Sandra Matz on How AI Reads Our Minds, and What It Gets Wrong
TBS CROSS DIG with Bloomberg
Can AI understand a person better than their spouse or family can? And if so, should that be a source of fear or of help? In this one-on-one interview on TBS CROSS DIG with Bloomberg, host Ryuichiro Takeshita talks with Sandra Matz, a professor at Columbia Business School whose book Mindmasters has just been published in Japanese. The Japanese edition was translated by the NTT Data team and includes additional chapters. Matz said those chapters were meant to add local relevance and practical recommendations for Japanese readers.
Matz's position runs through the whole conversation. AI already infers intimate things about people with striking accuracy, mostly from the "breadcrumbs" of data they leave behind. That ability can be used to exploit people or to support them. She argues that the right response is to understand and engage with the technology, not to reject it, while staying clear about where it falls short. In her view, those weak spots are non-verbal cues, cultural nuance, and the unusual "edges" of a person.
Two Kinds of MBA Students
Takeshita opened by asking how Matz's business school students use AI. Her answer was that they all use it, and that both students and faculty are still working out how to use it well. She sees the students splitting into two camps.
The first camp hands its critical thinking to AI. Matz said she is always surprised that these students think she won't notice, because the signs are obvious. When students give a chatbot the same prompt, it returns essentially the same essay. Each one may read well alone, but when she reads 20 of them, they say the same things in nearly the same words. The difference also shows in class. MBA courses depend on discussion and debate, and students who outsourced their thinking cannot keep up. When she probes a little deeper, she says, "they're out pretty quickly."
The second camp uses AI to make learning harder on themselves. These students treat it like a professor who questions them, finds holes in their arguments, and tests whether they would lose a debate against an opponent. Matz said most students currently belong to the outsourcing group. The students who actually work with AI, however, are the ones getting jobs.
Takeshita asked whether students fear that AI will make MBAs and business schools unnecessary. Matz said they are afraid, and she called that fear "not completely rational." She agreed it has become harder for MBAs to find jobs. Even so, she sees the same split here. Students who use AI to build subject-matter expertise, getting better at their work faster and in more depth, gain a large advantage. According to Matz, those students have multiple job offers, and better ones than graduates used to get.
A Toddler, Attention, and the Problem of Endless Praise
The conversation then moved to children and mental health. Takeshita noted that parents are worried, that some countries are banning social media for children, and that AI might be restricted next. Matz, the mother of a two-and-a-half-year-old, said she thinks about this personally. If she hands her son a phone for two seconds, he somehow already knows how to use it, and he gets upset when she takes it back. She described phones as "attention suckers" that deliver the next dopamine hit. Social media intensifies this, she said, because it adds social reward: the user gets positive feedback from other people, not just a video to watch.
She sees social media and AI as similar in some ways. Both can connect people, and both give an immediate sense that someone is listening. But she said she is "much more concerned" about the current generation of AI chatbots. They are trained, for understandable reasons, to be kind and constructive. Every question is brilliant, and the user is funny and smart. Even when a chatbot disagrees, it does so as gently as possible.
Her worry is that children who rely heavily on this will lose the ability to handle messy human interactions. Adults had to go through the friction of spouses and friends who disagree with them. That friction is painful and not always fun, but Matz called it "fundamentally human" and said it is how people grow and build meaningful relationships. If children never learn to handle it, she fears they will turn to AI every time. A friend might push back, get emotional, cry, or walk out, while an AI never will. She compared this to crack: great in the moment, probably not good in the long run. She said a body of research suggests this is already happening. Over time, people devalue their relationships with other humans and become more aggressive and more sexist, because their expectations have become unrealistic.
The Village That Knew Everything
Takeshita brought up an example from the opening of Mindmasters. Matz grew up in a German village of about 500 people, where everyone knew who had an accident and who was dating whom. Takeshita said his wife also grew up in a small village. She sometimes hated that environment, but people also knew her and were there to help when problems came up. It was a double-edged sword.
Matz agreed. Since moving to New York, she said, she has never experienced anything like it; there, people barely know your name. In the village, people who cared about her offered advice about relationships and about what to do after school. That was a remarkable support system, but also annoying. Villagers did not just want to know who she was dating. They tried to influence it, "pulling the strings" behind the scenes instead of openly giving advice. When Takeshita said this sounds like an algorithm, she agreed.
Algorithms now see almost everything people do, she explained, because every step leaves a trace: search history, a phone carried around the clock. She noted that people ask Google questions they would not ask friends or even partners, and that their conversations with ChatGPT can be "incredibly intimate." The result is that an entity people have no relationship with understands a great deal about their lives. Her example: she does not know OpenAI personally, yet it may understand everything about her life, and it could use that knowledge to exploit her.
AI as a Smoke Detector for Mental Health
Matz said mental health shows both sides of this most clearly. If a system can tell that someone is struggling, it can exploit them by selling things they do not need or pushing them further into despair. The same signal is also an opportunity. A phone might show through GPS that a person leaves home much less than before, or that they are interacting less with friends. That could mean nothing; the person might be on vacation reading a book. It could also be an early sign of emotional distress.
She compares such signals to a smoke detector. It is not a diagnostic tool like a visit to a psychiatrist, but it can indicate that something is off. If the signal is caught early and the person is pointed toward support, she said, they might be helped "before it becomes too late."
Takeshita said this matched what he had heard from parents whose children had mental health problems. They often said they wished they had known sooner. Matz added that seeing "smoke" does not mean confronting a child; it can be a reason to reconnect. She said the same applies to adults. Depression, for example, is likely to recur once someone has a history of it, and it is usually caught too late. She suggested that a person with such a history could nominate trusted people as "stewards" of their smoke alarm, such as parents, a spouse, or close friends. When the alarm goes off, it alerts them too, so the person does not have to dig their way out alone.
Takeshita said "engagement" is a key theme of the conversation. He admitted he is also afraid of AI, but since it is now the default situation, he feels people need to engage with it and bring human ethics into it. Matz agreed. Engaging builds an understanding of the limitations. Once people know that AI tends to be extremely nice, even sycophantic, they can deal with that and learn what AI is good and bad at. That literacy is being built gradually, she said, and it is "really hard."
How Much Can AI Know About Us?
Takeshita asked Matz to put a number, from 1 to 100, on how much of a person AI can understand. She said numbers are hard, but that AI is "really good," and was even before today's flexible models. Earlier machine learning systems were given a narrow problem and data with known answers. Even then, for traits that are easy to quantify and that people consider fairly private, such as sexual orientation, political ideology, and gender, accuracy was "in the high 90s." For a categorical judgment like liberal versus conservative, given all of a person's data, the prediction can be very accurate.
As a psychologist, Matz finds the more fluid traits more interesting: personality, values, and mental health, which are not a one or a zero. Numbers mean less there. What researchers can do instead is compare AI's judgments with those of people who should know someone well. When they do, Matz said, AI outperforms essentially everyone: co-workers, friends, and family members, including parents, children, and siblings. Her explanation goes back to Google and ChatGPT. Because technology feels somewhat anonymous, people reveal parts of themselves to it that they do not reveal to the people around them.
Language, Non-Verbal Cues, and Breadcrumbs
Takeshita pushed back. AI in 2026 learns mainly from text, he said, but he communicates with his wife through body language and time spent together, not only through messages. Why would something that learns from text do better?
Matz agreed that AI is still not good at reading non-verbal cues, and those cues matter a great deal in the moment. If she wants to know whether he is angry right now, reading his body language is very useful. Understanding someone's general tendencies or disposition is different: that requires sampling as widely as possible across many situations. This is where AI is strong, and not only with text. Text is how people interface with it, but it can also process search histories, GPS records, and other data in any format, then fit the puzzle pieces together. It sees what people intentionally communicate and everything else as well. Her example: if someone's phone keeps running out of battery, perhaps they are less organized than others. Takeshita said his phone does the same. People nearby rarely see these traces, she noted, though they have the advantage in a live conversation, where they can read the non-verbal cues.
Takeshita then asked whether these breadcrumbs can really tell the whole story of a person. Matz said it is never the full story. Media coverage often presents this as a "wonder weapon" that pinpoints exactly who someone is, but every prediction contains errors. Much of the data is noise. Someone might buy a gift for another person, or go somewhere for work rather than by choice. The power lies in combining all the breadcrumbs.
Takeshita raised Instagram posts made to show off. Matz said researchers call these "identity claims": statements of who people are and how they want to be seen. In her account, people tend to exaggerate in the same direction. Everyone wants to appear a little less neurotic and a little more extroverted, so the result is a "parallel shift." For that reason, it is not a big problem for AI. A true introvert is unlikely to appear more extroverted online than a real extrovert, because the extrovert is exaggerating too. The ranking between them survives. On top of that, AI has all the breadcrumbs people do not think about, such as a phone running low on battery or three trips to the grocery store in one day.
The Big Five: Where It Came From and What It Measures
Takeshita said one appeal of the book is that it combines traditional psychology with current science, and he asked Matz to explain the Big Five. She described it as a framework from psychology designed to explain and predict behavior across many situations. It was developed through language. Two psychologists went through the English dictionary and extracted adjectives, on the assumption that anything meaningful about how people behave and experience the world would be expressed in language. They then looked at how adjectives such as kind, generous, assertive, and determined cluster together. Does someone who calls themselves decisive also call themselves assertive? Does someone social also describe themselves as chatty? From these clusters came five dimensions that explain most differences in how people think, feel, and behave.
According to Matz, the Big Five is unusual because it holds up across cultures, languages, and contexts. It also helps explain things such as music preferences, vocational interests, and mental health. She stressed that each dimension is continuous, not a high-or-low category:
- Openness: curiosity about the world, intellectual engagement, and a strong interest in aesthetics and design, which she said is big in Japan. At the other end are people who are more conservative, down-to-earth, and conventional.
- Conscientiousness: sometimes called "the German trait." It ranges from organized, dependable, and reliable to, as she put it kindly, more flexible, or disorganized.
- Extraversion: sociability, enjoyment of other people's company, seeking stimulation, and optimism. More introverted people are content in their own company and do not need constant stimulation.
- Agreeableness: the interpersonal trait. Caring, trusting, and empathetic at one end; more competitive and critical at the other.
- Neuroticism: emotional life, from easily stressed and anxious to relaxed and calm.
Takeshita's Test Results
Takeshita had taken the personality test that morning and showed Matz his results. She said people who work in media are usually high on openness, because they are curious about the wider world rather than staying in front of the TV. His conscientiousness score was 2%, which Matz described as "a little bit more flexible." She said she uses the same phrase for herself. She called herself "the one German who didn't get high conscientiousness" and joked that maybe that is why she left for the US.
She used the result to make a broader point. People tend to assume high conscientiousness and low neuroticism are simply good, but every trait has an upside. People who are very conscientious often miss the bigger picture and struggle to adapt. Entrepreneurs, she said, benefit from some order in their spreadsheets but not if it stops them adjusting to new ideas. Takeshita's extraversion was high, which she said is unsurprising for a media personality. His agreeableness was also high, including trust and respectfulness, which she said matches her experience of Japan more generally. His negative emotionality, the neuroticism dimension, was somewhere in the middle. Takeshita told viewers they could take the same test on the book's official website.
Does AI Make the Big Five Unnecessary?
Takeshita asked whether AI can do better than such a test. Matz said the framework is useful because it allows comparisons across contexts. If you know nothing about a stranger, their Big Five profile gives you a rough idea of who they are. Once you know someone well, you tend to drop it. Nobody thinks about their spouse in Big Five terms, because they know the spouse in a more nuanced and complex way. She called the Big Five "a very pragmatic solution," not the most in-depth one. It is a generalization: people can be extroverted or open in many different ways, and that complexity is lost. AI, she said, is good at preserving it.
She explained why the Big Five stayed so useful before the latest models: there was always a human in the loop. In marketing, if the goal is only to predict which product a customer buys next, a model can use millions of dimensions, and reducing them to five would just add noise. But if a person has to decide how to speak to the customer or what motivates them, the Big Five helps, because humans cannot make sense of a matrix with a billion data points. Five is a manageable number. Takeshita mentioned the "seven plus or minus two" rule, and Matz agreed it roughly matches what the brain can handle.
Now that AI also generates communication and content, Matz said, the framework is often unnecessary. You can give AI everything known about a person and ask it to write the most persuasive ad for that specific person, and it can do that "pretty well." It needs a human-readable framework only when it has to communicate with a human. For example, if it wants to brief a shop assistant on how to talk to a customer consistently with the ad it produced, it has to translate its knowledge into something a person can understand.
Takeshita suggested that customers might eventually have their own AI agents, so that AI talks to AI without a human in the loop. Matz called that "a terrifying thought." She described this kind of AI as "your spouse on steroids": able to do much more than a spouse in some ways and much less in others, and familiar enough with a person that it needs a framework only to explain something.
Where AI Falls Short: The Average and the Edges
Asked where AI's knowledge is lacking, Matz said it "overindexes on the average." AI is trained to predict the most likely next element in a sequence, which by definition is the most common one. When it tries to understand a person, it drifts toward the average. In studies using "synthetic humans" to test how customers might react to a product mock-up, she said, AI gives a good sense of the average response, such as which of two options people prefer overall. What it does not capture well is the distribution. It loses the edges: the odd things about a person, or the moments when someone acts out of character.
This is where a spouse still does better, in her view. A spouse knows that a partner is generally calm but will "freak out" in one particular situation. Humans, she said, see much more of this complexity than AI can preserve, given how AI works.
Bowing, Culture, and Western Training Data
Takeshita said AI seems to assume that people express everything in language, while in Japan much communication is non-verbal, such as nuanced bowing. Is AI too dependent on language?
Matz expects this to change. She does not think the future will consist of typing questions into a chatbot; interaction will become more fluid. As computer vision and the models improve, she thinks AI will eventually pick up on such signals. But Takeshita had pointed to a deeper problem: even then, everything depends on the training data. If that data is mostly Western, showing how extraversion or politeness is expressed in a Western context, AI will default to those patterns unless cultural sensitivity is built into the model. She said many companies are trying to do this, but it is "not trivial." Cultural nuances often sit at the edges, and if AI is weak at the edges, they will be lost.
Takeshita returned to the theme of engagement. He suggested that Japanese companies, citizens, writers, and journalists could put more content online and help make AI more diverse. Matz said this is hard for individuals. It is unclear at what stage of training cultural sensitivity can be added, and if there is little data about certain minorities, it is not obvious how to build them into the model. People inside the AI companies also need to be thinking about these questions, and she said the companies are "still catching up a little bit."
Studying the Psychology of Models
On the future of AI and psychology, Matz said that as a scientist she finds it fascinating in both directions. AI can be used to learn about the human mind, and models can be studied as having psychologies of their own. They are not all the same. When Takeshita said ChatGPT and Grok feel different, she agreed they have "completely different personalities." She said there is a scientific push to look beyond outputs to the architecture: how activation patterns differ inside each model. That kind of inspection is very hard to do with the human brain, outside some patient cases. So she raised the possibility that we might come to understand the models better and learn about the human mind along the way.
She described the risks as widely discussed. One is "cognitive surrender," for which her MBA students are a prime example: people lose skills such as critical thinking. Another is the loss of the ability to handle offline relationships, which she said she is especially interested in. Most of her own research, however, is on the opportunity side. If AI makes people boring by pulling them toward the average, she asks how systems could be designed to do the opposite and enhance people's complexity. She expects the downsides to dominate public discussion over the next few years. Her hope is that scientists can identify uses that reduce these risks or support human flourishing.
Why Grok and ChatGPT Differ: Training Data
Takeshita asked why models behave differently: the training data, how they absorb it, or the length of training? Matz said "everything," and gave Grok as an example. She said it was trained essentially on Twitter. When she asked whether Twitter content was more sexist, Takeshita said he did not want to comment but understood her point. She said this shows directly how training data shapes a model.
In business settings, she noted, there is debate about whether models need to be so large. Smaller models trained on focused, high-quality datasets might work better than models trained on the whole internet, including Twitter and Reddit. She said this makes human-produced content valuable, citing her understanding that The New York Times now earns a lot by licensing its content. The data also shapes what she called the "personality" of a model.
Takeshita asked whether a model trained only on high-quality material, such as her book, paintings, or plays, would be better. Matz said it depends on the goal. If the aim is to capture human experience broadly, then Twitter and Reddit are useful, because they include the dark sides of human nature, which the model can then reproduce. If the model serves a specific business purpose, excluding data that would never be used in that context might improve it. For someone writing journalistic articles, she said, focusing on high-quality content rather than Twitter may make sense.
Will AI Produce a New Theory of Psychology?
Takeshita asked whether research on how models work will change psychology itself. Matz said she hopes so, but the work is technical and difficult, and "we're not doing a great job" yet. She described the field of mechanistic interpretability, which aims to open models and understand causally how they make decisions. Models hallucinate, she said, because they are probabilistic: they choose the most likely next element, and the result is not always the same. Once the causal chain behind a model's decisions is fully understood, she thinks we will be much closer to understanding psychology.
She added a twist. People do not fully understand the human mind either. When a person explains their own decision, the explanation may not be the real reason; she said neuroscience shows people often make up a story after the fact. So we may end up understanding models better than human minds, and the interesting work may be explaining the gap: cases where an explanation fits a model but not a human.
Asked whether AI will produce a breakthrough theory of psychology, Matz said she was "not entirely sure." She expects a great deal of strong AI-driven research in the natural sciences, where the task is understanding building blocks. Psychology, she said, is centered on subjective experience. AI might raise interesting questions or show how theories connect, but she finds it hard to imagine it producing a completely new theory no one has thought of. She thinks it may help explain how physiology relates to experience, how neuroscience works, and how consciousness works. Her conclusion: likely yes for the natural sciences, and for pure psychology, an open question.
AI mindology.
AI.
So, Professor Sandra Matz, welcome to our show.
Thank you so much. It's a pleasure to be here.
And I just praised your new book, Mindmasters. It's out in Japanese. So I read your English version as well, and I think this is also translated by NTT DATA and has special chapters, so it's fun to read.
Which is great because I think my concern was, how do we make it relevant for the Japanese market? And I think it was really fun to work with the NTT DATA team on just adding a bit of flavor and some practical recommendations.
So the topic today, we're discussing AI because we all know that ChatGPT can maybe read our minds, but at the same time there are also risks. So just to kick this conversation off, how are your students in the business school using AI?
They're all using it. It's so funny because I think the students grapple with how to use it effectively, and I think we as faculty also grapple with how do we help them use it. And so I think you almost see these two camps. Some of the students just use it to outsource all of their critical thinking. They have to do their homework. It's—
The students use AI.
I'm always amazed by how they think I would not notice. There's so many dead giveaways. First of all, the essays that they write, they all look super similar to one another, right? They give it the same prompt, so it spits out the same. And it might read nicely for them individually, but then I look at 20 essays and they all say exactly the same thing with pretty much exactly the same language. So I do think you can still tell.
You can also then still tell when we actually have conversations in the classroom. Because it's an MBA, there is not just that they do their homework, but then we have discussions and we debate. And so you can tell that the students that just outsource, they don't follow the conversation. And so when you start probing them and you try to go a little bit deeper, they're out pretty quickly.
Then there's the other students that really try to use it to make their learning more intense, if you want. So they almost have it as a professor questioning them and poking holes into their arguments and seeing whether they actually lose maybe the discussion with an opponent. And so those are the ones who really thrive. I would say right now, I think the majority falls more into the outsourcing bucket. But the ones that really do the work with AI, those are the ones that get jobs.
Are they afraid? Because people say if we have AI, we don't need MBAs anymore, we don't need business schools. Are they afraid? Because they have to get jobs.
I do think they are afraid, and not completely rational, I would say. I do think it has become harder for MBAs to find jobs, but that's also where you do see this dichotomy of the ones that I think use AI in the right way. They not only understand AI, but they use it to build their subject matter expertise. So they really become better at what they're doing much more quickly and much more deeply. And I think then it becomes a huge advantage. And so those are the ones who have multiple jobs and who have much better offers than they used to.
So I think throughout this talk we'll touch on several subjects like mental health. Parents are worried about social media, AI. Their kids are asking chatbots for everything, and it can affect children's mental health. Some countries are banning social media and maybe in the future ban AI. What's your take on that?
I'm the mom of a two-and-a-half-year-old, so I think about these questions.
Is he already using AI, or—
It's remarkable. You give him the phone for two seconds and he changes all of the—he just knows somehow magically how to do it. So I think it's going to be a battle because I think the challenge with phones, the same with social media, is they're just such attention suckers. And it's clear that you just get the next dopamine hit. So it's very hard for him. You can see it: he's two and a half and he gets upset when I take the phone.
So I think the same is true for social media, just intensified because we're suddenly giving social reward, right? It's not just like the movie's playing, but you suddenly get positive social feedback. And so with social media and AI, I think they're similar on some level in that they, first of all, sometimes connect you with other people, but they're also giving you this immediate sense of someone is listening. And so social media is one. I'm actually much more concerned about AI and the current version of chatbots.
What I think happens there is the way that they're trained, and for good reason, they're very kind. They're very constructive, right? They say, whatever question you ask—
They say, "You're the best."
"You're the best." Exactly. Every question that you ask is brilliant. You're super funny. You're smart. So the only thing that you get is this positive feedback. And even when they disagree, they do it in the most constructive way, right? And so what I'm concerned about is that the more kids are using that, that they just lose their ability to deal with messy human interactions.
You and I, we had to go through the friction of having a spouse, a friend that we disagree with. And it's painful and it's frictionous. It's not always fun, but it's in a way fundamentally human. And that's the only way that we grow and we have meaningful relationships and lives. And so what I'm concerned about is that if kids don't learn how to deal with that, they're going to turn to AI every single time, right? Why would you have a friend who pushes back and gets emotional and starts crying and walks out on you if you can just go to your AI?
And there's a bunch of research suggesting that this is happening. It feels good in the moment. It's a little bit like crack, right? Great in the moment, but probably not so good long term. And you do see this playing out already where over time we just devalue the relationships that we have with other humans. We become more aggressive, more sexist, just because we have expectations now that are unrealistic.
I like the example you gave to the audience of Mindmasters. In the beginning you mentioned that you grew up in a small village in Germany. 500 people only, very small. So 500 people, they all know what you are doing. If you have an accident, they know the next day. People know who you're dating, everything. And it was a great example because it's similar to AI. They know—I don't know if it's they or not—but they know everything. It's a bad thing and a great thing at the same time because my wife also grew up in a small village, and she sometimes hated that environment. But at the same time, people know you, so if you have some problems, people are there to help. So it's a double-edged sword.
Yeah, absolutely. And to me, I so resonate with the experience because everybody knows everything. It's a community that I've never experienced anything like it after. I now live in New York, and you barely know your name, exactly. But there was this incredible support of people knowing you, and the moment that people cared about you, they were there to give you advice on relationships and what to do after school. So it was an incredible support system, but also obviously incredibly annoying at times because they were not just trying to figure out who you were dating, they were also trying to influence who you were dating. So I think that is the tricky part.
Advice who you should date.
Behind the scenes, right? They're just pulling the strings. They're not even giving you advice. They're trying to push you.
So it's like an algorithm.
Yeah. And exactly the same is true for algorithms, right? So today algorithms now observe essentially everything that you do because you leave traces with every step that you take, whether that's your searches or the phone that you're using, you carry around with you 24/7. I always think of Google, right? We ask Google the most intimate questions that we wouldn't feel comfortable sometimes asking our friends or maybe sometimes even a partner. Take ChatGPT. We have conversations that are incredibly, incredibly intimate.
And so I think you're absolutely right in that that means that there is someone, and it's an entity that we don't have a relationship with, right? I don't know OpenAI personally, and yet there's someone who understands everything about my life. And now they can use that knowledge to exploit me. And I think mental health is such a good example because the moment that I understand, for example, that someone is struggling, I can obviously use that to exploit them and maybe sell them stuff that they don't need or push them even further into desperation.
But it's also an incredible opportunity to say, there is someone whose behavior seems to deviate from their typical baseline, right? Take your phone. If I see that you're not leaving the house as much anymore as you used to based on GPS, maybe you're not engaging with your friends as much anymore. Could be nothing. Maybe you are just on vacation having a great time reading a book. Or it could be that it's an early indication that you might be in emotional distress.
And so the moment that I get the signal, I always think of it almost like a smoke detector. It could be nothing, right? It's not a diagnostic tool the same way that you going to a psychiatrist is, but it could be an indication that something is off. And if I catch it early and then point you in the right direction to find the support that you need, well then maybe actually we manage to save you before it becomes too late.
That's a good point because I have interviewed some parents. Unfortunately, their children had mental health issues, and they always say that they wanted to know it beforehand. "Wish I had known that." Maybe it can be nothing, but it's good at least to see smoke.
Yeah. And I think seeing the smoke also means that you can engage, right? So it means that it gets pointed out that something might be off. So you don't have to now jump in and confront your kid about why there's—but it's a way of you trying to connect again. And I think this is true for kids. I think the same is true for adults.
Because oftentimes, if you take depression, for example, we know that once you have a history of depression, it's very likely going to happen again. And most of the time we catch it too late to really be supportive in the moment. So if you have a history of depression, you could just nominate someone, right? So you could say, I'm going to make you a steward of my smoke alarm. So the moment a smoke alarm goes off, it not only goes off for me, it also goes off for people who I trust. That could be my parents, could be my spouse, could be really good friends. And so now it's not just on me to try and dig my way out of this hole again, but I get the support from people around me.
I think engage is a key topic of this discussion today because it's important to know the risks of AI. And I feel I'm also afraid of AI, but at the same time it's a default situation, so I need to engage and put some human beings, ethics and everything into this area.
Yeah. And I think the more that you engage and the more that you understand, you also know the limitations. But if we say that AI has the tendency to be super nice and super kind and sick, well then there's a way for you to deal with them because you understand it. And you also know, here's what AI is good for. Here's maybe what AI is not so good for. And I think this literacy we're just gradually building, and it's really hard.
So I have two questions. One first is a quick question. So out of one to 100, 100 being the most, how many percentage of ourselves can AI understand, the human being?
Numbers are always hard, and I'll explain in a second why. So it's really good. So it already used to be good before we had the current forms of AI that are very flexible.
Before we had some—
Exactly. We had machine learning, which was essentially saying, I'm going to give you a very specific problem. I'm going to give you some data with some ground truth, and I'm just going to have you figure out how they go together. And already back then, if you take things that you can more easily quantify, like predicting your sexual orientation, political ideology, gender, things that we already find relatively intimate and we don't necessarily want everybody else to know about ourselves, your scale from one to 100%, you're in the high 90s, right? If it's a categorical judgment of are you more liberal or are you more conservative, I give it all of your data, it can make a very, very accurate prediction.
For me as a psychologist, the more interesting aspects are the ones that are more fluid. So it's not just a label, like personality, values, mental health. Oftentimes that's just not a one or a zero. And putting numbers there is harder because they oftentimes don't mean much. But what you can do, and what I think is an interesting comparison for us, even with a comparison of the village, is we can look at how good is AI at making some of these predictions about who we are, how we feel and so on, compared to other people in our space.
So if I asked your spouse, for example, your wife, what is the personality profile, mental health, values? I can now compare the AI to people who should know us pretty well. And what you see there is that AI outperforms essentially everybody. So it outperforms co-workers, friends, family members. And again, this includes parents, kids, siblings. They should know us pretty well because they spend a lot of time. But coming back to Google and ChatGPT, there's just aspects of ourselves that we reveal to technology because it feels somewhat anonymous that we don't reveal to the people around us.
So this leads to my second question. So AI at this moment, 2026, they're learning through text, languages. But this puzzles me because when I interact with my spouse, I'm not only messaging her, but also communicating with her using body language or spending time with her, everything. Does this not matter? Because why is AI better than us just learning from language or text?
I think you're absolutely right. What AI is still not good at is picking up on nonverbal cues, right? And usually nonverbal cues matter a lot for the moment. So essentially, if I want to figure out if you're angry, reading your nonverbal cues in the moment is incredibly helpful. If I try to get a bigger picture of who are you as a person, generally speaking, I would say a behavioral tendency or disposition, then what I want to do is I want to sample as widely as I can across different situations.
And what AI is good at, and that's also where it's not just text, right? AI can easily process text. That's what we use to interface. But it can also easily read or process your search history, your GPS records. So it can take everything, no matter what the format is, and then just piece the puzzle pieces together. And for me, that's the interesting part because it gets to see everything. It gets to see the things that you intentionally communicate, but it also gets to see everything else. If your phone is constantly running out of battery, maybe you're just not as organized as other people.
Exactly. Mine, too.
So there's all of these subtle cues that we don't really think about, but it can pick up just because we create these data traces. And that's something that oftentimes the people that we interact with don't really see, but they do have a leg up in
The moment when it comes to now we have a conversation and we get to read all of the nonverbal cues.
Because those are breadcrumbs, right? But for me, breadcrumbs are just crumbs. So it's not the whole myself. But is it true that those breadcrumbs can tell the whole story of a human being?
It's never the full story. When you look to the media, there's always this wonder weapon, and they can pinpoint exactly who you are. Of course there's error, right? So with every prediction, you're going to make mistakes. It's never perfect. But when I think of these breadcrumbs, it's essentially the combination of all of them that's powerful. There might be a lot of breadcrumbs that are just noise, right? Maybe you did something, maybe you purchased something not for yourself, but someone else. Maybe you went somewhere because you had to for your job. It wasn't your choice.
I post on Instagram. I try to show off.
Yeah, exactly. Social media.
Social media, yes. It's not myself.
But that's the interesting part, because essentially what AI can do and what we can do with data is we can see the signals that you explicitly send. Social media, we call them identity claims, right? This is you saying, here's who I am. Here's how I want you to see me.
And what we see there is that typically we all exaggerate a little bit in the same direction. So if we all want to be a little bit less anxious and neurotic, and we all want to be a little bit more extroverted, typically it's a parallel shift that you see where we all exaggerate in the same direction. Which means it's not actually that big of a problem for AI, because it's unlikely if I'm really a total introvert in the real world. You are more extroverted.
Okay.
I'm very unlikely to overtake an extrovert online because the extrovert is also exaggerating. So rank-order-wise, the AI still gets a sense that maybe you're more extroverted than me. Plus, we get all of the other breadcrumbs that we don't think about, right? I don't think about the signal that I'm sending when my phone's running out of battery or when I go to the grocery store three times a day because I can't...
Yeah. But it's a combination, and it can figure out.
And that's what it's really good at.
It's good at. Okay. So I think one of the interesting parts of this book is you're combining traditional psychology and cutting-edge science together. Could you explain to the audience what Big Five is and how is it different from the current situation in the era of AI?
Yeah. So Big Five, and this is really something we're showing to the audience.
Yeah. It's really coming out of psychology. And the idea of the Big Five is essentially, can we create a framework that allows us to both explain and predict behavior across a whole bunch of situations? And so the way that the Big Five—it's a personality framework—was developed was actually looking at language. There were two psychologists who went through the whole English dictionary and pulled out adjectives.
Language is very important in psychology. Okay.
Because it's how we communicate, right? Everything, the way that we think about ourselves, the way that we interact with one another, all comes down to language. So the idea was anything that's meaningful to us in terms of our behavior, in terms of how we experience the world, should be communicated and expressed in language. That's where they started, and they just pulled out all of the adjectives. So kind, generous, assertive, determined, and they just looked at how do they cluster together. If I say I'm very decisive, do I also say I'm very assertive? Or if I say I'm very social, do I also say I'm very chatty? And so they looked at all of these different adjectives, and they pulled out five dimensions that explain most of the behavior. The way that we conceptualize the Big Five, it's differences in the way that people think, feel, and behave.
And so what's interesting about the Big Five is that it's one of the only frameworks that generalizes across different cultures, across different languages, across different contexts, and it explains behavior in different contexts. So if I'm interested in trying to understand your music preferences or your vocational interests, your mental health, Big Five is a framework that explains quite a lot of that. So that's why it's useful.
Could you quickly walk us through? Beginning from openness.
Yeah. So openness, and then you also see how they essentially work together. Openness is the extent to which people are open to new experiences. So they're curious about the world. They're intellectually engaged. There's a strong focus on aesthetics, so design, which I know is big in Japan, and it's wonderful, as opposed to people who are a bit more conservative, down to earth, and conventional. And all of these dimensions are continuous. So it's not that you're either high on openness or low. It's like you're somewhere in between.
Then the second one is conscientiousness, which they oftentimes call the German trait, which is to an extent like you're very organized, dependable, reliable, or a little bit more flexible is a nice way of putting it, or disorganized. Extraversion is the one that I think we oftentimes use in our day-to-day language. So are you social? Are you excited about spending time with other people? Do you seek stimulation? Are you optimistic? And then people who are more introverted are just a little bit more content with their own company, and they don't need this constant stimulation.
Agreeableness is the interpersonal trait. So to what extent are you caring, trusting, empathetic, as opposed to more competitive and critical? And then neuroticism is the one that deals with our emotional life. So to what extent are you easily stressed, anxious, or just much more relaxed and calm?
I actually did a test this morning. Can I show it to you?
Yes, please. I was going to say anyone who works in media news is usually high on openness, because that means that you're curious about the world, right? So if you're curious about what happens in the world, not just in your own life, you're not just stuck in front of the TV, but you're actually reporting, that's a giveaway. Let me see what else we have here.
Conscientiousness. Yeah, your phone running out of battery. Conscientiousness is low. It's like 2%, which means...
2%. Okay. Which means not organized. Okay.
A little bit more flexible.
Flexible. That's a good way to put it.
That's how I say it for myself. Because I think I'm the one German who didn't get high conscientiousness. Maybe that's why I left for the US. But that also means that oftentimes people who are low on conscientiousness, they're not stuck in the details, right? We always have a tendency to say it's great to be conscientious, it's great to not be neurotic. But there's always a flip side where you do get benefits. So people who are overly conscientious, they oftentimes don't see the bigger picture, and they oftentimes don't adapt, right? If you think of entrepreneurs, it's great to be somewhat organized in your spreadsheet, but it's also not that great if you can't adjust to new ideas. What else? Okay.
Extraversion high, also not surprising as a media personality. Agreeableness, so that's the caring trait. You also score high on trust and respectfulness. I think that's also my experience of Japan more generally. And then negative emotionality, which is the neuroticism trait, somewhere in between.
Great. So anyone watching this can take this test. They can access the official website, mindm. So my question is, can AI do better than this test? Can AI know more about you?
It's actually a great point, because what is helpful about the framework is it gives us a way of comparing people in different contexts. If you meet a person in the street and you don't know anything about them, actually getting their Big Five profile is helpful because you do get a rough idea of who that person is. The moment that you get to know someone better, you can actually oftentimes discard the Big Five. You don't think about your spouse in terms of Big Five, right? Because you know them a lot more in a much more nuanced context, a more complex way.
Exactly. In a way, Big Five is a very pragmatic solution of trying to understand people. It's not the most in-depth one.
It's like a generalization.
Exactly. It's a generalization, and it's helpful because it gives you an idea, but it also means that you can be extroverted in so many different ways. You can be open in so many different ways. So we're losing the complexity. And what AI, I think, is really good at is preserving that complexity.
And the reason for why the Big Five was still extremely helpful before we had the latest AI models was, if you think of applications, there was always a human in the loop. So if you think of marketing, it was great if I could understand a consumer on a million different dimensions. If I just want to predict which product they should be buying next, then I don't need five dimensions of Big Five because it just adds noise.
If I think about how do I communicate with them in the best way, how do I address them, what are some of their underlying motivations, the Big Five was helpful because that's how humans can make sense of that, right? A human can't make sense of a matrix that has a billion different data points.
Five is a good...
Exactly. Five is like, we know five.
Seven plus and minus two.
Exactly. That's roughly what the brain can deal with. But now that AI is also creating communication and it's creating content, oftentimes you don't even need this anymore, right? So you can just say, "Here's everything that I know about you." And if I then tell it, "Well, now just create the most persuasive ad for you specifically," it can actually do that pretty well. It doesn't necessarily need the Big Five anymore, unless it wants to communicate with another human. But if now the AI wants to say, "Here's the message that I created," and I want to give a shop assistant in the store advice on how to communicate with you such that it's consistent, then it needs to turn it into something that's human.
So the human understands.
Exactly.
But in the future, maybe the customer will have their own AI agent, so it'll be AI to AI. So maybe we don't need a human in the loop anymore.
Yeah. It's a terrifying thought.
Are you saying that AI doesn't have to use this Big Five? It's like the extreme personalization of this world.
Yeah. Essentially, it's your spouse, right? I always think of it as on steroids because it can do a lot more than your spouse in many ways, and also much less in others. But it's someone who knows you so well that it doesn't necessarily need the framework unless it wants to explain something. Then it's really helpful.
Are there any areas that AI lacks knowledge?
I think what AI is not good at yet is it essentially overindexes on the average, right? So if you think about the way that AI was trained, it essentially predicts the most likely next word in a sequence of words, or generally element in a sequence of elements, which by definition is the most common one. So when AI tries to understand you, it essentially defaults to the average. If you look at studies where we have synthetic humans and we try to see, well, how would customers respond to this mug here if you put it out on the market, you typically get a really good sense of the average. So on average, people like this one better than that one.
What it's still not good at is getting a sense of the distribution. You lose the edges, right? Lose the edges of a person being weird because it's just not in the general repertoire of AI, or of you as a person doing something that's a little bit different. And so I think that's where the spouse still is doing better, because the spouse understands, well, okay, he's generally pretty emotionally calm. If I put him in this specific situation, he's probably going to freak out. So I think as humans we see much more of the complexity that AI, just by the nature of it, can't preserve as well.
Let's talk about cultural differences. Since you're in Japan, you have your Japanese edition. I feel that AI is built on the assumption that people express everything by language. But in Japan, we don't necessarily use language. We have a lot of unofficial nonverbal communication. I'm sure you have realized that we tend to bow, and it's very nuanced. Do you think there are cultural differences? I think AI is too dependent on the language model.
Yeah. And I think that's probably going to change. Even just the way that we interact with AI, right? I don't think the future is going to be us typing questions into a chatbot. I think it's just going to be much more fluid in the way that we...
Like verbal communication.
Yeah. And also the moment that computer vision gets better and these models get better, I do think they at some point can pick up on this. But you're still hitting on an important point. Even if it can do that, it depends on the training data. And if the training data is all Western, and it's like, well, here's how extraversion gets expressed in a Western context, or here's how politeness gets expressed in a way, then this is what the AI is essentially going to default to unless you bake cultural sensitivity into the model. And I know a lot of the companies are actually trying to do that, but it's not trivial. It's much harder because, as you said, there are these subtle nuances that oftentimes are on the edges. And if AI isn't good on the edges, then we're essentially going to lose them.
I think, again, engagement is the point today. NTT is doing a good job, but I think Japanese companies or citizens or writers, journalists should put something online and teach AI, maybe in the future, to make it more diverse.
Yeah. I think it's hard for us to do as individuals, right? Because there's always the question, at what point in the training can you actually insert some of these cultural sensitivities? Is it the training data itself? And if there is just not as much data on certain minorities, well, how do you then still bake it into the model? So what you also need is people in those companies thinking about these questions, right? And I think there the companies are still catching up a little bit.
So let's talk about the future of AI and psychology. How do you see the future of these two areas?
I think, and this is also something that comes out in the book, for me as a scientist, everything is incredibly interesting. I find it really interesting to see what can we learn about the human mind, human behavior using some of these tools. It's equally interesting to see, well, can we actually understand the psychology of these models, right? Because they're not all the same, and it's the first time that maybe we can learn something about the human mind just by observing the behavior of the models themselves.
I think OpenAI's ChatGPT and Grok, they're different.
I know, completely different personalities. And right now I think there's a lot of scientific push to not just look at how are the outcomes different, but really looking at the architecture. So how is the activation in the model different? And what's fascinating about this is it's very hard to do in the human mind, right? We can't just open your brain and poke. Sometimes you can do that with patients, but it's very hard to understand exactly what the...
brain is doing. So there could be a world where we actually understand the models better and learn something in the process about the human mind.
But for me, this capacity is essentially then saying, well, what are some of the risks and what are some of the benefits? And I think the risks are widely discussed, right? I think it's very clear that what we call cognitive surrender, like our MBA students, the prime example, is that we just lose some of the skills that we have. Critical thinking is one of them. And the challenge that we discussed of, like, we lose our ability to deal with our offline relationships, that's a big topic that I'm interested in.
But then I think on the opportunity side is where a lot of my research sits. So how do we actually use AI to enhance our complexity? So if we say that AI makes us boring because it goes for the average, well, then how do we design systems that don't do that? And so I think over the next couple of years you will see in public discourse a lot of the downsides. And my hope is that what scientists can contribute to is really, what are some of the use cases that either mitigate some of these risks or really boost human flourishing and experience?
So at this moment, why do you think those models act different? Is it the training data, or how they absorb data, or the length of the training?
Everything. But I think Grok is a great example, right? Grok was essentially all trained on Twitter. And if you just look at what happens on Twitter, is it more sexist? Is it just a lot more...
I don't want to comment, but I know what you mean.
So you can see directly, I think, how the training data impacts. And if you talk about applications of AI, right, if we think about it in the context of business, well, there's a lot of conversation happening around, do these models actually need to be that big? Or is there a world where the models actually become smaller just because the data that we use is not the entire internet, which includes Twitter and Reddit, but it's much more focused data sets of high-quality data sets?
And I think that's an interesting one, is to what extent can humans produce content, right? I think the New York Times now makes a lot of money by essentially licensing the content. So there's something that is deeply human and that the models actually need by having high-quality content. And that also then impacts what the personality, if you want, of the model becomes.
Do you think it affects, let's say, an AI model, it only absorbs high-quality content like your book, or a painting or art, or watches a play? Do you think it will get better, or is there a more complex structure in it?
Depends on what you're trying to solve for. But if you're trying to have a model that just kind of captures human experience across the board, then having people expressing themselves online on Twitter and Reddit is actually helpful, because even the dark sides of human nature, you just essentially replicate them and you can bring them out.
If you have a certain purpose that you want to use AI for in the context of business, then you might actually be better off just kind of shutting out the data that's not—you would never use that data in a certain business context. So in that case, actually not using it could make it better. And so really the question is, are you just trying to replicate the human experience in your model, or do you have a certain target in mind? So if you're writing journalistic articles, yeah, maybe you should focus more on high-quality content as opposed to Twitter.
By doing research on AI and figuring out how the model works, do you think that will also affect the areas of psychology? We get to know ourselves better.
I hope so. So I think right now there's a lot of research, and it's technical research, and it's still hard, and we're not doing a great job. It's called mechanistic interpretability, which is essentially trying, can we open the models and fully understand causally how they make their decisions? So it's not because, right, it's essentially the reason for why they hallucinate is it's probabilistic. It's just the most likely next part, and it's not always the same. The moment that we can really understand the causal chain of here's how the model makes its decision, that's, I think, when we become a lot closer to understanding psychology.
Now, you can also say we don't understand the human mind, but if I ask you to explain a decision, I don't know if that's actually the right explanation. We know from neuroscience that oftentimes it's not, and you just make up a story on the way. So I think we actually might end up in a world where we understand the models much better than we understand the human mind, and what we might end up explaining is the discrepancy. Right? So when do we understand a model, but it doesn't actually apply to the human mind?
What's your prediction? Do you think there will be a groundbreaking theory of psychology in the future because we have AI and we can experiment?
I'm not entirely sure. I think there's going to be a lot of incredible research in the natural sciences. So I think there, for sure, because there it's really, you just need to understand it. You need to understand the building blocks. In psychology, it's all about the human experience, right? So it might be able to say, here's an interesting question, and maybe here's how this theory intersects with this theory. But at the end of the day, because psychology is so much about the subjective experience, it's hard to imagine that it would just come up with this brand-new theory that no one's ever thought about.
In that case, I feel like it can help us just probably better understand how, let's say, physiology interacts with experience, how neuroscience unfolds, and how consciousness works. So I think on that front, on all of the natural sciences, probably yes. Pure psychology, not entirely sure.
Okay, professor, so we need to wrap up. I highly recommend this book to our audience. So any final comment, final words for our audience?
No, it's just such a wonderful time to be here.
Okay. Thank you so much. I really enjoy the conversation. Thank you. Okay.
Thank you so much.
Thank you so much.
Article published · Updated