Every breath you fake

From recorded future news and PRX, This is click here.

Artificial intelligence is supposed to help us understand the world, but what if it's starting somewhere else, with us? Not just what we click or where we go, but what we feel. AI can already plan your vacation, manage your calendar, even write thank you notes. But now, researchers are asking a more intimate question. Can a machine look at your face and What's going on inside your head? And maybe, more unsettling, could it be better at it than we are? From recorded future news.

This is Click Here, a podcast about the people making and breaking. This week, Karen Duffin follows a deceptively simple idea. Can a machine understand emotion? Субтитры сделал DimaTorzok Human judges are not really picking this up with their eyes, but the computer vision systems are it. That's after the break. Stay with us. I'm Karen Duffin and this is Click Here.

Every day we navigate our jobs, our relationships, getting safely home at night by doing something so instinctive, we probably don't even realize we're doing it at all. We assess people's body language. We decode furrowed brows. We measure half smiles. Trying to decipher what other people are feeling. Dr. Mark Frank has built an entire career on that. He's a professor of communications at the University at Buffalo, and he studies how emotion shows up on our faces, in our voices.

But his research didn't exactly start in a lab. When I was an undergraduate, uh, in order to make some money, I worked in a bar and I used to uh on the weekends to bounce. This was the 1980s. College bars were packed, testosterone thick in the air. It was Mark's job to figure out which guys were just loud and which ones were about to throw a punch. So reading people wasn't so much a hobby as it was a survival skill.

If you can see and you anticipate, you can see how this person's looking at that other person and that how this one's reacting and you can kind of intervene before trouble is afoot. And I thought I got really good at reading people. So good, he says, that the nights he worked the door, the number of fights went down. And what Mark learned scanning those crowds was simple. Bodies talk. And they often reveal truths that we're trying to hide, sometimes in the smallest of ways.

There were certain things like somebody coming around the corner if they were underage. There was always this little break in their stride, very t subtle but detectable. They told me, you know, better double check their ID, you know, all this other kind of thing. Mark took that instinct to graduate school, trying to answer this question. I'm really curious, like how good are we at reading people? And that led him to someone else who was interested in the same question.

And this is where I first come across the work of Paul Ekman. The man who put micro expressions on the map. If you've seen the TV series Lie to Me, it was influenced by Ekman's work. I would know if you were telling the truth, right? Ekman believed that emotions leave physical signatures on the face. So he built a system to measure them, muscle by muscle.

And by the time he was through, Ekman's map included ten thousand facial expressions. It became a tool for everyone from the FBI to therapists and even animators use it to draw emotions more realistically. After Mark graduated, he went to work with Paul Ekman in his lab. So I w well then worked with Paul Ekman for three years and just learned just so much. He's like the smartest guy I ever met.

Mark had developed an interest in deception, and this is one of the first things he and Paul Ekman studied together at a very basic level. They mapped the difference between a genuine smile and a fake one. And it turns out emotions have two components. The emotion itself, which is more of an internal experience. And then there's the outward expression of that emotion, a smile, a laugh, an eye roll, and the inward expression, the feeling, so to speak, that part is hard to control.

When you're feeling an emotion it comes from down in the limbic system and it travels through the extra pyramidal motor system, which works in a pulse, a ballistic like fashion. In other words, emotions fire through us automatically, almost like a bullet. Like say you remember something funny at a funeral. And it's inappropriate to laugh, but you feel that little you know like bulge coming onto your face and you got like suck your cheeks to try not to laugh.

But not every expression comes from that same deep, rapid emotion system. Sometimes we manufacture them. Like when someone asks you to take a more formal picture. Okay, look, I'm gonna smile one, two, three, right? That comes from the same part of your brain that causes you to raise your hand up or down or to give a thumbs up, things that you're doing deliberately. Mm-hmm.

Emotions are automatic. Expressions are often deliberate, which means you can manipulate an expression. So if you want to spot deception, look at those expressions. A fake smile, a forced frown. And what Mark and Paul Ekman found is that the difference between a fake smile and a real one is in something very subtle. Flow, specifically the flow when a smile reaches the eyes. A real smile they found unfolds smoothly in a predictable way.

So the flow, the dynamic qualities of the smiles, the duration tends to be more consistent, the onsets tend to be smooth, and so on. But when someone fakes a smile, that flow becomes slightly mechanical because the brain is assembling it piece by piece. And it turns out that humans are actually not that good at spotting when that flow is off. So Mark began to wonder would a machine be better at detecting that deception, at detecting when the flow is off?

About 10 years ago, he decided to test this, man versus machine. In the task he gave them, which was better at identifying real pain versus fake pain? The human judges were only about fifty five percent correct in telling which was real pain from fake pain. Well, the computer vision system became 85% correct. 85%. And this came down to that flow thing we talked about before. Because machines can spot the micro differences between a grimace that develops naturally versus mechanically.

Human judges don't are not really picking this up with their eyes, but the computer vision systems are able to pick that up. For about a decade now, Mark has been studying this distinction. When it comes to reading emotions, which parts are humans good at? And which parts might machines actually do better? And even trickier, could we teach machines to get better at this? It's something that feels so innately human.

When we come back, Mark puts that idea to the test, not in a lab, but in a classroom. We were trying to develop these computer vision systems to interact with children and to recognize when the children are upset, when they're attentive, when they're bored. Stay with us. Looking for more of the cybersecurity and intelligence coverage you get on Click Here? Then check out our sister publication, The Record, from Recorded Future News.

You'll get breaking cyber news from reporters in New York, Washington, London, and Kiev, among others. And you'll see for yourself why it attracts hundreds of thousands of page views every month. Just go to the record.media. It is called Internet. I usually Super highway. Security Why do things go viral?

Mark Frank started a project focused on children who have learning disabilities because there's a shortage of teachers trained to work with them. So Mark imagines AI acting almost like an extra teacher's aid. Spotting when a child is frustrated or when they quietly check out. Because, you know, if the kids aren't attentive, they're not gonna learn anything, right?

The system flags the distracted student, so the teacher can step in, and then the AI system gives the other kids something to do in the meantime. So creating systems that can interact with the child to facilitate their development, understanding, and so on and so forth is obviously I think a really positive thing. Though Mark's enthusiasm comes with some heavy caveats. For starters, he's worried about the data they're using to train the algorithm.

Well, this is one of the issues that if we've just sort of discovered. Like you read the papers trained on well-established data sets. You hear this phrase a lot. But those well-established datasets, often it's just a couple hundred images. And when you consider the vast range of human emotion, that is a very thin foundation. There's a lot of these data sets that were considered well established data sets.

We're not really normed on anything all that impressive. It might be a handful of graduate students just making their judgments. Okay, this is anger, this is fear, this is sadness. And that was a little bit frightening as we did a deep dive into it. And interpreting expressions is just very subjective. You know, a grimace to one person might read as anger to another. I mean just look at the captions of celebrity tabloid photos.

Here's celebrity X is upset at the camera person, right? Okay, well what exactly is upset? What emotion is that? What and sometimes you look at them go, well, no, that doesn't look like they're upset. In contrast, Paul Ekman's hand-coded system, it used very strict physical criteria, a mathematical computation that he called action units. For example, Ekman's fact system had, okay, with happiness is the action unit six plus twelve. With fear, it's got action unit twenty plus this plus this.

Not to mention, AI systems are fed in part at least on data from the internet, which is, of course, full of conflicting and deceptive material. So you can only imagine if if AI systems are trying to harvest the internet for these various things, it's going to be taken off in some really strange directions.

And even if you fixed that data problem, there's another challenge, content. Machines might recognize what an emotion is, but they can't tell you why someone feels it. You know, it's a two-step process. First you have to s recognize something, but then you have to interpret it. Okay, now was that the nervousness of the person who's a lying and afraid to get uh afraid of getting caught? Or is that the nervousness of the, you know, innocent person who's afraid of being disbelieved?

Machines don't know what happened five seconds earlier. And that missing context can make all the difference. I mean, just imagine a scenario like this. Please have. Your friend texts you at the last minute that they can't pick you up at the airport and so you're annoyed and you're waiting in a long customs line. And now you're standing there just frustrated. What if an AI system is scanning faces and reads that frustration as aggression and flags you as a potential threat?

And suddenly you're pulled out of line for questioning and facing possible detainment. You can see how that missing context becomes critical, especially as we start asking machines to make decisions based on the emotions they spot. And this is something that may not be hypothetical for long.

There's been a recent explosion in emotion reading technology, what's known as emotion AI, in sectors from healthcare to HR, like AI that analyzes emotional responses in people who have dementia and Alzheimer's. Researchers are building systems to decode their emotional cues when they can no longer communicate verbally. Over in human resources departments, AI is being developed that will scan applicants' faces during interviews.

trying to infer personality traits or suitability. Different sectors, all with high stakes. The difference between landing the job or not getting it, between faster medical care or care that lags. Mark also worries about what happens when our emotions become data, and that data becomes a commodity. Because that's what moves people, right? Emotions are about movement, you know, they're about behavioral intentions and what you're going to do. Companies already monetize our behavior.

And if data about what's happening inside us, our emotions, If that becomes a commodity too, Mark worries it will allow companies to press our buttons more effectively, not just make us shop more, but say make us angrier than we already are. And if you can start pushing the anger button, you know, the next thing you know, then, you know, people start to get hurt. Some companies, like the Spanish startup Neurologica, are already building systems to read emotions at scale.

Our technology is installed in soccer stadiums and airports and we can determine whether a crowd is interested, enthusiastic, bored, excited, and dynamically change advertising. In cybersecurity, your greatest fear isn't the threats you see, it's the critical signals lost in the noise. Every day, security teams sort through millions of potential threats. That's why Recorded Future exists, to give you precision and Our advanced AI detects patterns humans might miss.

While our threat intelligence experts, veterans of military and intelligence services, provide crucial context. With Recorded Future, you gain the confidence to identify critical threats and the precision to act before they become attacks. Learn by 1900 plus Rats faster and achieve three hundred fifty percent plus ROI within a year.

Given all of this, it's perhaps no surprise that regulators are wary of the technology. Particularly in the EU. They passed legislation in 2024 banning emotion AI at work and in schools, in part due to concerns it may exaggerate existing biases. Though the concern isn't just bias, which affects all AI, but something deeper. This technology is trying to interpret human feelings, something that's already hard for people to agree on.

And while AI is doing the interpretation, it's humans with all of their biases who teach the machines. AI makes the same mistakes that people do, and often in a more extreme fashion. Given all of Mark's reservations, it was hard to tell if he was enthusiastic about AI doing all of this or skeptical. Um, yeah, it's it's you know, each coin has two sides. And uh, you know, knives can be used to cut up a lovely meal or a human being.

And Mark is working hard to ensure that these systems fall on the helping, not hurting side of that coin. He's creating broader and more systematized data sets and focusing on efforts he thinks will truly benefit people, like those kids with learning disabilities. But here's the real key, Mark says. We have to get crystal clear about what humans are good at and what machines are. Humans have an inborn ability to read the room, so to speak, and we understand context.

AI excels at reading those tiny dynamic changes that help distinguish lies from truth. And the combination of those strengths could probably be helpful if it's done right. But Mark believes that machines alone can't be trusted. There is the risk that a lot of people are just gonna turn their brains off and just go with what the AI says. It was not designed to be a standalone system.

And Mark's biggest concern isn't just misinterpretation, it's something more existential. What happens if we stop exercising our emotional muscles, our empathy muscles? We've already seen the effects of letting machines carry our mental load. There are some recent studies, you know, kids using AI on papers and using things that turn their brains into these like souffle brains, I call it. You know, there's like nothing there. You poke it and pff it all goes away.

And Mark has seen early evidence that something similar happens with emotion and machines. There's a correlation between how much time you spend on social media and how good you are at reading subtle emotions. And the more time people are in these little social media things, the worse they are at reading emotions and just interacting with people face to face.

Evolutionarily, this is what we were designed for, right? We were designed for a face-to-face world where you interact with somebody and all five of your senses are engaged as you interact with people and so on. And so people want to sit back and just sort of let AI do this for them. There will be a cost to that.

We learn empathy, he says, through face-to-face interaction, by saying something hurtful and seeing its impact, or meeting someone we thought we'd hate and finding out actually we like them. and you have to have those experiences. Like exercise, you have to go to the gym. You gotta actually move something physically to build a muscle. Well you have to move your brain to do this.

Mark Frank started his career outside a bar, scanning faces for trouble. And now he's helping build machines that might do some of that work too. Reading faces, searching for signals and human emotion. But he's careful about what those machines should do. Because empathy, he says, isn't something you can just install. It's something you practice. Because recognizing emotion isn't the same as understanding it.

That was Karen Duffin. And maybe that's the line we hold on to here. Not whether machines can read us, but what do we lose if we stop reading each other? I'm Dina Temple Rastin, and this is Click Here. Looking for more of the cybersecurity and intelligence coverage you get on click here? Then check out our sister publication, The Record, from recorded future news.

You'll get breaking cyber news from reporters in New York, Washington, London, and Kiev, among others. And you'll see for yourself why it attracts hundreds of thousands of page views every month. Just go to the record.media.

Here's what you need to know about the tech world this week. It's Tuesday, April seventh. First, the federal government is stepping into a growing fight over a new kind of online betting. Mike Sealig runs the federal agency overseeing prediction markets. He says they're here to stay nationwide. Production markets are basically websites where people can bet on real world events, everything from sports to elections to military conflict.

Now, Illinois, Connecticut, and Arizona want to regulate these sites and limit how they operate within state lines. They argue they're basically unlicensed gambling. But last week, the Trump administration sued those states. And the agency that regulates such things, the Commodity Futures Trading Commission, says these platforms aren't gambling at all. They're financial tools, more like stock markets than casinos.

And they warned that if states make their own rules, it could open the door to fraud and weaker protection for users. Critics are convinced. Some lawmakers, especially Democrats, are now pushing to ban betting on things like elections and war altogether. So this is turning into a bigger question. Who gets to control this new kind of market? States or the federal government? And that fight could eventually land at the Supreme Court.

Next, Cambodia has extradited a man accused of helping power one of the world's largest online scam networks. A Cambodian and Chinese police joint task force arrested Chinese citizen Li Xiang. Li Xiong is the former chairman of the Wei Huan Group, a company that at its peak operated a cryptocurrency exchange, a bank, and an online marketplace. Researchers say that combination made it a kind of one-stop shop for criminal groups.

According to the US Treasury, the network helped move at least four billion dollars in illicit money, including funds linked to North Korean hackers. This all ties into a broader crackdown in Southeast Asia, where large-scale scam compounds have trafficked workers from around the world, forcing them to run investment schemes, sometimes called pig butchers. Authorities are trying to dismantle these networks, but analysts warn even when one is taken down, another often takes its place.

This is one of the biggest leaks in the AI world's history. This is not an April Fool's Day joke. And one of the biggest AI companies in the world is dealing with an embarrassing leak. Anthropic, maker of the AI assistant Claude, accidentally exposed some of the internal instructions behind its coding tool. The company says it wasn't a hack, just human error during a software release. No user data was exposed. But the leak did reveal something valuable, how anthropic guides its AI.

Those instructions, called a hardness, shape how the system behaves, what it does, and what it avoids. Now, competitors and developers have a clearer blueprint on how Claude works. And security experts say it could also give hackers new ways to probe for weaknesses. People digging through the code found some unusual features, including something Anthropic calls dreaming, where the AI periodically organizes its own memory.

and another mode where the system might go undercover and not identify itself as AI when posting code online. There was even a built-in digital pet named Buddy. Anthropic has since tried to pull the code offline. One cybersecurity expert called the leak embarrassing but not dangerous. And full disclosure here, Anthropic is a financial supporter of this program through paid advertising, but that support doesn't influence our editorial decisions.

And finally, a small but long awaited change from Google. Account holders will now be able to change their Gmail address name. In the past, if you wanted a new name, you'd have to start over with a brand new account. For years, if you didn't like your Gmail address, you were stuck with it. Now Google says U.S. users can finally change their email address without losing their inbox, contacts, or history.

It's a fix for a very specific kind of digital regret, like that username you picked in middle school that somehow followed you into adulthood. There are limits, you can only change your address a few times, and if you switch, you're locked into a new one for a year. Your old email won't disappear, and messages will still forward over. So yes, you can move on from smart guy 766, but you don't have to erase him completely. And for the record, my last name at Gmail seemed weird enough.

Click Here is a production of Recorded Future News and PRX. Today's show was written and produced by Megan Dietrich, Sean Powers, Erica Gaida, Zach Hirsch, and Casey Georgi. It was edited by Karen Duffin and Sarah Cavado, and fact checked by Darren Ancrum. Original music is by Ben Levingston, with additional music from Blue Dot Sessions. Our staff writer is Lucas Riley, our illustrator is Megan Goff, and our sound designers and engineers are Jake Cook and Jesse Neiswoner.

Find us on X or Facebook at Click Here Show. Leave us a voice message at sixty. CH talk. Sometimes we'll turn those moments into reporting, sometimes into a conversation. And sometimes into a future story you'll hear on the show. I'm Dina Temple Reston and thanks for listening.

Support for this program comes from Recorded Future. In cybersecurity, the biggest risk isn't what can be seen, it's what gets missed. Recorded Future analyzes billions of signals to help organizations stay in the Recorded future Know what matters, act first. Looking for more of the cybersecurity and intelligence coverage you get on Click Here? Then check out our sister publication, The Record, from Recorded Future News.

You'll get breaking cyber news from reporters in New York, Washington, London, and Kiev, among others. And you'll see for yourself why it attracts hundreds of thousands of page views every month. Just go to the record.