# Alex Todorov on "What Do We 'See' in a Face?" - Think Better Speaker Series ## Метаданные - **Канал:** The University of Chicago Booth School of Business - **YouTube:** https://www.youtube.com/watch?v=JHv5QGESAlw - **Дата:** 24.03.2026 - **Длительность:** 55:46 - **Просмотры:** 219 ## Описание On March 9, 2022, Chicago Booth's Alexander Todorov took attendees on a deep dive into the science (and pseudoscience) behind first impressions in a hybrid installment of the Think Better Speaker Series. Todorov explained what makes snap judgments about a person's character from just their appearance so compelling and misleading at the same time. Think Better speaker series: https://research.chicagobooth.edu/cdr/events/think-better Visit Mindworks: http://www.mindworkschicago.org Take paid studies online in the CDR Virtual Lab: http://bit.ly/CDR-LABS ## Содержание ### [0:00](https://www.youtube.com/watch?v=JHv5QGESAlw) Segment 1 (00:00 - 05:00) - [Nick Epley] So the purpose of this speaker series is to explore how insights in behavioral science affect society, shape policy, impact business, and also affect our everyday lives in ways that hopefully understanding will enable us to improve human life in some meaningful way. And tonight I'm delighted to welcome Alex Todorov, my friend, and one of our newest, maybe our newest, faculty member in the behavioral science group. Alex spent the last 20 or so years at Princeton, just a tad under, and he joined us in January of 2021, during that pandemic, space-time warp that both feels like forever ago and yesterday all at once. So we don't know if you've been here for a long time or a little time, but we're delighted to have you here, - [Alex] I don't know either. - with us in Chicago and I believe this is probably the first sort of biggish public event he's given since he's been here so we'll have to give him a warm welcome in just a moment. I've learned from Alex many times over, both from his research and from talking with him that what you think you see in someone often doesn't match the reality of that person. So for instance, in Alex, you might see someone who looks very smart, clearly academic, a cool academic at that, but there's a lot more to him than that. Turns out, Alex is also a wicked fast sprinter. You wouldn't have guessed that. So wicked fast that he is nationally ranked in the top 25 in his age group, in the United States. I won't tell you what age group that is, but if you don't like his talk tonight, you're not gonna catch him running after him to chase him down to tell them that. I also see someone, in Alex, I see someone looks like they were driven to be an academic given how well he's succeeded over the course, succeeded over the course of his career. But when I asked him, "Alex, how is it that you got here," here being at the University of Chicago? And he said, "A huge amount of luck, and randomness. " That is, it was not really an intentional thing, much at all. So Alex, Alex was in his second year in college, in Bulgaria, when the communist system there sort of broke up, and that allowed him to look elsewhere for places to continue his education. He got lucky and got an Open Society Fellowship that allowed him to go to Oxford at the same time that he got into a PhD program at the New School for Social Research in New York. He actually thought he should go to the New School first, instead of going to Oxford, and his advisor there told him that he was insane, that, "No no, you should go to Oxford," which he did, got him inspired in psychology, also convinced him that he needed to go to a slightly different PhD program. When he came back to New York, he benefited from a policy in New York of being able to attend classes at any school in New York, if you are, if you are in a graduate program. And one of them that got him connected to NYU, got him enrolled in the PhD program there, connected to a few advisors, and one lucky break after another happened. And now he's here, not necessarily intentional, but a lot of happy accidents brought him here. And then Alex, I also see just looking over the course of his research career, someone who seems very deeply focused on one particular, fundamental, problem in psychology or in social psychology, namely, how is it that the most basic of all social stimuli affects our minds, namely the human face? So I thought there must be some deep seated interest in the face, in understanding the face in Alex. And so I asked him about this yesterday. And he said, "No, actually not. " That early in his career, turns out he kind of stumbled on a really big finding, which he's gonna describe a little bit to you tonight. Turns out that a person's appearance affects and predicts election outcomes, that was surprising. And in talking with his colleagues at Princeton, one of them Nobel Prize winner, Danny Kahneman suggested to him that, "Hey, you probably wanna start a research program on something, rather than doing somewhat scattered research, and this face stuff seems kind of interesting. " And so that's exactly what Alex did. And one thing led to another, and after 20 years, Alex's work is really fundamentally changed the way we as psychologists, understand what people see when they look in the face of another person. And one of Alex's most fundamental insights, at least interesting insights from my perspective, is that there's a lot more to a person's face than you actually see when you look at them. In fact, what we see in somebody often doesn't reflect reality much at all. And I think Alex might even argue to us a little bit tonight that it almost never reflects the truth about the person. So with that, please join me in a hardy welcome of Alex, not just to this event here tonight, but to the University of Chicago itself. (applause) ### [5:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=300s) Segment 2 (05:00 - 10:00) - [Alex Todorov] Thank you, Nick. Oops. Thank you, Nick. I guess this was the most interesting introduction I ever had so, so what do we see in a face? Well, there's actually more than hundred years of good research on faces in cognitive neuroscience, in cognitive science, in social psychology, and faces are really special. So I'll just give you a couple of reasons. One, newborns with virtually no visual experience, prefer to look at faces than at equally typically complex objects. Two, primate brains contain neurons, and that selectively respond to faces, and these neurons are usually concentrated in a few selective regions in the inferior temporal cortex, so we, our brains are really specialized to process faces. We are face experts, from the moment we are born we are exposed to thousands of faces so very subtle differences in, in facial appearance we can detect it. So it's easy to discriminate between two faces. And we form impressions from extremely little information. So let me give you a demo. So I want you to fixate on the circle and then a face will appear very briefly on the order 50 to a hundred milliseconds, and let's see what you would see. So if I have to make a guess, was it's a man or a woman? - [Audience] Woman. - Woman. A race? - [Audience] Asian, - Asian. Emotion? - [Audience] Smiling. Exactly. So very rapid presentation, and yet you can extract this demographic information from it. So impressions from facial appearance extremely efficient. Typically about hundred millisecond exposure to the face is sufficient to read off demographic attributes, like race, age, things like emotional states, cognitive states like effort and exhaustion, focus of attention, eye gaze. It's not that you make this decision within a hundred milliseconds, that's the time it takes the information to get from your retina to the parts of the brain that process faces, but within half a second, you're basically done. And this is the case even if the signals are very subtle. So if I ask you, "Who is the male or who is the female here? " This is the gender illusion. How many people think that the face on your left is the male face? So there's a clear consensus. It's actually, it's exactly the same face in every respect accept that the skin is made a little bit darker of the male face. That's the gender illusion. As it happened in every culture that people have measured men within the same ethnicity, men tend to be darker than women. And you can look at the palm of the hands so it's not just about sun exposure or something maybe men just work outside. So there's just a, this seems to be a stable, biological difference. This impressions also could be fairly surprising. It turns out that information from the face could be even revealing about underlying health in the case of the elderly, that is how old one looks relative to their biological age. So if you look at this two faces, this was a very large study done on twins in Denmark, I believe over 7,000, and what they did they look at 70 years old people and they tracked them for a very long time, basically looking how long they're going to live. Now you see more of the twin that looked younger, you should be able to detect on the left side, and the twin that looked older. It turns out that next to the best predictor, especially when you old, how long you're going to live is obviously biological age. Turns out that the second next best is not a DNA test or anything like that but how old you look relative to your biological age. So people who looked younger actually lived longer. And at first sight, it's actually quite amazing. How is this possible? What is happening? But it's not really that surprising if you start digging inside behind the data, what makes you look younger? Well, the face after all is part of the body, right? So when you're aging, if your body is aging, your face is aging too. The first reason is genetic luck. There's nothing you can do about this. The second though, is socioeconomic status and access to healthcare. If you're wealthier, you have better access to healthcare you're going to look better. Lack of chronic conditions, like asthma. ### [10:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=600s) Segment 3 (10:00 - 15:00) Chronic sun exposure, not like vacations outside, but being a farmer or a fisherman. That's also very bad. Smoking turns out it's terrible for your appearance, among other things. So this, all of the things that will make you look older, and so you can think of these accumulated advantages and disadvantages in life, they do get imprinted on our bodies, and on our faces, which after all, are part of our bodies. But what a lot of my work has been on a set of inferences that are much more problematic. It turns out that again, after extremely brief exposure, people could make all sorts of inferences of complex and presumably stable personality attributes; like whether the person is trustworthy, competent, aggressive, and so on, and so on. In fact, these are the kinds of inferences who have a very, very long history of historic fascination. Nobody cared about the brain processing, the face processing mechanisms in the brain, but people always cared, at least since we started living in modern states, about can you read the personality of people from their faces? In fact, this is what I call the physiognomist promise. Physiognomy is the pseudoscience of reading character from faces. And the main promise is that you can actually infer others' character from their appearance. The first teacher in physiognomy is actually attributed to Aristotle, so it goes to ancient Greece, but it really, really picked up in, in the 18th and the 19th century. This was the cover of maybe the first Swiss Best Seller by Johann Gaspard Lavater, who was a clergy in Zurich and was extremely influential and prolific writer. He wrote these four volumes, essays on physiognomy, some of the best illustrators worked on the books. None other than Goethe helped him with the first edition. So it was a really very popular and highly influential book in Europe. Just to give you one brief anecdote, Darwin was almost denied the chance to take the historic Beagle voyage on account of his nose. And he can entertain his nose. Why? Because the captain of Beagle happened to be a fan of Lavater and he didn't believe that the person with such a nose would possess "sufficient energy and determination" for the voyage. Luckily for science Darwin made it. And he wrote in his autobiography, "but I think he was afterwards well satisfied that my nose had spoken falsely. " And that much more serious actually effect by the way, noses for were some reasons were huge in the 18 and 19th century, there was a whole branch of physiognomy called nosology, and there are whole books dedicated to reading character of people from their noses. So in some way, this is all funny and it's a ridiculous notion, but the physiognomist promise is alive and well. Primarily because we are all naive physiognomist's forming and acting on impressions from faces, forming instantaneous impressions and acting on them. So let me give you a few demonstrations. Imagine that you're walking into a party and this are the first two people you're going to see. How many people will approach the face on your left? Okay. Clear majority, right? You, you can change it in a hiring context and say, imagine you're hiring a salesperson; who would you hire would be the same decision. Now, all of the faces that I'm showing you were not created by an artist: they're generated by a mathematical model that visualizes our impressions of different traits. In this particular case, this is our impressions of extroversion and introversion. So the, you have a prototypical extroverted face on your left, prototypical introverted face. If you're going to a party and your objective is to have fun, this is the person you would likely approach. You might be wrong for all you know, but not knowing anything else, that's what you are going to do. Well, who would make a better CEO, the person on the left or right? How many people think the person on the, your right will make a better CEO? Again, so it's a clear consensus, maybe not as much as the first one, this is a model of perceived competence. And so you make the face look more competent and it has a, a resulting effect on decisions. You can apply to real faces; I can tell you this, the twins, the evil and the good twin, who is the evil twin, and chances say, of course, the face on your left would be the evil twin, or you can make it positive and say, if you're hiring a security guide, a guard, who would you hire? ### [15:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=900s) Segment 4 (15:00 - 20:00) And in this case, we just put a little bit of masculinity in the face, and once you introduce it comes with all of this other stereotypes; the person is physically stronger, aggressive, and so on, and so on. And now to your cookies, this is who is more trust worker, is hiring a babysitter, who would you hire? Again it would be, I am not even going to ask, but obviously would think the person on your right, this is from a new model. One of my post-doc was a leading figure, Stefan, who is sitting here in the first row, and I'll tell you about this model in a second. So what is important is that this impression from face is immediate and compelling. You don't have to, you don't stop to think for a minute and say, well, let me compare the eyebrows, different facial angle and come up with a decision. Let's just to have an instantaneous decision. In fact, we can measure this. I already show earlier demonstration with demographic characteristics, but you get exactly the same face with this complex social judgment. So I can show you a face, then view something called perceptual masking, because you want this to interfere with your perceptual processing. So really know that the stimulus is presented for the time you intend to present it. And the basic finding is that it can start with, from subliminal exposure when nothing happens, people at chance, you go to 30 milliseconds, people starting agreeing with their judgment if they have unlimited time, and by the time it's 160 hundred and 170 milliseconds they're done, they don't need more time. This is kind of literally single glance impressions because that is the time it takes to move your eyes. And eye movements is the way how we collect information in the world. The second thing is that the impressions from faces are consequential. So again, I'll ask you to focus on the next slide. Pay attention, who is more competent? Again, if you do it in a experimental context that's enough exposure for people to make a judgment, This were actually gubernatorial candidates, I believe in race in Virginia in early 2000. And in a series of studies, Nick mentioned this study, the first study that I stumbled upon and made me truly interested in studying faces, we simply had actually naive judgment of Princeton students looking at faces of politicians. They didn't even know they were looking at politicians. Actually think this was one of the elections was when Barack Obama was running for the US Senate. So we removed pictures of familiar, famous people. And it turned out that just using these judgments, we can predict about 70% of the election outcomes. Works just as well, just enough for gubernatorial elections, too. Political scientists have followed up on this work and now we know that the effect is really driven by people who know next to nothing about politics, and who watch TV, which makes a lot of sense from psychological point of view. We talk about heuristic processing. So they are substituting a complex decision, complex question, how competent is the candidate, with a very easy question, how competent they look? So in a sense, they're looking for the right information in the wrong place, because it's easy. And that's for you, the story of why we have psychological heuristic in decision making. There are many other findings. These are not from my lab group, from other groups. Competent-looking CEOs are more likely to get high compensation packages. They're not necessarily better CEOs, but they get high compensation packages. Trustworthy-looking borrowers are more likely to get loans and with lower interest rates, this was done on a website, I believe prospect. com, but the loan's not trivial up to $25,000. And they have the credit card, the credit history of the person, they have history of the previous defaults, so lots of information, good information that should lead you to better prediction. And yet putting a picture was good. In fact, people who had a picture made them more likely to get a loan. So having a picture is better than not having a picture. And then having the right picture is the best. So these people got more likely to get loans and with low interest rates. Work that is much older, dominant-looking cadets are more likely to achieve high military ranks and the list goes on and on. And there are lots of experimental demonstration, this sort of effects in the real world, but you can think of the first demonstration. If you run a study, you can see that there's a very clear consensus what the judgments are going to be. And lately, this is the part that is actually a bit, well, not a bit, quite worrisome. You have lots of new technological startups promising profiling personalities based on facial images alone ### [20:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=1200s) Segment 5 (20:00 - 25:00) you have scientific publications claiming that they can read from facial images alone, criminal inclination, sexual orientation, political orientation and so on from. So I will address this at the end of my talk. So physiognomy is here; has never gone away, just this new clothes, the Emperor's new clothes. And yet the promise is deeply flawed and disturbing in its implications. Why is that? Well first, why is the physiognomist promise so appealing? Well, it promises an easy solution to a complex problem. And the complex problem is how to figure out the intentions and capabilities of other people, of strangers. That's, you're stuck with strangers since the moment, I think Jared Diamond had a great line that once we started living in modern states, maybe 13,000 years ago or something along this line, it was the first time in human history when you when encountering strangers, you did not have to kill them. You know, before the decision was easy. But now it's difficult. Like, "What do I do with this person? " It's this is interesting in, in historical context. So in early 20th century, Katherine Blackford and Arthur Newcomb had a bunch of books that were super influential. And they were the modern physiognomy of early 20th century. So on the arguments were not based on Lavater's really flimsy logic, but it was based on evolutionary theory. Of course, that was also very flimsy, but nonetheless, the arguments were different. And in fact, Katherine Blackwell had some great ideas you can think of her as a precursor of the modern human resources departments. So the idea here is how to solve the problem of finding the right person for the job. And she had specific ideas, forget about, just look at them, and you can do this character analysis based on appearance, from the interview, whatever they say, it doesn't matter. Letter of recommendations? They're worthless. That's not true. They're not high predictive. They're much better predicted than structured interviews. And so this had quite a bit of influence in the early 20th century. And this was exactly the time where psychologists were working saying, "Wait a minute, there's not that much accuracy in this lay judgment. How did you come up with this? " Well, this is from modern work. So here, this is a scatter plot, to illustrate from a, this is a meta-analysis of thousands of studies looking at the relationship between impressions from unstructured interviews. Like there's a difference from a structured interview, which has been validated, and maps into specific abilities, and unstructured interviews. The act, the correlation is positive by the way, it's about. 15, but it's, I mean, human eye start seeing correlations when they're in the order of point. 60. So there is a signal there, but it's extremely weak. It's like four times weaker than a letter of recommendation because letter of recommendations, they just have a larger sample of observations. So I think this is quite illuminating. I just generated to see what exactly this means. If you ask people to estimate what's the correlations between their impressions and how the person will do on the job, it's about. 60, and that's what people like but I think, and Lerov called the interview illusion, thinking that I actually, you learn a lot about the person. You learn nothing basically. Here's another thing that if you haven't read the "Moneyball," that's a fantastic book, or you can also watch the movie with Brad Pitt, as, obviously, Billy Bean, but the book is much more illuminating. Why is this interesting? Well, the whole book is about the unlikely success of Billy Beane, who is the general manager for Oakland A's baseball team. By the way, I know nothing about baseball and still it's fascinating book, but why he's been so successful, he will work with much smaller revenues than the other teams, and yet he would be consistently one of the best teams. And what he was doing, he was exploiting the prejudices of appearance. So he would find players who were under their market value because they didn't look the part. So here is quotes from Michael Lewis book: "Billy Beane was looking for those guys who for their whole career had seen their accomplishments understood with an asterisk. The footnote at the bottom of the page said, he'll never go anywhere because he doesn't look like a big league ball player, young men who have failed the first test of looking good in uniform. " I mean the, actually the scouts, they had the pro with the good face, and Billy Beane had the good face. Turned out that he was a failure as a professional player and he took this lesson and then he decided to move to management ### [25:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=1500s) Segment 6 (25:00 - 30:00) and people were all surprised. And then he was spectacular, a fascinating character. And another thing from Michael Lewis, "What begins as a failure of the imagination ends as a marketing efficiency: when you rule out an entire class of people from doing a job simply by their appearance, you're less likely to find the best person for the job. " So things have turned quite a bit if you look at the data from the arguments of, in Newcomb and Blackford. Generally predicting human behavior, it's a very difficult business. So when you're thinking about this, there's just to put you a few things in your mind. The typical correlations between a validated personality measure, that is a questionnaire like extraversion, or neurosis, or depression, there are lots of validated que, they are psy, they have good psychometric properties. And yet the correlation between these measures and specific behaviors, it's on the order of. 30. Okay? So that is very low. The correlations between a general attitude, how much I care about politics, and specific behavior, like will they go and vote in the midterm election, it's about. 20. The correlations between two behaviors that presumably map the same traits, for example, cheating on a test and lying to your parents, the correlation is. 20. Okay? So why would one think that you would get a single glance at a face and you get better prediction from things that actually, you know, like behavior, the behaviors there, lots of observations, thousands of observations, personality measures that are validated and so on and so on. Well, the second reason why the physiognomist promise is so appealing is because it actually does agree with our intuitions. We have this rapid efficient impressions and they seem to be shared with others. And, and here, this is important caveat; if it's others like us, is anything that you visualize would be based on the population of strangers. And so there would be differences between members of even different families, different friends group, different ethnicities, different cultures, and so on and so on. There would be some agreement, but there would be always more agreement within the people who know each other. In fact for the intuitions and a lot of the work that we've done, and we did this about 15 years ago, we can actually visualize these impressions. So all of the faces that I showed you in the beginning, they were generated by different statistical models. And the logic is very simple, but you can think that each face is just a set of numbers. We can randomly generate lots of thousands of different faces, and then we can throw them at people, ask them to rate them on anything that you like: attractiveness, trustworthiness, and then we can build a statistical model that captures what's the variation. Is there something systematic in the original random variation in the faces? This is what we call in psychology data driven analysis. That is, we don't manipulate anything a priori. I don't say, well, let me manipulate the smiling, the distance between the eyes, none of it. We just let the features vary randomly. And this is what you get. So this is the model of trustworthiness. And the score on, on the running score would be the perceived trustworthiness, according to the model. Very trustworthy, six standard deviations. And now it becomes untrustworthy. We always validate these models but one thing that I love about this research is at the moment you see the stimuli, you know that it works, right?. We always generate now novel faces. We manipulate them based on these models. We give it to new participants and make sure that in fact, the model predictions match into human judgments. And they do now, what did you see there? I mean, once when the face becomes more trustworthy, it starts smiling. We actually made absolutely sure that the face were emotionally neutral, yet very subtle difference in the face that suggests smiling it's picked up. So the model is like an amplifier, right? You pick up whatever signal is, and then you turn up the volume and you see, oh, oops, emotional expression's emerged. The face becomes more feminine, the trustworthy face. It's not a very strong signal. And it goes with the gender stereotypes. We generally trust women more than men. Okay? So here's one of my favorite models. The extroversion, and the face becomes really extroverted. So happy to see you. And now it's become introverted. As I said, this was a, this was a work that we've done, a while ago and way before the huge advances in computer science in the last 10 years ### [30:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=1800s) Segment 7 (30:00 - 35:00) what people call deep neural nets. And this is the work that Stefan has been heavily involved. So all of these faces here are not real. They might look real. I'm sure you probably have seen it, like every other month there's article in the Times, or another publication of hyper-realistic faces. So this was generated from a model that has been trained on 70,000 faces. And so you can generate faces that are hyper-realistic. They might resemble to a face that you know, but they're not real. And we can manipul, generate as many as we want. And here they're simply ordered based on their perceived trustworthiness on the top, kid's are perceived as way more trustworthy, and then smiling people with direct gaze, they are perceived as trust and the least trustworthy, usually men, they don't have happy expressions, they usually there's no direct eye gaze and so on and so on. So this is the basic task: we generate random faces, we ask people to rate the faces, and then again, each face is just a set of numbers, we use the same conceptual logic, and then we built the model. And here's the model of perceived trustworthiness. So we can apply to novel faces and, becomes a little bit younger, expression, of course changes, and you can do this for everything. So here's the, just another example. This is the least and the most outgoing faces. Again, these are randomly generated hyper-realistic faces. And then we can apply this to a new face and you can see expression changes of course, the eye gaze becomes direct. Notice it's not just the face; the sunglasses become much brighter and here's your model of outgoing. And we can do this forever. Like it works beautifully for skinny, or not skinny, for age, for of these attributes, but it works also for these sorts of social attributes. Now, what is really, really important is to realize what the models are doing. So the models capture systematic biases and judgments. They visualize our stereotypes. They're not capturing cause that's what an extroverted person looks like, that's what a trustworthy person looks like. I'll elaborate on the argument as I go along. They visualize stereotypes of a particular group. So all of this models is the primary, there was, I mean, originally a lot of this was done in Princeton and then the rest was done in (indistinct) so the vast majority are Caucasian participants. In fact, we had judgment, looks like me, familiar and they all cluster together, right? So it is specific to the group that generates the judgments. They do not capture personality characteristics. And this is the difference in the work that we do and the physiognomist's work. In the physiognomist's promise the, there's this deep assumption, that confuses the impression, your cognitive and emotional response with the real thing. It also assumes that there's one-on-one mapping from faces to facial images. Now, if you just think about your pictures, if you're not fully narcissistic, there are pictures that you like and pictures that you don't like, right? So it just takes a thirty secondth of reflection that images don't represent us in the same way. That's why we are always like to have great photographers who can make us look much better than we actually look most of the time, right? And as a result, it underestimates the huge effects of image variations in impressions. So let me, before I give it this demonstration. So why, so, but why is it, why do we form this? What's the source of this first impressions? Where they coming from? Well, they're grounded in momentary mental states. So you saw emotional signals emerging very clearly in the different models. They're grounded in stereotypes like masculinity, femininity, age. They're heavily dependent on a specific and the image variation. So they're constructed from cues that have some significance in the immediate situation. Let me tell about voluntary mental states. I think this study was done in Sweden. I just really like the study. I haven't done it, but, obviously the same person, right? Now if you show these two pictures to two different groups of people and you ask them to rate the person, is he more attractive? So the person, I think it's on your right, right? People think, "Well, he's not very attractive. He's not very smart, you know, he's kind of like, I don't know, there's something with him. " Do you know what the difference is? ### [35:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=2100s) Segment 8 (35:00 - 40:00) - [Student] The weight there, this one is like a little bit chubbier, I mean, he's not chubby, but like sleep deprivation. - Sleep deprivation. So same person. You go to the lab, take your picture while you have slept very well then 24 hours, you don't sleep. And then your picture's taken under the same standardized conditions. That's the difference. And that's a momentary mental state. I mean, the guy is exhausted. You can see. And when, and in fact, when we are sleep deprived, we don't look very good. Moreover, I mean, we are also not very smart and we do stupid things when we are sleep deprived. So know that this inference, not very smart, it's actually accurate, for the momentary state here now. And that's the reason you don't wanna be sleep deprived or exhausted when you're engaging in a complex task. So the inference is fine, but saying that, "Well, that's what the person is like," that's a huge inferential leap. It's the same person, just two different mental states. So it is fine in the immediate situation, but it's not warranted as an inference about the intelligence of the person. You have overgeneralization from stereotypes. Baby-faced individuals, there was a lot of work. Leslie Zebrowitz from Brandeis University has done lots of, lots of work, way before I was in this field, showing that people make all kinds of inferences based on, of, impressions of adults whose faces some more baby-face like or not. And as one example, they're perceived as naive and not physically tough. She has done some research with longitudinal research with teenagers, which suggest otherwise, presumably they're working hard to overcome the stereotypes so they tend to be smarter than their peers, and more likely to get into trouble. Individuals with feminine faces are perceived as more trustworthy, that comes from our gender stereotypes, individuals with masculine face are perceived as more competent. As it happens we can, using power models we can really tease apart what's the competent stereotype. One is attractiveness, which benefits women on average, but you can actually remove, literally subtract the attractiveness and see what's left, and its masculinity. And in some ways, it is a sad fact but it's not surprising from a statistical learning perspective because if you're exposed to more and more male leaders, that's where the stereotype comes from. Even worse actually, we can now we can sort of inject masculinity in real faces. And when you're making a male face more and more masculine and perceived as more competent. And if you do this for a female face, at some point, it breaks down, because of the gender stereotypes, the gender stereotypes and they tend to be much more prescriptive for women than men. It's like, you don't, that the stereotype is going against dominant-looking, masculine women. So that's a characteristic associated with leaders, but if it's too much in a woman, you get a negative response. The second, so this is where it's all, this is like lots of findings showing you how you have this specific inputs that influence your impressions. The other thing that is really important to know, and it's not immediately obvious that image variation, even if it's random, can generate completely different impressions of the same person. We are oblivious to this fact because image variation relatively unimportant for familiar faces. I walked in and the person who was at the reception and said, "Oh, you're coming for the conference? " The speakers, like, "Oh yeah, but that's, I didn't recognize" like, of course not, you don't know me. Next time you might recognize me, but it's recognition is difficult. And first impressions about unfamiliar faces. So if I ask you, who do you like more the, from the two women on the top side probably will guess the woman on the left. How about on the lower side? Well, probably the woman on the right side. Well, actually this is the same woman, this row, and the first column is, is another woman. Now notice that if you are familiar with this two individuals, would be totally trivial for you and you'll go with the person you like, because of what they like you and what you know about them. But for us, we know nothing, so you end up making this judgment. So there's just image variation. This was a study done by actually the Cannon Lab, the photography lab in Australia, they brought five professional photographers and they meet Michael, who is an actor, and each photographer is given a false piece of information. That's like a social psychology experiment right now. ### [40:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=2400s) Segment 9 (40:00 - 45:00) And so some are told Michael is self-made millionaire. Some are told he's a ex-convict, he's a psychic. And then they spend some time to draw the essence of Michael. And then they go and the photography is beautiful and totally different. And they're like, "What, what, how did you come up with this one? " And then it all makes sense because who is the self-made millionaire here? (general laughter) And who is the ex-convict, right? It's the same person basing on these false preconceptions you end up with very different images. If I ask you, "How many individuals are shown here? " Eh, it would be more than three. The modal guess is five. It's actually two. Why is this so difficult? Because you don't know them. Now, this guy, if I have to form an impression of this guy, I don't want to hang out with him. He looks like a skin head. It is the same guy. He's actually very talented social psychologist, done beautiful work on familiarity. He's in the "Best Psychologists in the UK. " Notice that this is totally different contrasted with familiar faces. There is just as much variation in this images. Even more like length of the hair, facial hair, age, yet, if you know who Brad Pitt and Leonardo DiCaprio, boom, it's instantaneous. And not only that, but it's not only that you recognize them, but also the recognition comes with knowledge, the movies you've seen them, the attitudes, political; and this what whether you like them or not, it really unlocks your memory. But that's the case when you know them. When you don't know anything about the person, the images don't provide any magic path to what the person is like. It is an illusion based on our experience with familiar faces. And in the case of unfamiliar images, it just, they can unlock only our stereotypes. Okay. I'm rushing a little bit here because I'm running out of time. Let me tell you a little bit about the new applications, machine learning applications. So here's one, I'll just focus on once specific study and this study, the, you, the algorithm is trained on images of people on online dating websites. And in fact predicts much better than chance the sexual orientation of the users. Let me say that I have no problems with the, the defining history of we do leave digital trace of power, personality, and preferences, and this could be detected by powerful algorithms. And that's a fact, right? Now the problem is that you have a confound, what we called in experimental psychology, the images are self-posted, right? They're selected. They're not people. The users didn't randomly select an image and like (makes a puffing sound) that's my, you know, dating website image, right? You presumably think carefully what to select. So the detection is not about honest facial signal of sexual orientation. The other problem with this algorithm is it's like a big black box. So you get a decision that is better than chance, and then you're trying to figure out, but what's in the box, right? And so what is it exactly that you're picking up? And so they, you're trying to double guess. So she creating a composite image of your heterosexual male and female faces, and also of the straight and gay faces, male and women. Okay. So what can you see? Well there's clearly differences in grooming, right? Male faces, shaven, unshaven, female faces, makeup, eyeliner, but it also appears that there are morphological differences, chin, nose, forehead. And that's that presumably may reflect differential hormone exposure. That was the, that was the interpretation in that paper. Now notice like the straight guy has a much bigger chin and smaller forehead. And that's me, if you can recognize me by now. (general laughter) So we do signal many things about ourselves, right? And some of this signaling is quite deliberate, but some is fairly implicit. You just really want specific aspirations. You wanna belong to particular groups. That's reflected in how you dress. That's reflected in lots of other things that you might not even be aware. A powerful A I algorithm can capture these cultural signals, no problems. I mean, they're there to be captured. But this is very different from saying, "Well, this actually there is a biological signal. There's like a direct mapping from something, you know, the broadness of your chin in your particular preferences. " That is where there's no, no strong data to suggest this, at least so far. ### [45:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=2700s) Segment 10 (45:00 - 50:00) So we do rapidly infer from faces things that most of the time are not very controversial, like age, gender. But in what we do, we construct this personality impressions from this first-order inferences. You can think of this basic mental states, emotional states, but even the first-order preference inaccurate always overestimating the role of the face. Let me just show you one more study. So these are tennis players after they just won or lost a point. (general laughter) Now, if I ask you, "Who won and who lost? " We know people are totally at chance. In fact, I have to put it in my notes because I never remember, but one, four, and six lost, and two, three, and five won. (general laughter) Now just the face by itself, doesn't give you any information. You know there's high arousal, something, and this, you get with pain, pleasure, all of this stuff, you know, there's high arousal, but you don't know what the direction of the emotion is. The body actually completely determines the accurate inference. We never did the careful analysis, but when you lose usually you're like this and when you win, you're like that. And, but once having the face and the body doesn't help you much. If you just show the body, that does the trick. Now, though, this is interesting. When you show the intact images of the people in experiments, they say, "Well, I think the right guess comes from the face. " Why? Because that's where you pay attention. The disambiguation is very rapid and you immediately know what's happening, but we know it's not in the face, know that we don't conduct this kind of experiments in everyday life, right? You really need to break down the stimuli to figure out what's the informational signal. And there's no way to learn. There's just, unless you do this sort of a experiment. So why do we have all of this? I think I'll be on time. Well, in the absence of good information, which is whenever you interact with strangers, and you don't have any reliable information, we rely on first impressions to infer the intentions, their capabilities. And we rely on lots of things. It's not just facial appearance, expressions, accent, accents are big, I mean, there are lots of great people working in the psychology department on the role of accents and perceptions, voice, body language. You get exactly the same first impressions from voice, even from single words, and single sentences, the same way you get this for faces, attire, context of the situation, and so on and so on. And all of this is integrated very rapidly. In a sense so it is the way I like to think of the functional role of first impression is a shortcut of dealing with the complexity of living with strangers. Some of these impressions might be accurate here and now. If somebody is disgruntled, they're probably not going to help you. And, and somebody is kind of happy looking, maybe on average, they're more likely to help you. But inaccurate, but it is inaccurate as guide about characteristics that are stable over time and situations. And with this, I will conclude with Georg Christoph Lichtenberg, who was probably one of the most interesting and least known 18th century thinkers. He was the first chair of experimental physics in Germany, posthumously credited with introducing the aphorism in German language, a fascinating character. And the reason he's very interesting to me, he wrote, and I said that completely took apart Lavater's ideas and I said physiognomy and you'll see why he's credited with his writing abilities. As he put it, "First impressions lend the smallest possible knowledge the greatest possible appearance of it. Consider someone wise who acts wisely, and do not be misled by regularities on the surface. " And with this, thank you. I'm happy to take questions. (applause) - [Mark Temelko] Thank you, Alex. We have some microphones for people who have questions, but there is, there are some questions from our viewers online, and I just wanted to start with one of those, "How stable are first impressions? " - [Todorov] So this is another thing that it's actually very easy to change your impressions, if you have the opportunity to interact with the person and observe them. But if you don't, there's a problem, because what's problematic with first impressions is that they kind of preclude opportunities, right? I decide I'm not gonna interact with this person. I'm not gonna get into economic transaction because I don't trust them. That's a problem, if in fact that's a potential cooperative partner, right? They just preclude an opportunity, will never learn. So I think this is their largest effect. ### [50:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=3000s) Segment 11 (50:00 - 55:00) But if it happens, so if you do an experimental game where you actually measure actual cooperation and you do repeated interactions, people very quickly figure out, "Oh, okay, this person is, might look trustworthy, but they don't behave trustworthy," and they change their behavior accordingly. - [Audience Member] So, so following up on that, how stable are these sorts of impressions in different cultures? For instance, in Northern Europe, where people are really used to women leaders and prime ministers, do they interpret a woman's competence differently? - [AT] So this is very, I mean, culture is hugely important. And if certainly in Northern Europe, you have a way high political representation of women. And so you get somewhat different effects what's predicted in different elections and different cultures. I mean, it's also, we have done one study. This is somewhat different, but you're absolutely right. I mean, stereotypes change, when reality changes. Often it's not the case that attitudes first change and behavior follows, typically behavior change and attitudes follow. So if you have a higher women representation, all of this starts changing, but we've done a study for example, with Israelis and Japanese young women. And then you can create a more from very typical Japanese to typical Israeli. And you ask them to say, "Well, how trustworthy is the face? " And guess what? The more Israeli it becomes, the more trustworthy he looks to Israelis, the more Japanese it becomes, the more trustworthy for Japanese. And it's not the case for attractiveness. Attractiveness actually, the mixed ethnicity faces have a little bit of a advantage. So, so culture is definitely important in what you expose. I mean, one way, always good to think about how do you end up with specific preference in faces is to think of faces that they, your statistical diet of faces, what kinds of faces you're exposed in your everyday life? And that would change what's typical. And we typical, we like things that are typical with trusting and so on and so on. - [Audience Member] Yeah, there's been a lot of news lately about facial recognition technology, especially in societies like China. And I'm just curious what impact or connection there is between your research and what's going on there. - [Todorov] Yeah. So I've been following some of this. I mean, it is related and it's very interesting because, you know, like people who have thought into the psychology probably was you have thought this, you know, but we were 10, 15 years ago, we would say like computers are much better than humans in serial, you know, but pattern matching, forget it. Look at faces, and then that's all history. So as a general rule, these algorithms are more powerful than humans, but then there's lots of training biases. And this, there's two things that make an algorithm good. One is the architecture of the algorithm and the other is the training space. And there have always been competitions, full competitions best face recognition algorithm. One year, the best one from Europe and from Asia. And again, proprietary information, you're trying to guess what they're doing. Turns out the Asian algorithms do better on Asian faces. European European faces. Why? Because of the bias in the training data set. And often with minority recognition you have all kinds of problems. And there are lots of, they can do better on average, but there's some difficult decisions where this even the best algorithms fail. So it's always people, the recommendation this is just another piece of evidence, but it seems that the practice is much more than that because people put so much more faith in this kind of technology and, generally in face recognition. I mean, another side effect, the major reason for wrongful convictions: eyewitness testimony. It accounts for about above 70% of wrongful convictions. So. - [Mark Temelko] I think we have time for one more question here. - [Audience Member] My question's kind of related to the culture question. So how quickly do you notice the stereotypes of a society or culture change based on the social movements? For example, in the US, like there's more awareness around like women empowerment, for example, or like more social movements around race and gender. Do you see the societal stereotypes related to faces change and how quickly can that happen? - [Todorov] Oh, I don't know about, this is a much tougher question. I mean, there's a lot of work that are changing stereotypes, you know, let's say gender, over the last 50 years. ### [55:00](https://www.youtube.com/watch?v=JHv5QGESAlw&t=3300s) Segment 12 (55:00 - 55:00) And there's big changes, for the better. And that, and that seems to really, the changes becomes with when suddenly have a big participation of the women in the labor force. Whether that would change, I mean, I would think that would also change perceptions. I mean, will change also impressions from faces. Again, it's just, it's whatever models you expose to, and going back to the same idea of the, your statistical data faces. But it takes time. I mean, like you look, you know, this research, that decades of data and yeah. Okay. - [Nick Epley] All right. Thanks so much, Alex. (general applause) - [Todorov] Thank you. --- *Источник: https://ekstraktznaniy.ru/video/45819*