For the first time, researchers have found a way to non-invasively translate a person’s thoughts into text. Using fMRI scans and an AI-based decoder trained on a precursor to ChatGPT, the system can reconstruct brain activity to interpret the gist of a story someone is listening to, watching or even just imagining telling. Ian Sample speaks to one of the team behind the breakthrough, the neuroscientist Dr Alex Huth, to find out how it works, where they hope to use it, and whether our mental privacy could soon be at risk. Help support our independent journalism at theguardian.com/sciencepod

This week, scientists revealed a system that takes computers a step closer to reading our minds.

We're actually kind of reading thoughts out of someone's head.

But 1 day, the approach made possible by chat-GPT-like technology could be used to understand the thoughts of people with serious neurological conditions.

There are a lot of people who have lost the ability to communicate for various reasons, often because of strokes or brain surgery to fix other things like brain tumours.

Then again, at the speed AI systems are developing, who knows what the future holds?

Geoffrey Hinton, the British-Canadian scientist dubbed the Godfather of AI, resigned from Google this week, warning of the potential dangers of artificial intelligence.

The issue is now that we've discovered it works better than we expected a few years ago.

What do we do to mitigate the long-term risks of things more intelligent than us taking control?

And how do we safeguard our mental privacy?

I'm the Guardian Science Editor, Ian Somerville, and this is Science Weekly.

Alex Huth, you're an Assistant Professor of neuroscience and computer science at the University of Texas at Austin.

And you and your team have just done something really quite fascinating.

You've made an AI decoder, but tell us in a nutshell really what that is and what it does.

Yeah, so we created a system that takes brain scans and reads out the words that somebody is hearing or thinking.

We record people's brain activity using functional magnetic resonance imaging, fMRI, and then we use some algorithms that we've trained on a lot of data from that person to read out words that they are hearing if they're listening to someone speak or words that they're just thinking in their heads.

Like if they're trying to tell a story to themselves without actually moving their mouth, we can read out at least partially what they're trying to say.

So give me an example of sort of what this AI decoder could do.

I mean, if I was thinking just for example, the sentence, Science Weekly is the best podcast around, what might your decoder churn out?

So Interestingly, our decoder doesn't get every single word.

It's a very kind of sloppy, messy signal.

But what it can get is the gist of what you're trying to say.

So for example, it might come up with, my favorite podcast is Science Weekly.

What our decoder is getting at, it's not really just language.

It's not pulling out the syntax or something about the words that you're thinking.

It's getting at the meaning of what you're thinking, the semantics.

I wondered if you could just tell us what your sort of starting point was.

Where were you when you began looking at all of this?

Yeah, so there's a couple different threads that we kind of pull together here.

1 is there's been a lot of work done on what we call invasive brain decoders.

This involves having neurosurgery to put electrodes actually into your brain that can directly record neural activity.

This is of course only appropriate when people are suffering from a serious malady like locked-in syndrome, for example, where they've had a stroke that makes them completely unable to move or communicate.

Of course, that requires neurosurgery, which is a big leap.

Most people are not going to undergo surgery to do this kind of thing.

The other strand of research that we're kind of pulling on is a lot of work done in fMRI, in functional magnetic resonance imaging.

You can go in an MRI machine, anybody can do that, and just get your brain scanned.

It's just measuring these increases in blood flow that happen as a consequence of neural firing.

But thus far, most MRI decoding has been pretty limited.

Limited to getting it like, which of a few words is somebody thinking?

If you're reading a sentence, There's been some more sophisticated things happening in the visual decoding world, where people have subjects look at images inside an MRI scanner, and then they can read out, actually reconstruct the image that the person was seeing.

And we're using kind of some of those same ideas and technologies that are used in the visual decoding space for this language decoding in our work.

So let's dive a bit deeper into how you actually achieved what you're reporting in this latest paper.

First of all, you have to train this AI system.

And I believe that you were getting people in the study to listen to podcasts.

So the data set is 1 of the really critical parts of how the system works.

So we need to map out where all these different kinds of concepts are represented in the brain.

You know, if we're gonna read out the sentence like, my favorite podcast is Science Weekly, then we need to know what parts of the brain will activate when you think about podcasts, when you think about this particular podcast, for example.

And to do that effectively, we need to just have people listen to a lot of language, like a lot of words, so that we can see where these different things are activated in the brain.

So the way that we did that was we had our subjects go back in the MRI scanner over and over and over again, over the course of months, even years, and just listen to podcasts.

That's our basic kind of experiment that we do.

So we collect all this data for our subjects.

In this study, we used about 16 hours of data.

So this is about 16 times they go back in the MRI scanner and listen to stories.

And we have these detailed transcripts of all these podcasts, so we know exactly which words were spoken and when they were spoken.

So then we can use that to then build this mapping, right?

So we build these models that take, you know, what were the words that the person was hearing and predict the brain activity from that.

So you take fMRI images as people listen to podcasts and train the AI, which was based on a precursor to chat GPT, to associate particular patterns of brain activity with particular phrases or sentences.

And then when a person tells a story in their head, the system looks at the fMRI images and works out statistically which words and phrases are most likely to generate the kind of brain activity the scans show.

The way that this decoding works, it's not like how you might imagine it, where we stick in the brain data and out pops words on the other end.

We're doing this strange kind of backwards model.

And this is just outrageously effective, like shockingly effective.

It works so much better than any methods that we had before for modeling how the brain processes language or understands language.

This is, I mean, 1 of the kind of biggest developments in our fields in the past decade, half decade certainly.

So you've, at this point, got a system which can, you've essentially trained it to the point that it can start to say, okay, this kind of brain activity is the kind of thing you get from these kinds of words.

How do you then test out how well this works in people?

Yeah, so we tested it in a few different ways.

So the basic thing that we do is we just take brain responses from the person listening to a different story that we didn't use in our training set.

And then we put the brain responses into this decoding process and generate the words that might have generated those brain responses.

And then we can compare those to the actual words of a story that the person heard.

So we can say, how accurate was this decoding?

And that was the first thing that worked and was very exciting.

Of course, it's not, again, incredibly precise.

It doesn't give us back every word that the person heard.

It gives us this weird kind of paraphrase gist of what the person was saying.

And what then were you trying to do with the silent movie work?

Yeah, so this was kind of an out of left field idea.

Like, wouldn't it be cool if this worked?

But really, it was a test of kind of what kind of information we're getting at with this decoder.

It was trained on people listening to stories.

So there's a lot of language structure in the stories that people are listening to.

So there's a lot of syntactic components.

So we wanted to know, is our decoder getting at that low-level language stuff, or is it getting at something really higher level?

So that's why we did this movie experiment where we had people just watch a silent video.

For some of the videos we used, we actually had these audio description tracks.

These are tracks that are included for people who are blind to be able to watch movies that describe what's happening on the screen.

For some of these movies, we had these audio description tracks.

When we decoded words from the brain data recorded while people were watching these movies, we could then compare it to the audio description tracks and ask, was it similar statistically?

And it was, it was actually decently correlated with what was in the audio description track.

And just looking at it by eye, you can clearly see there's a lot of mess here.

But surprisingly, it really does kind of describe events that are happening in the movie, which honestly, yeah, I was quite surprised by this.

And I think it means that, yeah, what our decoder is getting at is something pretty deep.

And can you use 1 decoder that works on anyone or are the brain signals very specific to each person?

You cannot have 1 decoder that works on anyone or at least not with our current technology.

We thought about this a lot because, of course, this whole thing is kind of creepy, right?

We're actually kind of reading thoughts out of someone's head, which I feel like is crossing a line into something that a lot of people will be kind of unnerved by.

So we spent a lot of time thinking about the privacy implications of this kind of technology.

So we tried to take data from 1 subject and map it to a different subject's brain and the results were just bad.

It was, I guess, slightly above statistical significance, but you couldn't read anything from the text.

On the 1 hand, we were kind of bummed about this.

It's like, ah, that makes it much harder to use.

But on the other hand, I think this is actually kind of good for privacy reasons, right?

I don't think we necessarily want to have that kind of technology in the world, at least yet, before we have a good legal framework around it.

The other kind of privacy thing that we tested that I think is really interesting was, suppose you have this model for a person already, can you actually trust always what it's telling you?

Can you treat it as a truthful readout of what's happening in their brains?

So to really kind of push the limits of that, we had the subjects try to make it not work, try to think other thoughts, try to resist the decoder, we called it.

So in this experiment, we had them listen to a little story.

And then while they were listening to the story, we had them do different things like mental math, like count backwards by sevens, name as many animals as you can, or try to ignore the story and tell a different story inside your head.

The math 1 didn't actually work that well, but the naming animals and telling a story, both of these really disrupted the decoder.

We were unable to read anything out of those brain scans, you know, so we didn't get the story that the person was listening to.

So I think this is also important in the sense that it gives people agency in this kind of situation.

This technology is obviously, you know, in the very early stages, the research stages, but what sort of applications do you see it being used for?

Yeah, so the kind of straightforward application that we would love to see this used for is as an assistive device, right?

So there are a lot of people who have lost the ability to communicate for various reasons, often because of strokes or brain surgery to fix other things like brain tumors, etc.

There are plenty of reasons why people might have damage to parts of the brain that would enable them to speak.

So that's a place where we think this could really help and maybe help people without having them go through neurosurgery.

But this being an MRI scanner makes that pretty impractical, right?

This is like a giant multi-million dollar machine that sits in a multi-million dollar facility with plenty of full-time staff to manage it.

So we think this probably has to move to a different kind of brain imaging technology, which we're investigating right now, some other possibilities.

The technology that we're a bit more excited about is something called functional near-infrared spectroscopy or fNIRS, which works on a similar principle to the photoplethysmograph, the little clip they put on your finger in the hospital to measure your blood oxygenation.

Basically it's shining different wavelengths of infrared light to measure blood flow.

Oxygenated and deoxygenated blood have slightly different colors.

And you can see that if you shine a light through it.

So in FNIRs, you do the same kind of thing, but shining light through a person's skull.

So our skulls are actually somewhat transparent to infrared light.

So you can put some LEDs on someone's head, you can shine the light through, and then measure it at a bunch of locations.

And what this actually measures then is something much more akin to the fMRI signal.

So in that setting, we're kind of hopeful that our technology might adapt nicely to that.

We're working with some folks who are FNIR's experts on trying this out.

And that is something that, even if it's not kind of a consumer device today, it has the potential to 1 day be something that you could use at home.

1 of the big news stories, of course, this week was the so-called godfather of AI, Geoff Hinton, leaving Google and joining those, sort of, warning about the dangers of this sort of technology.

And certainly, as we were saying, mind-reading AI certainly sounds like it could get into that territory of a horror film.

I mean, I'm wondering how you see us reaping the benefits without falling into the sort of 1984 style pitfalls in the future.

So this is 1 of the reasons that we really pushed hard on these questions of mental privacy in this work is that we were worried about this, right?

We thought about this, we knew that this was kind of a scary thing.

So we wanted to really lay out what are the limits on this thing.

And I feel like if we can kind of publicize that information, if we can tell people when does this not work and how can people make it not work, then it makes it much less sort of useful as an instrument in that way.

For example, the polygraph, the standard lie detector, I'm sorry, I'm going to swear for a moment.

The polygraph is absolutely pseudoscience.

It works partially because people believe it works and they get nervous.

So we don't want this kind of misuse to be applied to our technology.

That's why we're kind of trying to spread the word about what it can and can't do.

But these limits are not, of course, set in stone.

So we said that we can't map a model from 1 subject to another subject and we can't read out something if the subject is trying to prevent us, but those things won't necessarily always be true.

As the models get better, as the data gets better, whatever, as the algorithms get better, it might be possible to do these things.

So this is why we also talked about in our paper how we think that legal protections should be put in place for mental privacy and we think this should be done proactively before it actually becomes a problem that's affecting people's lives.

Which, I don't know, expecting anything like that to happen proactively is probably, I don't know, unrealistic, but I would love to think so.

You can read my colleague Hannah Devlin's article about this on theguardian.com.

Before you go, please don't forget to subscribe to The Guardian's podcast series Cotton Capital, which is looking at The Guardian's links to transatlantic slavery.

Episode 5 follows Guardian journalist Lanre Bacaray as he travels to Manchester, the city where the Guardian newspaper was founded, to learn more about Black Mancunian history.

Search for Cotton Capital wherever you get your podcasts and hit subscribe.

The producer was Josh Inchana, the sound designer was Tony Onochukwu and the executive producer was Eli Buri.


Speaker 0: This is The Guardian. This week, scientists revealed a system that takes computers a step closer to reading our minds.

Speaker 1: It's kind of creepy, right? We're actually kind of reading thoughts out of someone's head.

Speaker 0: But 1 day, the approach made possible by chat-GPT-like technology could be used to understand the thoughts of people with serious neurological conditions.

Speaker 1: There are a lot of people who have lost the ability to communicate for various reasons, often because of strokes or brain surgery to fix other things like brain tumours.

Speaker 0: Then again, at the speed AI systems are developing, who knows what the future holds? Geoffrey Hinton, the British-Canadian scientist dubbed the Godfather of AI, resigned from Google this week, warning of the potential dangers of artificial intelligence.

Speaker 2: The issue is now that we've discovered it works better than we expected a few years ago. What do we do to mitigate the long-term risks of things more intelligent than us taking control?

Speaker 0: So how does the mind decode at work? Where could the research take us next? And how do we safeguard our mental privacy? I'm the Guardian Science Editor, Ian Somerville, and this is Science Weekly. Alex Huth, you're an Assistant Professor of neuroscience and computer science at the University of Texas at Austin.

And you and your team have just done something really quite fascinating. You've made an AI decoder, but tell us in a nutshell really what that is and what it does.

Speaker 1: Yeah, so we created a system that takes brain scans and reads out the words that somebody is hearing or thinking. We record people's brain activity using functional magnetic resonance imaging, fMRI, and then we use some algorithms that we've trained on a lot of data from that person to read out words that they are hearing if they're listening to someone speak or words that they're just thinking in their heads. Like if they're trying to tell a story to themselves without actually moving their mouth, we can read out at least partially what they're trying to say.

Speaker 0: So give me an example of sort of what this AI decoder could do. I mean, if I was thinking just for example, the sentence, Science Weekly is the best podcast around, what might your decoder churn out?

Speaker 1: Yeah. So Interestingly, our decoder doesn't get every single word. FMRI is very slow. It's a very kind of sloppy, messy signal. But what it can get is the gist of what you're trying to say.

So for example, it might come up with, my favorite podcast is Science Weekly. What our decoder is getting at, it's not really just language. It's not pulling out the syntax or something about the words that you're thinking. It's getting at the meaning of what you're thinking, the semantics.

Speaker 0: I wondered if you could just tell us what your sort of starting point was. Where were you when you began looking at all of this?

Speaker 1: Yeah, so there's a couple different threads that we kind of pull together here. 1 is there's been a lot of work done on what we call invasive brain decoders. This involves having neurosurgery to put electrodes actually into your brain that can directly record neural activity. This is of course only appropriate when people are suffering from a serious malady like locked-in syndrome, for example, where they've had a stroke that makes them completely unable to move or communicate. Of course, that requires neurosurgery, which is a big leap.

Most people are not going to undergo surgery to do this kind of thing. The other strand of research that we're kind of pulling on is a lot of work done in fMRI, in functional magnetic resonance imaging. That's the technology that we're using. It doesn't require surgery. You can go in an MRI machine, anybody can do that, and just get your brain scanned.

It's just measuring these increases in blood flow that happen as a consequence of neural firing. But thus far, most MRI decoding has been pretty limited. Limited to getting it like, which of a few words is somebody thinking? What topic is a sentence about? If you're reading a sentence, There's been some more sophisticated things happening in the visual decoding world, where people have subjects look at images inside an MRI scanner, and then they can read out, actually reconstruct the image that the person was seeing.

And we're using kind of some of those same ideas and technologies that are used in the visual decoding space for this language decoding in our work.

Speaker 0: So let's dive a bit deeper into how you actually achieved what you're reporting in this latest paper. First of all, you have to train this AI system. And I believe that you were getting people in the study to listen to podcasts. Tell us what you did there.

Speaker 1: Yeah, that's right. So the data set is 1 of the really critical parts of how the system works. So we need to map out where all these different kinds of concepts are represented in the brain. You know, if we're gonna read out the sentence like, my favorite podcast is Science Weekly, then we need to know what parts of the brain will activate when you think about podcasts, when you think about this particular podcast, for example. And to do that effectively, we need to just have people listen to a lot of language, like a lot of words, so that we can see where these different things are activated in the brain.

So the way that we did that was we had our subjects go back in the MRI scanner over and over and over again, over the course of months, even years, and just listen to podcasts. That's our basic kind of experiment that we do. So we collect all this data for our subjects. In this study, we used about 16 hours of data. So this is about 16 times they go back in the MRI scanner and listen to stories.

And we have these detailed transcripts of all these podcasts, so we know exactly which words were spoken and when they were spoken. So then we can use that to then build this mapping, right? So we build these models that take, you know, what were the words that the person was hearing and predict the brain activity from that.

Speaker 0: So you take fMRI images as people listen to podcasts and train the AI, which was based on a precursor to chat GPT, to associate particular patterns of brain activity with particular phrases or sentences. And then when a person tells a story in their head, the system looks at the fMRI images and works out statistically which words and phrases are most likely to generate the kind of brain activity the scans show.

Speaker 1: Yeah, that's exactly right. That's a nice way to put it. The way that this decoding works, it's not like how you might imagine it, where we stick in the brain data and out pops words on the other end. We're doing this strange kind of backwards model. And this is just outrageously effective, like shockingly effective.

It works so much better than any methods that we had before for modeling how the brain processes language or understands language. This is, I mean, 1 of the kind of biggest developments in our fields in the past decade, half decade certainly.

Speaker 0: So you've, at this point, got a system which can, you've essentially trained it to the point that it can start to say, okay, this kind of brain activity is the kind of thing you get from these kinds of words. How do you then test out how well this works in people?

Speaker 1: Yeah, so we tested it in a few different ways. So the basic thing that we do is we just take brain responses from the person listening to a different story that we didn't use in our training set. And then we put the brain responses into this decoding process and generate the words that might have generated those brain responses. And then we can compare those to the actual words of a story that the person heard. So we can say, how accurate was this decoding?

So that was kind of step 1. And that was the first thing that worked and was very exciting. Of course, it's not, again, incredibly precise. It doesn't give us back every word that the person heard. It gives us this weird kind of paraphrase gist of what the person was saying.

Speaker 0: And what then were you trying to do with the silent movie work? Because I know you brought that

Speaker 1: in as well. Yeah, so this was kind of an out of left field idea. Like, wouldn't it be cool if this worked? But really, it was a test of kind of what kind of information we're getting at with this decoder. It was trained on people listening to stories.

So there's a lot of language structure in the stories that people are listening to. So there's a lot of syntactic components. So we wanted to know, is our decoder getting at that low-level language stuff, or is it getting at something really higher level? That's like, what is the idea? So that's why we did this movie experiment where we had people just watch a silent video.

For some of the videos we used, we actually had these audio description tracks. These are tracks that are included for people who are blind to be able to watch movies that describe what's happening on the screen. For some of these movies, we had these audio description tracks. When we decoded words from the brain data recorded while people were watching these movies, we could then compare it to the audio description tracks and ask, was it similar statistically? And it was, it was actually decently correlated with what was in the audio description track.

And just looking at it by eye, you can clearly see there's a lot of mess here. It's worse than the language decoding. But surprisingly, it really does kind of describe events that are happening in the movie, which honestly, yeah, I was quite surprised by this. And I think it's pretty cool. And I think it means that, yeah, what our decoder is getting at is something pretty deep.

Speaker 0: And can you use 1 decoder that works on anyone or are the brain signals very specific to each person?

Speaker 1: Yeah, absolutely not. You cannot have 1 decoder that works on anyone or at least not with our current technology. We thought about this a lot because, of course, this whole thing is kind of creepy, right? We're actually kind of reading thoughts out of someone's head, which I feel like is crossing a line into something that a lot of people will be kind of unnerved by. So we spent a lot of time thinking about the privacy implications of this kind of technology.

So we tried to take data from 1 subject and map it to a different subject's brain and the results were just bad. It was, I guess, slightly above statistical significance, but you couldn't read anything from the text. It made no sense. On the 1 hand, we were kind of bummed about this. It's like, ah, that makes it much harder to use.

But on the other hand, I think this is actually kind of good for privacy reasons, right? I don't think we necessarily want to have that kind of technology in the world, at least yet, before we have a good legal framework around it. The other kind of privacy thing that we tested that I think is really interesting was, suppose you have this model for a person already, can you actually trust always what it's telling you? Can you treat it as a truthful readout of what's happening in their brains? So to really kind of push the limits of that, we had the subjects try to make it not work, try to think other thoughts, try to resist the decoder, we called it.

So in this experiment, we had them listen to a little story. And then while they were listening to the story, we had them do different things like mental math, like count backwards by sevens, name as many animals as you can, or try to ignore the story and tell a different story inside your head. The math 1 didn't actually work that well, but the naming animals and telling a story, both of these really disrupted the decoder. We were unable to read anything out of those brain scans, you know, so we didn't get the story that the person was listening to. So I think this is also important in the sense that it gives people agency in this kind of situation.

Speaker 0: This technology is obviously, you know, in the very early stages, the research stages, but what sort of applications do you see it being used for?

Speaker 1: Yeah, so the kind of straightforward application that we would love to see this used for is as an assistive device, right? So there are a lot of people who have lost the ability to communicate for various reasons, often because of strokes or brain surgery to fix other things like brain tumors, etc. There are plenty of reasons why people might have damage to parts of the brain that would enable them to speak. So that's a place where we think this could really help and maybe help people without having them go through neurosurgery. But this being an MRI scanner makes that pretty impractical, right?

This is like a giant multi-million dollar machine that sits in a multi-million dollar facility with plenty of full-time staff to manage it. So we think this probably has to move to a different kind of brain imaging technology, which we're investigating right now, some other possibilities. The technology that we're a bit more excited about is something called functional near-infrared spectroscopy or fNIRS, which works on a similar principle to the photoplethysmograph, the little clip they put on your finger in the hospital to measure your blood oxygenation. Basically it's shining different wavelengths of infrared light to measure blood flow. Oxygenated and deoxygenated blood have slightly different colors.

1 is more red and 1 is more blue. And you can see that if you shine a light through it. So in FNIRs, you do the same kind of thing, but shining light through a person's skull. So our skulls are actually somewhat transparent to infrared light. So you can put some LEDs on someone's head, you can shine the light through, and then measure it at a bunch of locations.

And what this actually measures then is something much more akin to the fMRI signal. So it's actually the exact same signal. It's this bold blood oxygenation signal. So in that setting, we're kind of hopeful that our technology might adapt nicely to that. We're working with some folks who are FNIR's experts on trying this out.

And that is something that, even if it's not kind of a consumer device today, it has the potential to 1 day be something that you could use at home.

Speaker 0: 1 of the big news stories, of course, this week was the so-called godfather of AI, Geoff Hinton, leaving Google and joining those, sort of, warning about the dangers of this sort of technology. And certainly, as we were saying, mind-reading AI certainly sounds like it could get into that territory of a horror film. I mean, I'm wondering how you see us reaping the benefits without falling into the sort of 1984 style pitfalls in the future.

Speaker 1: So this is 1 of the reasons that we really pushed hard on these questions of mental privacy in this work is that we were worried about this, right? We thought about this, we knew that this was kind of a scary thing. So we wanted to really lay out what are the limits on this thing. And I feel like if we can kind of publicize that information, if we can tell people when does this not work and how can people make it not work, then it makes it much less sort of useful as an instrument in that way. For example, the polygraph, the standard lie detector, I'm sorry, I'm going to swear for a moment.

It's bullshit. It doesn't work. It's pseudoscience. The polygraph is absolutely pseudoscience. It works partially because people believe it works and they get nervous.

So we don't want this kind of misuse to be applied to our technology. That's why we're kind of trying to spread the word about what it can and can't do. But these limits are not, of course, set in stone. So we said that we can't map a model from 1 subject to another subject and we can't read out something if the subject is trying to prevent us, but those things won't necessarily always be true. Right?

As the models get better, as the data gets better, whatever, as the algorithms get better, it might be possible to do these things. So this is why we also talked about in our paper how we think that legal protections should be put in place for mental privacy and we think this should be done proactively before it actually becomes a problem that's affecting people's lives. Which, I don't know, expecting anything like that to happen proactively is probably, I don't know, unrealistic, but I would love to think so.

Speaker 0: Thanks so much for coming on, Alex. Thanks again to Alex Huth. You can read my colleague Hannah Devlin's article about this on theguardian.com. Before you go, please don't forget to subscribe to The Guardian's podcast series Cotton Capital, which is looking at The Guardian's links to transatlantic slavery. New episodes are released every Monday.

Episode 5 follows Guardian journalist Lanre Bacaray as he travels to Manchester, the city where the Guardian newspaper was founded, to learn more about Black Mancunian history. Search for Cotton Capital wherever you get your podcasts and hit subscribe. And that's it for today. The producer was Josh Inchana, the sound designer was Tony Onochukwu and the executive producer was Eli Buri. We'll be back on Tuesday.

See you then. This is The Guardian. You

See all Science Weekly transcripts on Applepodcasts

How AI is making non-invasive mind reading a reality