Deepfakes

Calling all young tech creators. Have you made something cool with code? Maybe a game, a robot, or a website? Then come take part in Coolest Projects 2026, a free event where we celebrate young digital creators like you! Join us in person on Saturday the sixteenth of may at the Life Centre in Bradford, or take part in coolest projects online from anywhere in the world.

Visit coolestprojects.org to find out more. Welcome back, young detectives, to Mysteries of Science. This is the podcast that never takes anything at face value. So from the archives today we're diving into a digital Who Dunit all about deepfakes. Uh, in this episode you'll find out all about AI generated trickery and you get the chance to test your detective skills. Can you tell the difference between real and fake? Hello and welcome to Mysteries of Science.

And I'm the editor of The Week Junior Science and Nation. which is the monthly magazine from the team behind the week junior. And I'm Michael, the Acting Deputy Editor. On this podcast, we explore the strange phenomena and bizarre events that have left scientists scratching their heads. And despite their best efforts, remain well and truly unsolved.

In recent months, there's been an outbreak, let's call it an outbreak, of pictures and videos and audio clips across the internet, which look and sound. The Pope wearing a big white black. jacket and there was um a video of the former US President Barack Obama talking about healthcare. So these these videos and pictures they look and Real, but in fact they are 100% totally fake. They're known as deep fakes. Yes, now I've heard about these, but why would How do they want to do that?

But how can we believe what we see? Well Michael, why don't we gather together some very real experts and find out the truth behind deep fakes? This is Mysteries of Science. Okay Dan, with deep fakes it sounds to me like you don't know what's real or who to trust. So how are we gonna find experts for this one?

Thankfully, Michael, at the Week Junior Science and Nature, we've made some really good friends over the years, um, including the Raspberry Pi Foundation. So readers of the magazine may recognise them uh from our coding team Uh Raspberry Pi is a UK charity which encourages young people to get involved with computing and digital technologies. So as we're talking about world.

than them. Hi, I'm Ben Garside and I'm a content creator at the Raspberry Pi Foundation. Welcome to the show, Ben. Thanks for joining us. So tell us, what exactly are deep fakes? Sure. Yeah absolutely. So Imagine that you have a really cool tool that lets you take videos of people and change their faces or voices to make it

sound or feel like or look like it's someone else. So that's essentially what a deepfake is. Um and all deepfakes are created using this thing that everyone seems to be aware of now which is called artificial intelligence or AI. Yes, that's right. AI, that's something else we've explored on the podcast before, I believe, my. Yes, season four, episode three to be precise. Is your computer smarter than you? If any of our listeners wanted to go back and check it out, oh wait until you finish.

Yeah, of course. Um and AI is basically, if I understand it correctly, AI is basically machines that can think. Yeah, so an AI is a computer that's able to do something which would ordinarily require human intelligence. Right, so so deep fakes are then fakes create

But how does the AI do it? How does it learn to make a realistic copy? of a real person. Well, there are different ways in which deepfakes are created, but importantly for all AI models, deepfakes or not, they all rely on lots and lots of data to be able to create these models in the first place. Um so if we're trying to create a deep fake, so let's say for example a video pretending to be someone else.

Ideally the videos that you're using or the recordings or the images would show the person in different lighting conditions, different facial expressions and so on. Okay, so basically you're training the computer how to recognize this particular person, but how does Take all of that information and then make something out of it that

Most deepfakes are created now. There were different approaches, but most of them are used call with by taking this approach using things called Now GAN stands for general adversarial networks. Now I really, really love this term, particularly around adversarial, because it's really descriptive about how the process actually works.

Um and I don't know about you, but when I hear the term adversary, I I kind of well, it means two people that oppose each other, but I always think of like superheroes that has like an arch nemesis or something like that. Yeah, I love that. Pfff like Batman and Joker or Superman and Lex Luther. Well, who would our two adversaries be here and how

How are they combining to create a deep fake? So in relation to deepfakes, guns have these two adversaries and one of them's called the generator and one of them's called the discriminator. Now the generator will start off by generating a completely random image. I mean unrecognizable from anything, just a bunch of pixels, uh not at all like the person we're trying to recreate.

Now that image is then sent to the discriminator. Now the purpose or function of the discriminator part of the algorithm is to compare that image that's been generated by the generator with all the other data it has in this training data. And if it doesn't look anything like uh the person we're trying to recreate, then the discriminator will send that back to the generator with um a bit of feed.

saying, no, it looks nothing like uh what we want and try again. And that process goes over and over again until eventually we ended up with something that the discriminator can't actually tell the difference between what's being created by the generator and actually the the data it has in its in it in the training data. That sounds exactly like how we work, Dan. I come to you with the work that I've done and you say, nope, that's nothing like what we want, try again.

Yeah, that's that's so true. Well, you know, when you're making magazines and podcasts, Michael, it's all about getting it to be the best it can be. Um actually I wonder if this reminds any of our listeners Of their teachers, maybe they should start calling their teachers the discriminator. Well, just like school students, AI needs to learn too in this.

process is often referred to as machine learning or deep learning. And that's actually where the name deep fake comes from. It's a combination of deep learning and fake. Okay, so so Ben's explained to us what deepfakes are and how they're made.

But what about how they're used? You often hear about the ways that they're used to like trick and deceive people, but I'm wondering are there any positive ways? But that they can be used. I remember going to the cinema to see this film called Star Wars Rogue One, and I think that was back in twenty sixteen. Uh and for people who aren't uh so interested in Star Wars, this film was positioned to be the like the backstory to the first Star Wars film.

Episode four, which was released I think in the seventies, so definitely like forty years before Rogue One was created. Now for some characters in the film it didn't really matter that they had different actors, like Darth Vader who wears a cape and a mask. For example, Princess Leia, she appeared in this Rogue One film looking exactly and sounding exactly like she did in this film uh from the 1970s.

So that was a really cool for me, that was a really cool use of um deep fake technology. Oh, that's interesting. I'm a massive Star Wars fan and Rogue One is actually one of my favorite films in the whole series. And I remember being

blown away by the special effects when I saw it, but I'd never actually considered it as an example of deep fake technology until now. Yeah, that makes total sense actually. I I'm guessing that films are are really one area where deepfakes can really sort of be used to all sorts make loads of different effects, like deaging actors or bringing characters back to life.

in ways perhaps that wouldn't have been possible before. Exactly. Though I guess in those instances, people know that what they're watching isn't real. We're looking at fictional characters here, but When deepfake technology is used to recreate real people, then it can get a bit darker, a bit murkier. So how can we tell the difference? Well, Michael, I think it's time.

We brought in our next expert. Hello, my name is Kimberly Mai. I'm a PhD researcher at UCL and my research is focused on machine learning and artificial intelligence. Welcome to the show, Kimberly. Now Kimberly is something of a deep fake detective, and recently her team did a study into how good humans are at detecting deep fakes, in particular deep fake speech. So what we wanted to Measure was how well humans can detect speech defects. We also wanted to measure if

there's any difference in detection capability between languages. So we looked at English and Mandarin Chinese. And thirdly, how well and thirdly, um, if humans can be trained to get better at detecting deepfakes. So what we did was we conducted an online study of 500 people and we asked them to listen to 20 clips um twenty clips spoken either by a

Real person or by an AI and ask them to dis to decide whether the clip was real or fake. And what we found was that humans could only detect speech deep fakes 73% of the time. There wasn't really any difference. difference in detection capability between English and Mandarin and training people, so how we did that was we let people listen to some examples of speech deep fakes before doing the real task. Um that only helped slightly, but it didn't really improve before

Seventy-three percent of the time is basically like spotting three out of every four fakes. So I mean that's not terrible really, but it still means you're being fooled. one out of every four times. So I think there's there's a bit of room for improvement there. Um Kimberly, why do you think people struggle to tell real speech and fake speech apart? So I think the reason why people weren't very good at detecting deep fakes was when we analyzed their responses, is that across English and Mandarin

people tend to rely on um intuition to make decisions. So for example, they would listen to a clip and they would say, Oh, the clip sounds quite natural, so it it must be real rather than, I guess, definitive things like for example mispronunciations or strange intonations.

interesting. I guess we do tend to fill in the gaps a bit, don't we? Like if we're reading a a sentence, we might not notice straight away that a word's been misspelled or that words are in the wrong order, because our brain kind of anticipates and makes predictions based on what it's seen before. Oh wow, Michael, right. I've not heard any of these clips before.

And I'm quite intrigued to see how I perform, like how I how easy it is for me to tell what's real and what's fake. So why don't we take some of the clips from Kimberly's study, um, add our producer Adam can play them and we have to guess Like a game of real old rubbish, which one is fake and which one is real. Sounds great. Let's do it. 1964. Make several significant recommendations in this field. Do you wanna go first, Michael? I'm gonna say that was

Fake. Me too. Yeah. I felt convinced. Utterly convinced that was fake when I was listening to it. Adam. So that was clip 5A, and that was fake. Yay! Come on. Feeling feeling positive. Feeling good about us. Let's play clip seventeen B. Which carry the major responsibility for supplying information about potential threats. Oh that was tricky. That wasn't as clear cut at all, I don't think.

Um well could be either. I'm going to go for real. Yes, I'm gonna plump for real. I thought it was a little bit fake, but then overall the the impression I got was that was a real person. I'm gonna go for fake again. Drum roll. 17B was real. Well, you know, it's a very important skill as a journalist to be able to uh tell tell real news from fake news, isn't it? So seventeen B was real. Here's though what seventeen A sounded like, which is the fake version.

Which carry the major responsibility for supplying information about potential threats. And here's again the real version of it. Which carry the major responsibility for supplying information about potential threats. And it don't sound a million miles of there there was a slight like The beginning of supplying it sounded a bit different.

Yeah, I think that was very little difference then. Perhaps the real woman was a little bit smoother, like the way she said potential was a little bit robotic in the fake one. I don't know. Wow. That's that's incredible. Right. I mean clearly people and uh by people I'm including myself, we're gonna need some

help when it comes to detecting deep fakes. Well Dan, thankfully scientists are working on programs and software to help us separate fact from fiction. Deepfake detectors if you will, and they're using the exact same machine learning processes that are used to power the deepfakes.

Yeah, that's correct. So these um defect detectors are also machine learning algorithms. So what they do is they listen to thousands of examples of audio clips and they're asked to dis to classify whether something is real and fake. the the quirks or differences that make f fakes sound fake, which uh us as as humans aren't very good at doing. Amazing. So we're using the power of deep fakes

against themselves. So, you know, their greatest strength is also their greatest weakness. Sounds like some sort of exciting blockbuster trailer for a superhero movie. Yeah, attack of the deep fakes. Um but don't worry if you don't have a deep fake detector at home. Uh Ben has some good advice for you on how you can stay vigilant online. and tell what's real and what's not. I think the key is that we need to learn to be a bit more discerning about the media that we consume.

Um, you know, if I think back to when I was younger, um Photoshop was the tool that was used to fake things, they used to like trick people. Um and people came really good at it. So I think there were loads of fake photos being put out there where we had no clue whether or not they were real or not. But I think very quickly as a society, we learned instantly not to straight away trust a photo that we're looking at.

And, you know, even before the internet and social media, we used to get news delivered to our our doorsteps by a newspaper. Um but then the internet came around and all of a sudden there's so many unverified sources of information out there. So, you know, w we just had to get better and better at learning to not necessarily just trust the first thing that we

I think the problem at the moment is broadly we trust video because it's hard to fake. But as deep fakes become better and better, they're becoming you know, we we maybe need to learn that uh we can't just trust every video that we see.

And we need to think about where did this video come from? What newslet what um news outlet did it come from? Um, you know, was it just sent to you on WhatsApp or has it been shared lots of times on social media? So just thinking about where that information might come from and trying to verify that.

I think's really important. Very good advice there from Ben. Yes. And obviously one place where you can always trust what you see, read or hear is the Week Junior Science and Nature magazine and our podcast Mysteries of Science. Absolutely. Now I think it's time to get out our old friend, the Mysteriometer. Scientific scale from zero to a hundred with zero meaning we know nothing, and a hundred being we know everything there is.

to know. So I wonder where our experts think we are on this scale when it comes to deep faith. I think we're very clear on how this technology works and we... we know ultimately the end goal of this, if people keep developing deepfakes, they make it indistinguishable from real life. You know, we we see these things on T and we don't know if it's a real actor or a real person. So I think we're very clear on that.

I think where we're unclear is what we want out of it and what implications that might have. you know, how are cyber criminals going to make the most of this technology? Um and how do we deal with that? So I think that pushes that scale right right down to the middle. I would put us at around fifty. So I think the technology to detect deep fakes is already quite advanced and people

image generators and um chatbots and stuff like that, they've become they've come a long way in the past couple of years. I think a lot of work n still needs to be done on developing um better detectors Generalized better to unseat. unseen or unheard environments and there's still a lot of work to be done on that. Right, so fifty, fifty-fifty, slap bang in the middle. We know what deep fakes are and how they work.

But what the future of them is remains a mystery and it's a very exciting place to be. Yes, this technology is just starting to be used, so we don't know how it will change in the future. Will it be used for good? Will it be used for bad? We'd love to to hear from you. How do you think we should use deepfake technology?

To send us a voice note, head to funkidslive.com forward slash mysteries and hit the big red button. And assuming that you send us a real voice note and not a deep fake one, then your message could be heard in a future episode. And uh speaking of your messages, don't forget Get to join us in two weeks' time for our final episode of season six, where we'll be answering your space mysteries for World Space Week. Until then, stay curious.

Thanks for listening to this podcast, which is made by the same people that make The Week Junior magazine. You can get six free issues of The Week Junior or three issues of The Week Junior Science and Nature for£5. By heading to theweek junior dot co dot uk forward slash podcast.