From “MIT proves ChatGPT makes you dumb” to miracle brain implants, neuroscience headlines are everywhere — but how do you know what’s real? To help us not only answer those questions but learn how to answer them, we invited Ashley, a neuroscientist and professor at UC San Diego, to help improve our “neurobullshit detectors.” Using real examples from Neuralink hype to mouse studies and misleading brain scans she shows how to separate headline from substance.
Transcript
I’m going to do the bold thing and take out the microphone, like a standup comedian or something. Hi, good morning, this is my first Monktoberfest and it’s been wonderful, so just thank you already in advance.
[applause]
So the title does give it away a little bit here, but I want to set the stage. So within your skull, you have an amazing bit of hardware. It weighs about three pounds. It operates at about 10 bits per second, and within that organ, you’ll find the neurobullshit detector an understudied and frequently underdeveloped piece of tissue.
My name is Ashley and I’m a neuroscientist and educator, I’m a professor at UC San Diego, and I want to tell you a little bit about how to use this piece of equipment today.
So why am I here? A few things. You’re going to find a lot of neuroscience information out in the world and it’s seductively creeping into everything and I want you to have the tools to suss that out.
I’ve been teaching about neuroscience for more than a decade and somehow that hasn’t solved everything, so I thought maybe I’d come to a tech conference and maybe you people with all of your tools could figure it out and so I’m hoping that maybe I’ll plant a little seed today that will go somewhere after this place and the third thing is that well, you have a brain, and you probably want to know how to use it. There’s an asterisk here, though, because although you probably have a brain, because you’re here, you made it here, how much brain you have is actually unknown unless you’ve taken an MRI recently, so I want to give you something more to worry about today. Which is that, not that long ago, this was about ten years ago, someone showed up to the hospital with a little bit of slurred speech, not unlike the speech I heard last night a little bit and showed up and had a huge chunk of brain missing. There’s supposed to be stuff happening in the back here.
So this person had what was called cerebellar agenesis, so they don’t have a cerebellum and they’ve never had a cerebellum but it didn’t show up until this person was in their 40s or 50s, it’s not the only case of someone missing a significant chunk of brain tissue unbeknownst to them. And my favorite thing on this wags the headline that says 24-year-old pretty much OK without cerebellum.
So we’re going to talk today — there’s drinks that are meant to help your brain, you can buy a hair drier that’s called a neurohair drier, you can even get a med bed, I’ve heard that’s a thing.
So how do you know that it’s actually legit, right? So what I’m going to do today what educators call active learning so I’m going to show you a headline and I want you, on a scale of 1-5 or however many fingers you have, I want you to tell me what you think about these headlines, so we’re going to do three of them. The first one here says brain implant translates thoughts to speech in an instant. OK, on a scale of 1 is bullshit, 5 is very accurate, what do we think?
Oh, got a wide range. We’ve got a skeptical crowd here. I actually see a good range. A little bit more on the lower end of the spectrum. OK, we’ll keep that in mind.
This one says brain cells that shape social choices identified.
MIT proves ChatGPT makes you dumb. This is not the way I thought that was going to go. OK. OK. Interesting, really interesting.
- This is why we do this, because now I know where you are. OK. So what we’re going to do here is work through each of these as examples. And you know, the whole scheme of this is to say, OK, in this landscape of headlines, how do you know what is actually real? So let’s start with this first one: So brain implant translates thoughts to speech in an instant. OK, so what we’re working with here is a brain computer interface, so this is a device that it take a brain signal, analyze it, translate it into a message that can be carried out to om so device external to the device, this could be a prosthetic arm or and sometimes these are called brain machine interfaces or BMIs, OK
There are several different types of these. Noninvasive, semi-invasive and invasive. Coincidentally, three different types of engineering management, right?
[laughter]
[applause]
And so on the noninvasive side, we have EEG and MEG. These are things not in your brain, they’re listening to your brain. We’ll talk about those in detail in a moment. Then we can do semi-invasive. I think this is weird that this is semi-because they’ve taken the skull off. That’s pretty invasive. But in the grand scheme of things, it’s in the middle. They’ve put in electrodes on the brain surface,
And in invasive, we’ve actually stuck electrodes into the brain.
Fundamentally they’re all doing the same process. So the first thing they do is they record brain or muscle signals. The second thing they do is that they process and decode, so though interpret those signals and the final thing is they output these things to some sort of effector, like I say, an arm, leg, a video game, whatever it may be.
And neuroscience, this is one — actually as it turns out, this process of actually trying to decode what the brain is trying to say, turns out to be a little easier than we thought, and I think that’s amazing.
So, you know, this has been a long landscape of research. I’m going to tell you a little bit about it. But, you know, about four years ago, someone from mostly your world stepped into my world and made a big land claim here and said we’re going to fix this whole thing, we’re gonna do it. You may know who this is. A guy by the name of Elon Musk, right? And he said the first neura link product is going to — they did some amazing things, in 2021 they released their first videos, which is this monkey playing mind pong which was really cool and the neuroscience community was called to action and said, wait a second, we’ve been doing this for more than 20 years. And I don’t know how many academics you know, but this is a shit post, right?
[laughter]
Yeah, like, we have done it. So shots fired, right and the whole neuroscience community was lit up. So in 1965, Thelma Estrin, a woman with great arms, I will say, formulates requirements for a BCI. So makes this very formal. Here’s what we need in order for this technology to exist.
A little bit later, UCLA begins work in monkeys, so as a first non-human animal model to get this working. Then in the 1990s, we start doing work on humans.
1998 is the first time we actually do this invasively in a human. Yeah, when Elon Musk was like a teenager.
And in 2000, we get the first formal definition of what a BCI is and then finally we get a clinical trial in the early 2000s, and then in 2016, neura link is formed. So we’ve been working on this shit for a longer time * and almost all of this is federally funded research that Elon Musk is capitalizing on. Yup.
So I’m going to show you a video, because I cry like every time I see this. It’s amazing. This came out of UC Davis just earlier this year. This is a patient who has ALS or Lou Gehrig’s syndrome. You’re going to actually see him speak for the very first time, using this implant.
>> Demonstrations of the process took weeks or months to get working, the first time he tried to use the system just worked.
>> OK, ready? That first time that he tried to speak and that word appeared on the screen, … he started crying. I started crying. His family started crying. Everyone — we had to pause. We stopped the rest of that, everyone composed themselves, all right, let’s do it again and it kept working, so yeah, it was really special, it was a special moment.
>> At this point, we can control what somebody is trying to say about 97% of the time. Which is better than a lot of commercially available smartphone —
>> So it’s not that we’re reading people’s minds, we’re not detecting their inner thought or their inner monologue or even their mind’s voice, we’re really detecting their attempt to prove their muscles and to talk.
>> People who do not have the ability to communicate quickly are often isolated and lonely, having something like this will give them more of an opportunity to be more themselves.
[applause]
>> So, incredible stuff, and this is only getting better. The actual — the part of this that an instant is actually true. But this is about a group doing similar work and they really have shown that in about 20 milliseconds you can translate what someone is trying to say. So it is almost instantaneous. So believe it or not, this is very, very real and very, very cool. How do you know it’s real? You can see the outcomes for yourself. There’s a video. It was conducted by neuroscience researchers with experience. You can trace this back to the 1960s, there’s a history here. And finally, multiple groups have done this, and have doing this, including an industry and I don’t want to down-talk the evolving nature of this in industry, too, and that’s amazing too. I want those to exist and I want credit to be given where credit is due.
So very, very cool work here, so I’m glad to say that this does exist and yeah, we can stand behind it and be very excited about it.
So let’s go ahead and take the second one here. So brain cells that shape social choices identified. OK, so what is this about? So this is a little bit of a case of just some nice headline-writing, right? So this is a case of something like this, where, you are know, this is the paper headline, oh, really fancy, it sounds really nice. I don’t know if you can read the subtext, but it says a neuron cluster in mice shows choices in — OK, what does the actual paper behind this headline? Well, the paper is parva albumin interneurons — the press release writer did a pretty good job er hot so let’s talk about what’s in this paper, then. So in order to do so, we’re going to think about the different levels at which we we can understand the brain. We’re going to think about behavior and cognition and networks and when we think about this layer of understanding, we’re thinking about what is it like to listen to the brain from the outside? Sort of like listening to a sports stadium, maybe the Red Sox and the Yankees, and you’re hearing the crowd roar. So you company find are there exciting things happening, what’s the tenor of the stadium so that’s our higher level understanding.
And then you can go deeper, circuits and neurons. So this is my local women’s soccer team, San Diego wave, we have a very gay supporter section, it’s great and you can go in and hear what’s going on in that specific section, right? So every time our keeper saves the goal, this section is roaring and you can hear what’s going on there. Have
But then we can get more specific information, right? We can listen to what single neurons are doing, maybe even what a synapse or a channel is doing in this, right? And that’s the equivalent of me seeing my friend on TV with his kid drinking a beer and this is real. On national TV.
So you get different information from a single person than you do the supporter section than you do the whole stadium, right? But altogether these are very important pieces of information in order for us to understand what’s going on in the brain, right? Typically neuroscientists are observing a particular window here, we’re observing this within a particular context, and so from me, like my Ph.D. was in this level, sort of systems, circuits, neurons and usually because we don’t have a choice. I have to choose a level here. I can’t measure all of these things simultaneously.
So this is kind of where I would be.
So this paper is very interesting, and it’s an example of this. It’s a specific context in which we’re understanding something.
So need nonhuman animal models to study these mechanisms. I mean what are the roles of different proteins, molecules, and even different circuits. And I need a process of what’s the causality? Is there a necessary component? What components are sufficient, etc. So in almost every case, we need a nonhuman animal model and it doesn’t have to be a model. It could be a batch of cells, it could be what’s called an organoid, which is that you leave cells in a dish and they turn into an organ, it’s very sci-fi, but it’s real.
It’s a particular kind of neuron, in one brain region, helping us understand a social behavior, but in this case, in mice, right? So we have to always ask, does the animal model tell us what we need to know in a human model. And that’s the question we always have to ask ourselves.
So we use these nonhuman animal models to create contexts. We want to understand what happens when we tip this over, what else tips over in that system and we especially need to do this again when we can’t study those contexts in humans. Like, I can’t ethically give someone Parkinson’s disease or cancer, right? If I want to go into their brain I need to have a context with which to study it.
I want to give you a really specific example of sometimes where these contexts come from and how we discover them, and this one’s just wild. So in the 1980s, this synthetic opioid was being used by some folks, it’s called MPPP, and somebody messed up at some point and instead of making MPPP, they made MPTP, which is a different compound. And a bunch of students that took these MPPP showed up at an emergency room with Parkinson-like symptoms. As it turns out for reasons that we don’t fully understand, this drug actually creates a Parkinson-like context in the brain. It kills the same exact cells that Parkinson’s now kills. So what we can now do and what researchers do on a regular basis, is take this toxin, inject it into a mouse or say a monkey and study what it can do, because that can tell us what happens with Parkinson’s.
So we can then use that to something that we care about, which in this case would be Parkinson’s. Someone wrote a book about this if you’re interested.
>> So all that said, obviously nonhuman models are limited. There’s only so much we can do with a nonhuman model, right? So once again I’m going to ask you to kind of give me your opinion here. So in a moment I’ll have you vote, so just a classic thumbs up or thumbs down, so using only a mouse model, just mice, could we do the following?
OK, first one is determine the cause of schizophrenia. Thumbs up or thumbs down.
All right. I see mostly thumbs down and I would mostly agree. I think singlehandedly with mice alone we’re probably not going to do this. We might get some clues, but we’re not going to singlehandedly get a cause, if there is even a single cause, which is maybe a bad premise in itself.
>> No. two, identify genes that might contribute to schizophrenic phenotypes? Thumbs up or thumbs down? Yeah, I think we’ve studied those in mice and they teach us what happens —
And finally describe cell types that encode perceptual differences? Thumbs up or down? Different cells in the brain that to help us understand perception? A little more mixed on this one? This one is pretty good. Mice have a different perception than us in the world. We have to take that into consideration, there’s hard to study for example hallucinations in mice. You can, but it gets a little tricky. So this paper is an example of maybe a mild-hype headline but pretty limited science. But it gives us some insight. So once again, it was conducted by neuroscience research, there’s a visible trajectory that I didn’t talk about, and finally there’s some replicable here. So other folks are done this research *
With that, let’s go to the third paper, the juicy one, right? This is a paper I have been trying to avoid in all honesty, however, it keeps showing up in my feed in like a number of different ways. I don’t know if anybody else has seen this, right? So AI harms real people. MIT just completed the first brain study, this is shutting down your brain. This is like the stuff that’s showing up on my, you know, just typical Instagram feed and this is the type of information that’s getting out in the world. I’m not going to go into a super deep dive if you want to hear my wife and I talk about this for an hour and talk about more than just the neuroscience, we did on our podcast, change, technically. In this case, you deserve better brain research. But let me give you a quick breakdown of what the neuroscience in this paper is doing. So first and foremost, this is a preprint and preprints are great for science. We saw the importance of preprints during the pandemic, right? So bio archive has been around since 2013, but in 2020, we saw the most preprints on bio ARKIV — there’s other preprint servers, ARKIV and med RXIV is — preprints have really taken off. If you were to look in 2018, you’d see like a crazy spike like I said going into 2020. So these are important, but, there’s a big, big but here, which is that they are not peer reviewed. But it is the best thing we have at the moment for us to say whether or not science makes sense to other scientists, and so this is the major downside of preprints and it’s just a think that we always have to think about it with a grain of salt. So er I’m here basically peer reviewing this paper for you all, right? So what is the main neuro claim in this paper. There’s lots of other claims, but we’re going to focus on the brain stuff. The claim is that participants showed weaker neural connectivity. Fundamentally this is the thing they would like you to believe.
Let’s think about connectivity and so to think about connectivity, we need to think about plasticity, because fundamentally this is a claim about your brain changing when you use ChatGPT. So this is a question of plasticity. So how does plasticity happen? The first thing I want you to know is that the structure of your brain doesn’t change very much. Like, the gross structure of it doesn’t change very much. There are major exceptions to this, of course, if you have some very extreme form of dementia, for example, here is advanced Alzheimer’s, which we heard about yesterday, this causes dramatic changes in the structure of your brain. Not depending on whether you went for a walk today, right? That’s not going to happen it. So most changes in plasticity is not happening — it can happen via the expression of your genes. So you have a genome and the expression of that genome can change over the course of your life on some time scale.
The thing is that we can’t actually measure either of those things. We can’t measure synapses in individual neurons and we can’t measure gene expression in individual human beings, so we are fundamentally limited in our ability to measure plasticity of that kind. We can do things like you see on the album cover here, we can do things like measure the tracks in your brain, the big tracks, but these are not single neurons, these are big bundles of neurons all together.
So what can we do? What we observe is not nature itself, but nature exposed to our nature of questioning. I think this is a fundamental idea of any study. We’re observing it through a window, right? So we’re going to observe the brain again through a window and in this paper they use this technique called EEG and EEG going back to our analogy of the stadium is like standing outside of the stadium and listening to what’s going on. It’s a very coarse measure of brain activity. And EEG is a few centimeters away from where the stuff is happening. There’s a couple of things that make EEG kind of a difficult technique and the first is you’re not really up against the neurons themselves, the things producing the electrical current. You’re like a centimeter or two away, depending on the thickness of your skull. So you’re already kind of far from the source, that’s a problem, so you’re listening again to the stadium.
So EEG activity looks something like this, two get these kind of squiggles, these are different channels and stuff is going on over time and we can do things like try to infer what’s going on in between these channels, so we can infer if those channels are more correlated in time using some fun causality, we can ask, OK, these two channels are probably doing something together. Maybe they’re connected. And we call that functional connectivity. So we can generate an activity matrix and we can say, OK, we think these molds of the brain are connected based on the fact that their activity is correlated together.
So this is something similar to what I just showed you, a quote-unquote normal neurotypical brain. And this is what it looks like in epilepsy so you see a ton of activity here and also a lot of asynchronicity. And actually more connectivity than a neurotypical brain.
So it’s not always clear whether more is better, right? So I showed you that in an epileptic brain there’s more functional connectivity. When we think of something like size, there’s some good evidence that brain area increases in size when you do things like learn how to play the piano, when you get good at basketball. There’s a good study with London cab drivers having a big spatial area of the brain.
But there’s also some studies that show that if you overwork you get increases in your brain matter. What does that actually do? Thinking about connectivity I showed you that epilepsy shows increase in connectional activity, but something that we can also think about over the course of brain development, you actually lose connections. So in some ways, having fewer connections might actually be better, like a more streamlined highway system basically in your brain.
But the conclusion of this paper is that using something like ChatGPT lowers connectivity.
So at the end of the day you need a theoretical framework and you need a development of ideas, is a theory for why the brain is the way it is and what it means to be off its kilter, right?
I want to say, too, that as EEG headsets get quicker they get easier to use and then show you the data, right so this is just going to get easier and easier, you can get an EEG headset for like 500 bucks and do your own study and claim that, I don’t know, productivity is related to alpha or something, right? So my message to you is I need you to be more skeptical. Because it’s not always clear what lesser or more means or alpha or etc.
So the question shouldn’t always be what is happening in the brain, right? I’m a neuroscientist, I love brains, but sometimes it depends on what we care about. If it’s the case of ChatGPT, we might care about students cheating, the loss of critical reasoning, and so that’s human behavior. So let’s do. Human behavior work instead of putting a brain on it, right. So this study didn’t test ChatGPT leads to dumb. It didn’t demonstrate changes in neural connectivity once you get into the science of it. Maybe, but it’s still hard to understand what these studies led to.
Sometimes the best neuroscience research — most of the research I talked about in this talk was done in the UC system which is near and dear to my heart. I’m hitting time so I’m going to skip past. Just other guidelines. Don’t trust fMRI research. Someone put a salmon into an F MRI *scanner, we can talk about that. Finally, very few things apply to all brains, so if you’re a left-handed person, your brain is actually probably a little bit different than the right-handed people in the room in terms of where your language dominance is. So if you want to talk about these things, come talk to me. I am online and you can find me there. Thanks so much for your attention.
[applause]
