of doing it,” says Lee Miller, a Northwestern University neuroscientist who works on brain interfaces for spinal cord injury. “Beyond that, not so much.”

Progress has been slow. In 2011, a Nature Neuroscience article introduced “Stevenson’s law,” which tracks how quickly we can learn to record more neurons. The number doubles only about every seven years and so. The “law” is just a rough guideline, but even the Wait But Why explainer states that at this rate it’ll take until the end of a century to get to one million. (Last year, DARPA started a challenge to get scientists to work on the same project of recording a million neurons simultaneously.)

Implantation will be difficult.

Right now, implants require brain surgery; Musk keeps saying that this brain-computer interface will need to be non-invasive so that people won’t be afraid to get it. He hopes that one day it will become a fairly affordable and common process, like Lasik eye surgery.

People are always asking if you can get the results of a brain-machine interface from the skull or the skin, says Bouton. “But if you really want to listen to the conversation going on between the neurons in the brain, you have to be close,” he adds. It’s like being in a football stadium. Trying to get get signals from the skin is like standing at the 50-yard line and trying to listen to a conversation of two people in the top section. For things to really work, you need to get a lot closer, and that means inside the body.

We have to make implants that last

So it’s impossible that the interface will be completely non-invasive. But implantation might be less invasive than cutting open the skull and scalp. Musk offered the idea that something can come in through the arteries and then unfold inside the body’s vascular system to interact with the neurons.

The question here, says Bouton, is whether the electrode can still get a good signal by the time it goes through the blood. It’s possible, but there will be challenges when it comes to signal quality.

“They’re talking about injecting this spider-webby sort of thing through the veins into all the little tiny capillaries of the brain and not causing a stroke, and that sounds hugely invasive to me,” adds Miller, the Northwestern neuroscientist. This kind of procedure is dangerous, so safety will matter, particularly for people without disabilities. The reason most brain machine interfaces have been tested in people with disabilities is that they stand to benefit the most — and so will accept some risks in order to be able to, for instance, sip beer on their own again. These risks make less sense in people who aren't disabled and just think a brain implant would be cool. “That’s why I’ve expressed real skepticism that this is ever going to turn into a gaming interface in this century, though we shouldn’t minimize how critically important this could be to therapeutics,” Miller says.

There are other ideas for implantation, too. Researchers like NeuraLink scientist DJ Seo have worked on neural dust, or really tiny wireless sensors that can theoretically be implanted into the brain and then communicate via ultrasound. Charles Lieber at Harvard University is working on neural lace, a type of delicate electronic mesh that you can inject into the brain. Other teams are working with silk interfaces.

The interface has to last decades.

And then there’s the wetware problem: the body is a corrosive place. Right now, most brain implants less than five years — a real problem, since getting surgery every five years is nobody’s idea of a good time. Blood in particular seems to cause a lot of wear and tear.

That’s not all. The body often rejects these implants. Electrodes often don’t last because electrodes are made of metal and they’re put into the brain, which is a soft tissue. Imagine if you put a metal thing in jello and then you just shake the jello — that metal might damage the jello, which here means burst capillaries or worse. Sometimes the scarring around the electrode implants makes it harder for the devices to work.

The solution might be to make something that is, like the body, soft. Unfortunately, it’s trickier to implant something that’s soft. It might be possible to invent a kind of plastic that’s hard at room temperature and soft at body temperature — but that hasn’t been perfected yet, says Bensmaia.

In other words, the thing Elon Musk wants to have ready in four years relies on an organ we don’t fully understand and technology that doesn’t exist yet. While it’s always useful to have more people working on these tough problems, you shouldn’t expect a “wizard’s hat for the brain” any time soon — and certainly not on Musk’s timeline.

4/25/17 9:52 a.m. EDT Correction: The first brain interface was implanted in 1998. An earlier version of this article said it was implanted in 2006.