התמלול המלא של 476 What Happens When AI Starts Doing the Thinking? With Eric So

their capabilities. 00:05:12.240 --> 00:05:16.560 So, for example, my dad was wonderful with math. 00:05:16.800 --> 00:05:19.839 He could sort out all kinds of things in his mind. 00:05:20.160 --> 00:05:31.040 He was a little frustrated that his sons were good at math, but not that good and sometimes needed to use tools or calculators. 00:05:31.279 --> 00:05:44.000 And then when I see my girls, they have zero capability of working out math in their heads, but even the SATs that they have to take, they can use their calculators. 00:05:44.240 --> 00:05:48.720 So is this sort of the next version of having a calculator? 00:05:48.879 --> 00:05:52.639 What is wrong with us not knowing how to do math in our heads? 00:05:52.800 --> 00:05:55.759 So, what do you see as being wrong with this? 00:05:56.079 --> 00:05:57.279 I would take it in two parts. 00:05:57.439 --> 00:06:04.000 The first thing I would say is that there's a lot of past technologies like calculators that have undermated certain parts of our thinking. 00:06:04.240 --> 00:06:11.120 I think the critical difference of AI in this era is that AI offers to do the complete chain of thinking. 00:06:11.279 --> 00:06:22.879 So there's a famous quote that calculators can only calculate, they cannot do math, in the sense that calculators can do simple arithmetic, but they don't really choose the problems to be solved. 00:06:23.040 --> 00:06:26.800 They don't really understand the broader context, they don't think about how to apply it. 00:06:27.040 --> 00:06:35.759 With the rise of AI, in particular agentic AI, these systems are increasingly offering to remove us entirely from the thinking process. 00:06:36.079 --> 00:06:45.279 And I think that really poses a threat to young developing minds in the sense that I think people of a certain vintage, so I'm 46. 00:06:45.360 --> 00:06:51.120 I certainly remember struggling through problems growing up, and that's really a big part of how I develop my mastery. 00:06:51.279 --> 00:06:54.079 I suspect the same as you and many of your listeners. 00:06:54.319 --> 00:07:12.000 The thing I really worry about is the people who are developing some foundational mastery or attempting to are outsourcing that struggle component that was so critical for our development to an AI system under the belief that the output is sufficient evidence of mastery. 00:07:12.160 --> 00:07:14.000 And that's really not the case. 00:07:14.319 --> 00:07:16.319 I couldn't agree with you more, Eric. 00:07:16.480 --> 00:07:24.319 A lot of times I find myself writing and in struggling with trying to convey a concept and write about it. 00:07:24.480 --> 00:07:27.759 I end up understanding the ideas much better myself. 00:07:27.920 --> 00:07:29.920 So that struggle is really important. 00:07:30.399 --> 00:07:34.639 But I still want to understand what the issue is with this. 00:07:34.879 --> 00:07:43.920 In that couple of months ago, Eric, my family and I were visiting Orlando, and now they have Waymo's running all around. 00:07:44.079 --> 00:07:46.399 We took a Waymo, it was fantastic. 00:07:46.639 --> 00:07:51.120 So I could visualize a day when people don't need to drive. 00:07:51.279 --> 00:07:52.879 Maybe some people will drive. 00:07:53.199 --> 00:08:05.519 And that automated system that doesn't allow humans to think, at least based on current safety tests, is a lot better than the systems where humans are thinking and acting. 00:08:05.759 --> 00:08:16.720 So again, why should we protect that thinking, that struggle, that acting on the human side when the machine, in this case, can do it so much better? 00:08:17.519 --> 00:08:18.720 I would say a couple things. 00:08:18.959 --> 00:08:27.439 One is that there are many parts of work that are not directly separable from the work that you can produce with AI. 00:08:27.680 --> 00:08:33.039 So, for example, I have many students who will generate a really polished presentation with AI. 00:08:33.200 --> 00:08:45.120 But then when it comes to the final portion of that deliverable, that is standing up and defending it and having discussions with people, they really struggle through it because they've skipped the sort of the foundational component. 00:08:45.360 --> 00:08:57.919 And I think that reflects many parts of work where you need to make sure you understand what you're talking about or what you've developed in order to really complete your side of the business transaction. 00:08:58.240 --> 00:09:06.159 And in the absence of that initial struggle of really thinking through what goes onto the paper and how it's formulated, I think you lack that ability. 00:09:06.399 --> 00:09:09.840 These are what economists often refer to as sort of strong bundles. 00:09:10.000 --> 00:09:15.519 You can't really separate the output from the knowledge necessary to defend it if you want to have value. 00:09:15.679 --> 00:09:29.759 I think that it speaks to this really important point that in the absence of being able to think independently, to really stand outside the AI, it really begs the question on a personal level: why would I pay you the salary that you hope to command? 00:09:29.840 --> 00:09:33.679 In the sense that I can hand off all of these tasks to an AI system. 00:09:33.840 --> 00:09:41.840 So from a personal level, I think it's critically important to develop that mastery as a means to have something that you can contribute. 00:09:42.000 --> 00:09:46.480 It's a big reason why you're in the room and why people pay you the salaries that they do. 00:09:46.879 --> 00:09:52.080 I wonder, Eric, are universities well suited to develop that mastery? 00:09:52.320 --> 00:09:59.360 I agree with you in that when I use AI and I use it sometimes too much to your point. 00:09:59.519 --> 00:10:04.879 Sometimes I catch myself outsourcing thoughts before I should. 00:10:05.200 --> 00:10:20.399 That said, I can gauge the output, the response based on my experience, based on my judgment, which has been developed

over a series of years, experiences, and time. 00:10:21.120 --> 00:10:34.480 Part of what I hear from you is that younger people also need to struggle with ideas to develop that judgment in order to be able to add value to the output that AI gives. 00:10:34.799 --> 00:10:42.480 But I don't know if our education system is well suited to help them become capable in doing that. 00:10:42.879 --> 00:10:48.000 I think that AI poses a huge threat to the existing models of education on many fronts. 00:10:48.159 --> 00:10:52.639 And I think it really poses an important challenge that education adopts around AI. 00:10:52.799 --> 00:10:59.840 There's no doubt that future leaders will require people to use AI, push people, and it'll be part of our work. 00:11:00.080 --> 00:11:06.000 And so it's not a matter of yes, no, but I think it's really a matter of training people to use AI well. 00:11:06.240 --> 00:11:09.840 As I talk about in the book, AI isn't inherently good or bad. 00:11:09.919 --> 00:11:11.679 It really depends on how you use it. 00:11:11.840 --> 00:11:18.480 And I would go a step further to say it really is, it matters a lot about how we interface with the technology. 00:11:18.720 --> 00:11:32.240 Meaning a lot of choices that we make, both in terms of querying these systems, but also how we see the information that is presented, really can change whether we grow through using AI versus weaken our abilities. 00:11:32.480 --> 00:11:50.480 The additional thing I would say is that there's a lot of argument in favor of giving people real-world problems to solve with AI and then using the educational sphere to help inform the right approaches, but also to bring people together to think about those problems like jointly. 00:11:50.639 --> 00:11:59.519 And I think that's really where education can help in assembling people with a common problem, an idea, give them a training ground for developing this expertise. 00:11:59.679 --> 00:12:04.240 So I think education can help quite a bit with that, but I think you're totally right. 00:12:04.399 --> 00:12:16.320 Then it's current formulation, the way that you and I perhaps were grew up in the standard education system probably is not very well suited because it's so based on the deliverable that now you can hand off to AI. 00:12:16.720 --> 00:12:18.799 And it is a significant issue, Eric. 00:12:19.039 --> 00:12:37.519 Part of what I see is when you add AI with the fact that now through pendants, glasses, earbuds, it can take in all the information in our world and therefore give us recommendations and then we can act. 00:12:37.759 --> 00:12:49.919 We can to a certain extent become just the transmitters of the AI information that is coming into us if we don't take ownership over our own thinking. 00:12:50.159 --> 00:12:55.279 And one of the points that you make is about AI gravity. 00:12:55.600 --> 00:13:10.480 So the concept of AI gravity is really this idea that AI really pulls at people and many aspects of

the way that we are as humans and our economic systems that kind of nudge us towards more and more AI use. 00:13:10.720 --> 00:13:12.399 So, what are some of these components? 00:13:12.559 --> 00:13:26.559 One is simple a biological reality that thinking is like it turns out to be like an incredibly expensive biological habit in the sense that our brains are this paradox of biology. 00:13:26.639 --> 00:13:33.440 They represent only 2% of our body weight, but consume over 20% of our calories, even when we're resting. 00:13:33.679 --> 00:13:45.200 And when you put people into these high stress situations where it's really intellectually demanding, consumes a lot of those calories that our body and our evolution has trained us to try to conserve. 00:13:45.440 --> 00:13:50.639 So it's really no surprise that people feel this urge to have AI do their thinking for them. 00:13:50.720 --> 00:13:53.600 It's hardwired as a temptation into our thinking. 00:13:53.759 --> 00:13:56.559 The same thing is true of many economic systems. 00:13:56.799 --> 00:14:06.080 If you look at the way society operates, much of how we're compensated is not on the basis of necessarily the level of where we sit, but rather a comparison. 00:14:06.240 --> 00:14:08.240 How do we stack up against other people? 00:14:08.399 --> 00:14:24.799 And so when you put people into a situation where I don't really know if you're using AI or not, but I worry that you're using it to produce more, to get ahead, to generate more reports, it really creates this arms race, this concern that if I don't use it, I'll fall behind. 00:14:24.879 --> 00:14:30.559 And it puts further pressure on me to want to use these tools and outsource even more of my thinking. 00:14:30.799 --> 00:14:43.200 And then you combine that with the fact that detection is really difficult with AI systems in the sense it's very hard for me to figure out if you're using those glasses or those pendants, unless I really kind of press further. 00:14:43.440 --> 00:14:54.720 And I think when you think about the situation that people are put in, where it's a strong biological temptation, there's huge economic rewards in terms of productivity that can be accomplished if you use it well. 00:14:54.960 --> 00:14:58.080 You are not really penalized or detected if you use it. 00:14:58.240 --> 00:15:02.879 And it just creates this ecosystem in which it's very hard to resist. 00:15:03.039 --> 00:15:09.360 And so one of my key messages in the pull-up is that not enough to say to people, just go think for yourself. 00:15:09.519 --> 00:15:18.480 I think it really doesn't acknowledge or give credit to the fact that there are many societal systems that are putting emphasis on people to use it. 00:15:18.799 --> 00:15:26.000 And that's really where the book tries to urge people to think about system design as a means to address some of these issues. 00:15:26.240 --> 00:15:31.120 It's not about wagging your finger or having a stern lecture with young people in your life. 00:15:31.279 --> 00:15:43.679 It's about recognizing the situation they're in and helping them to develop strategies where they can use AI, as is going to be increasingly expected, but to use so in a way that preserves what makes us valuable. 00:15:44.000 --> 00:15:48.080 So I would love to touch on that systems design a little bit more, Eric. 00:15:48.240 --> 00:16:00.960 Because on one end, I have been encouraging, and a lot of my guests talk about how to get people in the organization to actually embrace AI and use AI more. 00:16:01.200 --> 00:16:03.519 You talk about competitive ratchet effect. 00:16:03.840 --> 00:16:10.080 So there are organizational pressures, there are pressures from everyone around us. 00:16:10.320 --> 00:16:16.720 But you are saying that we need to think about systems design to make sure it's done well. 00:16:17.039 --> 00:16:19.200 I think I would say a couple pieces to it. 00:16:19.440 --> 00:16:29.679 One is that I think any reward system that incentivizes that sort of measures people based on how much they use AI is destined to create this problem. 00:16:29.759 --> 00:16:32.480 That basically you just outsource more of your thinking to AI. 00:16:32.639 --> 00:16:33.759 It's a simple tenet of life. 00:16:34.080 --> 00:16:37.440 The whole concept of economics in a nutshell is people respond to incentives. 00:16:37.600 --> 00:16:41.840 And so when you give people incentives to use AI more, naturally that emerges. 00:16:42.000 --> 00:16:43.360 What's the counter proposal? 00:16:43.600 --> 00:16:55.919 The counter proposal is to measure and reward people for the things that you really care about that they can accomplish with AI, but that place priority on what makes you valuable. 00:16:56.080 --> 00:17:07.519 So for example, if I create an incentive system where it's just the number of presentations you can generate, you just upload these to AI, it could generate a hundred different presentations in any given day. 00:17:07.680 --> 00:17:34.000 But if you create incentive systems where it's not just about the presentation, but it's effectiveness of the delivery, the ability to generate sales or convert those to actual revenues, then I think it forces people to deal with this reckoning where just having AI generate these things for you is not what's going to get you ahead because you're going to miss out on the other pieces outside of the presentations, which are really valuable. 00:17:34.240 --> 00:17:44.480 And I would also add to this that I think what I've noticed is that the people use AI often to outsource the things that really make them valuable. 00:17:44.640 --> 00:17:53.680 So in the book, I talk about the fact that it's useful to really think carefully about what are the things that define your value as a thinker. 00:17:53.839 --> 00:17:56.079 I call them signature thinking skills. 00:17:56.160 --> 00:18:01.839 So maybe it's a lawyer's like rhetorical ability or sort of a physician's

diagnostic eye. 00:18:02.079 --> 00:18:09.519 And to really be thinking about designing your work with AI in a way that preserves those signature thinking skills. 00:18:09.680 --> 00:18:21.039 I'm all in favor of using AI for the parts of my work, which may be rote or like unimportant for the purposes of doing my signature work, but not to outsource everything. 00:18:21.200 --> 00:18:26.000 And it really makes a difference in terms of which things you decide to outsource to AI. 00:18:26.160 --> 00:18:35.200 So part of what I encourage people to do is to engage in an exercise where they think about what are the things that they want to have as signature thinking skills. 00:18:35.359 --> 00:18:40.640 And to me, if you work backwards, it writes the rules for how you should be using AI. 00:18:40.799 --> 00:18:49.039 Because if you're using AI to do those tasks that are like really that cultivate signature skills, then you're really just risking and losing those things. 00:18:49.200 --> 00:18:52.400 And I think it clearly poses a threat to our value. 00:18:52.880 --> 00:19:02.400 You draw an important distinction between what we can accomplish with AI and what we remain capable of doing without it. 00:19:02.720 --> 00:19:09.759 So what are the types of signature thinking skills that we should focus in on? 00:19:10.160 --> 00:19:12.400 I think it varies a lot by profession. 00:19:12.559 --> 00:19:29.759 So, in the sense that myself as a professor, one of the things that I really value in myself is in my ability to understand the connections between different streams of research and how they fit together, the sort of the pieces they tell us, and you know, but also the gaps between them. 00:19:30.000 --> 00:19:33.440 But if I were to have AI synthesize all those papers. 00:19:33.519 --> 00:19:43.200 So like I go to like notebook LM and have it generate a podcast, it's really removing myself from that process of kind of investigating or doing the deep synthesis. 00:19:43.440 --> 00:19:56.160 And like many of my colleagues have reported, they found that has weakened their ability to do that critical work, the thing, identifying the right questions, pulling the pieces together to be able to articulate that to other people. 00:19:56.400 --> 00:19:58.720 And so that's an example from my own work. 00:19:58.880 --> 00:20:04.400 But you can imagine that for many professions, there are critical things that are important and things that are not. 00:20:04.640 --> 00:20:15.279 For example, in my work, I think it's really important to be able to figure out what are the right questions to ask, and how do you design, for example, an experiment to test that. 00:20:15.440 --> 00:20:23.119 But what's less important to me are things that, like maybe searching and retrieval, like kind of finding papers that might be relevant. 00:20:23.200 --> 00:20:26.160 I think I have no problem outsourcing those things to AI. 00:20:26.400 --> 00:20:29.599 The same thing is true for like

certain coding tasks. 00:20:29.759 --> 00:20:37.039 It's really important that I know what I want to code, how to evaluate the output and how to structure tests. 00:20:37.200 --> 00:20:42.240 But that's different than the simple mechanics of sitting down to actually write out that computer code. 00:20:42.400 --> 00:20:45.920 I don't think of that as being very critical to who I am. 00:20:46.160 --> 00:20:52.480 Now, a software developer may feel differently, but that's what I mean by it's very situational or occupation specific. 00:20:52.960 --> 00:20:57.039 It takes a lot of intentionality to do what you're talking about, Eric. 00:20:57.200 --> 00:21:05.200 And I use the simple example of you said you grew up in Silver Spring, pretty close to where I live. 00:21:05.359 --> 00:21:14.079 I grew up in the DC area in Maryland, Montgomery County, and how I use Waze quite often. 00:21:14.319 --> 00:21:18.960 A few times I didn't use Waze, I got stuck in traffic jams. 00:21:19.359 --> 00:21:24.559 So, in essence, I learned from the AI that I better follow the AI. 00:21:24.799 --> 00:21:34.880 Now I follow the AI a lot, to the point that if I don't pay attention, sometimes I drive my daughter for volleyball to new areas. 00:21:35.440 --> 00:21:39.599 If you ask me to drive there without ways, I wouldn't be able to. 00:21:39.839 --> 00:21:53.839 Without being focused on what you're talking about, it's pretty easy for us to defer all of the decision making to the AI rather than the struggle of the thinking. 00:21:54.319 --> 00:21:55.440 Yeah, I think that's right. 00:21:55.920 --> 00:21:59.680 I think that AI presents us with a trade, and like any other trade. 00:22:00.480 --> 00:22:06.240 The sort of the appropriateness of that trade depends entirely on what you're giving up and what you're receiving. 00:22:06.480 --> 00:22:09.599 I think with GPS, I think it offers a lot of convenience. 00:22:09.759 --> 00:22:12.960 There's an argument to be made, it offers a great deal in terms of safety. 00:22:13.119 --> 00:22:18.720 And so I understand why people make this trade-off and why there may be less of a concern. 00:22:19.119 --> 00:22:28.000 I think, though, in the context of our intellectual work, I think it's just too critical to offload and to risk-threatening. 00:22:28.079 --> 00:22:36.240 It's too important to our functioning and society and who we are and our value that it doesn't present a palatable trait to me. 00:22:36.400 --> 00:22:43.039 And you're absolutely right that you have to be intentional about it simply because otherwise it invites a lot of short-term thinking. 00:22:43.200 --> 00:22:47.039 I just offload something to AI, and in the short term, it does work. 00:22:47.200 --> 00:22:51.680 But cumulatively, those hacks really add up to threaten to who we are. 00:22:51.920 --> 00:22:57.759 And oftentimes people don't realize it until they're in a moment where suddenly they can't access the technology. 00:22:57.839 --> 00:23:07.119 They don't have access to the AI chat bug, their AI pendants or glasses are not working, and where they really feel the pinch of having not gone through the struggle. 00:23:07.519 --> 00:23:11.359 As you also say, it's hard. 00:23:11.519 --> 00:23:16.480 It goes against some of our human instincts and human desires. 00:23:16.799 --> 00:23:21.599 Yeah, in the book, I talk about different rules or principles for thriving in this era. 00:23:22.079 --> 00:23:32.480 One of them, the first, is to value the struggle in the sense that, as we alluded to earlier, struggle in our society is often treated as a problem to be solved. 00:23:32.559 --> 00:23:35.200 It's often the basis of products that we are sold. 00:23:35.359 --> 00:23:37.920 But the reality is the struggle is formative. 00:23:38.079 --> 00:23:41.359 You may have heard the saying, there is no learning without struggle. 00:23:41.519 --> 00:23:46.079 And I think it's in part requires a reframe of how we think about struggle. 00:23:46.240 --> 00:23:49.759 And if you think, if you actually look back in history, there's precedent for this. 00:23:49.920 --> 00:23:59.599 So post-industrial revolution, our society had to be reformulated in the sense that you no longer needed to be physically active to be economically productive. 00:23:59.759 --> 00:24:04.559 And the implication of that is that many people saw their physical bodies atrophy. 00:24:04.880 --> 00:24:11.599 And what this required was an intentionality to develop physical exercise as something you do deliberately. 00:24:11.839 --> 00:24:17.599 It's crazy to think about in modern society, but there used to be a period where people did not exercise on purpose. 00:24:17.759 --> 00:24:19.200 It was something they were paid to do. 00:24:19.359 --> 00:24:20.799 But we understand this now. 00:24:20.960 --> 00:24:26.319 We go to the gym, we take the stairs, we feel the burn, we realize we're doing something good for ourselves. 00:24:26.480 --> 00:24:31.039 And I think that's part of the reframe that this technological shift is asking of us. 00:24:31.440 --> 00:24:46.640 Now, one of the things that you mention that counters the way a lot of people talk about AI, Eric, is that typically we've been talking about great human plus great AI. 00:24:47.039 --> 00:24:51.680 So that augmentation means much better performance. 00:24:52.000 --> 00:24:54.640 But you call that dream team fallacy. 00:24:55.119 --> 00:25:04.160 Yeah, the dream team fallacy conjures images that if I bring elite people and elite technology and I put them together and I get incredible results. 00:25:04.400 --> 00:25:06.960 But the data often tells us something different. 00:25:07.200 --> 00:25:15.359 And the reason is that technology and our human brains, they're not additive in the way that you might hope, or even multiplicative. 00:25:15.599 --> 00:25:17.039 They're really kind of reactions. 00:25:17.200 --> 00:25:25.759 And there's so many evolutionarily wired aspects of our thinking

that makes us often worse off when using the technology. 00:25:25.920 --> 00:25:27.039 So what's an example? 00:25:27.279 --> 00:25:33.759 A simple example is that when you work with an AI system, they often provide you with a recommendation. 00:25:34.000 --> 00:25:45.599 But because our brains are wired to engage in shortcuts, we often anchor on the first thing that the AI presents us without critically evaluating whether it's true. 00:25:45.920 --> 00:25:55.119 And what that often does is it removes the value of me being involved in the first place, if I simply have, if I just latch onto the first thing the AI says. 00:25:55.359 --> 00:26:00.079 Additionally, there is a tendency of us in becoming complacent. 00:26:00.160 --> 00:26:08.319 Sometimes they call this complacency bias, is that if I work with an AI system that gets good enough, I stop bothering to check if it's right at all. 00:26:08.480 --> 00:26:11.200 And so I remove my thinking from the process. 00:26:11.359 --> 00:26:14.799 And so you're not really getting this like elite human anymore. 00:26:14.880 --> 00:26:18.960 You're getting the distracted or thoughtless human combined with the AI. 00:26:19.119 --> 00:26:23.039 And that removes the value of having a person there altogether. 00:26:23.279 --> 00:26:36.240 And so one of the striking conclusions of my research and sort of this broader field is that contrary to a lot of people's intuition, actually embedding simple things like friction in your interface with the AI can help a lot. 00:26:36.400 --> 00:26:37.599 And so, what do I mean by that? 00:26:37.759 --> 00:26:42.400 Is that oftentimes the models, if you ask them a question, will immediately give you an answer. 00:26:42.480 --> 00:26:44.799 And the human brain naturally anchors on that. 00:26:44.880 --> 00:27:04.400 But instead, if the model either pauses for you to think, it forces you to think first, or if it prompts you with additional questions that can help you really get to the substance of the issue, not only do you get better outcomes, but you get retention in the sense that because I struggled through it, I'm much more likely to remember it afterwards. 00:27:04.640 --> 00:27:09.759 And so that's not what we think of typically in terms of a dream team, slapping the pieces together. 00:27:09.920 --> 00:27:21.039 But it turns out the evidence shows us we have to be very thoughtful as educators, as parents, as leaders in designing interfaces that are going to bring out the best of both the technology and the human. 00:27:21.359 --> 00:27:25.200 Eric, I've now become more hopeful as I have listened to you. 00:27:25.359 --> 00:27:34.960 I had read the book, and I'm now becoming more hopeful, both thinking about the analogy for the exercise that you mentioned. 00:27:35.119 --> 00:27:40.160 And my undergraduate degree, actually, before going to business school, was in human nutrition. 00:27:40.240 --> 00:27:43.759 And then I had done some graduate research in human nutrition as well. 00:27:44.079 --> 00:27:54.160 Our human biology is suited to want fat, salt, and sugar, which is why we crave it in the environment. 00:27:54.480 --> 00:28:10.079 However, as we have become more aware of the impact of it, to a certain extent, we exercise control over how much fat, sugar, and salt we consume trying to be healthy. 00:28:10.319 --> 00:28:36.319 So it's by being aware of what our tendencies are, what to a great extent some of these providers benefit from, which is engaging us on an ongoing basis, then we can add friction and we can do things in a way that benefits us rather than benefits them or detracts from our capabilities and abilities. 00:28:36.720 --> 00:28:38.000 Yeah, I completely agree. 00:28:38.240 --> 00:28:58.319 On the point about usage, one of the things I really worry about and talk about in the book is this sort of downward spiral that I sometimes see where someone begins to use AI, and slowly, because they're using AI, they're not practicing that particular task, and they lose fluency or they lose proficiency with that task. 00:28:58.559 --> 00:29:07.680 And as a result, we become less comfortable doing the task without AI, which in turn makes us reach for AI even more, and then just kind of spirals. 00:29:08.000 --> 00:29:21.920 And I've seen in presenting this book at various audiences, I've had so many people, both business leaders, but also parents, come to me and just say, I've seen AI take a toll and where people can no longer think without it. 00:29:22.000 --> 00:29:24.079 They just become dependent on it. 00:29:24.319 --> 00:29:32.000 And I think it's in part the book is trying to raise awareness of this sort of potential downward spot if we don't step in. 00:29:32.319 --> 00:29:42.799 It also raises a point that you alluded to, which is that oftentimes in the moment, we're poor judges for what we really understand and what we own. 00:29:42.960 --> 00:29:46.400 There's this notion of processing fluency that I talk about in the book. 00:29:46.640 --> 00:29:52.960 You see someone deliver this polished lecture, or you watch a YouTube video and suddenly you feel like you understand it. 00:29:53.200 --> 00:30:01.519 But when you actually task those same people with trying to explain it or answer questions about this struggle, AI presents us with something very similar. 00:30:01.680 --> 00:30:07.039 We feel as though we understand it because AI explained it very polished in a polished way to us. 00:30:07.200 --> 00:30:14.799 But when you look at the data, it shows us that oftentimes we mistake the ease of us ingesting it with our real command in it. 00:30:14.960 --> 00:30:18.160 And that's one of the things I'm trying to raise awareness of as well. 00:30:18.480 --> 00:30:21.359 You do an outstanding job of that, Eric. 00:30:21.519 --> 00:30:27.680 As a fun aside, uh a couple of weeks ago, I got an email as a podcaster. 00:30:27.759 --> 00:30:43.519 You get pitched all kinds of things, but I got

pitched on an AI tool which listens in real time to the podcast conversation and gives you questions based on what the guest said. 00:30:43.759 --> 00:30:47.200 I can just step aside and the AI interview you. 00:30:47.440 --> 00:30:51.119 Yeah, I think that's really what I'm worried about broadly. 00:30:51.359 --> 00:30:54.640 And it's not, I want to put it just about capabilities. 00:30:54.880 --> 00:31:05.039 Part of what I talk about in the book is how AI is changing our sort of view of the world, our understanding of what is truth, and our ability to relate to others. 00:31:05.200 --> 00:31:12.000 So there's a lot of research, just fascinating research about how AI changes our interactions with other humans. 00:31:12.160 --> 00:31:22.880 One of the central tendencies of AI, because it is trained to get your approval, those little thumbs up buttons, like when you work with ChatGPT, it really wants to make you happy. 00:31:23.039 --> 00:31:26.319 And so that means that the models are programmed to be sycophantic. 00:31:26.559 --> 00:31:30.480 More likely than not, if you suggest an idea, it's going to tell you it's a great idea. 00:31:30.799 --> 00:31:38.400 But this kind of creates this new type of echo chamber, a personalized one that I think takes what we've seen with social media to another level. 00:31:38.559 --> 00:31:44.240 And it surrounds ourselves with an AI system that just naturally says that everything we do is great. 00:31:44.480 --> 00:31:55.200 And where I've seen this really come into play is imagine you have two business partners or spouses, and they are in some sort of disagreement, and they both go to their AIs and say, Here's my position. 00:31:55.279 --> 00:31:58.000 And the models say, Oh, I can't believe that other person. 00:31:58.079 --> 00:31:59.279 They're completely wrong. 00:31:59.519 --> 00:32:04.720 And what happens is they come back to the table, even more entrenched in their original thinking. 00:32:04.880 --> 00:32:15.759 And so it's not only that it can send us into sort of this like more of an echo chamber, it creates divides between other people in the sense that I think I'm right, I think you're wrong. 00:32:15.920 --> 00:32:19.920 But I also just mentally really like that AI is always kind to me. 00:32:20.000 --> 00:32:21.200 It says nice things to me. 00:32:21.359 --> 00:32:25.920 So I don't like talking to the human who disagrees or tells me my ideas are bad. 00:32:26.240 --> 00:32:27.200 That's true. 00:32:27.359 --> 00:32:49.359 And if you look at sites like Character AI, Replica, which is much more romantic relationships, and just see some of the testimonials people have of the relationships they've established with AI, you can see why, to a certain extent, we then lose the capability to deal with another human where there is friction. 00:32:49.599 --> 00:32:59.119 It doesn't matter how happily married you are, there's always friction in that no matter how effectively the team works with each other.

00:32:59.359 --> 00:33:05.200 So there are capabilities we need to maintain to be able to interact with other humans. 00:33:05.519 --> 00:33:06.640 Yeah, absolutely. 00:33:06.960 --> 00:33:19.440 You may have seen in headlines recently that China is putting in policies in place to crack down on AI romantic partners in part because they're so worried about this problem that they're concerned about population growth. 00:33:19.759 --> 00:33:30.640 And I think it's a simple reflection of the fact that AI is so tempting to us, not only to outsource our thinking, but to use it as a means to fill our need for social engagement. 00:33:30.880 --> 00:33:37.200 I talk about this in the book, but there's all this fascinating research about what goes on in our brain when we interact with these systems. 00:33:37.440 --> 00:33:57.839 And it turns out that flattery that we get from AI, even if we are very rational about it and we know that it's a machine and it's flattering us, it just tends to trigger this chemical reaction in our brain in a way that is almost unmistakable or undistinguishable from real human interaction with the where we get genuine flattery. 00:33:58.000 --> 00:34:01.119 And I think this kind of invites us to keep coming back for more. 00:34:01.200 --> 00:34:07.200 It only makes us more addicted, and then we turn to it not just for personal things, but intellectual things. 00:34:07.440 --> 00:34:12.639 And this is really, I think, where people can become just taken over by AI if they're not careful. 00:34:12.960 --> 00:34:20.719 It's hilarious because ChatGPT compliments me when I ask it to disagree with me. 00:34:20.960 --> 00:34:22.800 So you're absolutely right. 00:34:23.039 --> 00:34:28.159 Where the way the interactions work, on one end, you laugh at it. 00:34:28.320 --> 00:34:31.760 On the other end, as you just made the point, we like that. 00:34:31.920 --> 00:34:33.599 As humans, we like that. 00:34:33.920 --> 00:34:36.239 So there is something for us to be aware of. 00:34:36.400 --> 00:34:42.400 But what I wonder, Eric, is you also used AI to write this book, and you mentioned that. 00:34:42.639 --> 00:34:52.320 So how do you recommend for people to approach both what they use AI for, also how they use it? 00:34:52.639 --> 00:34:57.360 Yeah, in the book, I talk about four sort of general rules or principles. 00:34:57.599 --> 00:35:01.280 One is to value the struggle, which is a mentality shift that we talked about. 00:35:01.599 --> 00:35:05.519 The second is to value who you are without AI. 00:35:05.599 --> 00:35:18.079 And what that means is recognizing that many important life moments happen outside of the digital sphere, and that we need to preserve those capabilities in order to function when the technology fails. 00:35:18.320 --> 00:35:23.039 A third, which speaks more to your point, is to reinvest your cognitive surplus. 00:35:23.280 --> 00:35:31.599 So what oftentimes I see people do is they'll use AI as simply a means to do the same things, but just a little bit faster or without thinking. 00:35:31.840 --> 00:35:40.000 But if you look back in history, the most transformative technologies were not about automating processes or doing things faster. 00:35:40.239 --> 00:35:42.880 They were for unlocking new human capabilities. 00:35:42.960 --> 00:35:49.199 Like you think about the airplane or the microscope, these things allowed us to chart new territory and thinking. 00:35:49.360 --> 00:35:51.840 And that's part of what I advocate with AI. 00:35:52.000 --> 00:36:01.920 The time that AI saves you for rote tasks is to use it to continue to grow intellectually, to challenge your thinking and to push yourself further to take on new tasks. 00:36:02.239 --> 00:36:06.079 The fourth rule is to make AI your cognitive trainer. 00:36:06.480 --> 00:36:07.679 What does that really mean? 00:36:07.920 --> 00:36:13.599 It means that AI is this incredible tool for learning if you use it properly. 00:36:14.079 --> 00:36:23.280 In the sense that AI can be personalized, it can be a useful tutor, but it requires the discipline to not use it as a simple answering machine. 00:36:23.679 --> 00:36:29.039 There's this long-standing notion in education research called the two sigma problem. 00:36:29.280 --> 00:36:43.199 And what that refers to is this very persistent gap where personalized tutoring often gives people educational outcomes that are two standard deviations better than the bulk class where everyone gets the same material. 00:36:43.360 --> 00:36:48.559 So if you have a really dedicated tutor, they can meet you where you are, you're just far ahead of others. 00:36:48.719 --> 00:37:07.039 And what research shows is that when you use AI with the right scaffolding, where it's a tutor, not an answering machine, it can meet you where you are, it allows people to achieve these remarkable gains in learning, in educational productivity, but it requires the discipline to not just take the easy way out. 00:37:07.199 --> 00:37:14.400 And so I talk about these four rules throughout the book as a means to frame how we go about approaching this era with AI. 00:37:14.559 --> 00:37:23.679 Some of that is about individual agency, some of that is about designing policies that don't encourage people to just use AI for the sake of using AI. 00:37:23.920 --> 00:37:34.320 And some of that is about designing interfaces between humans and AI that they sort of pay tribute to this dream team myth that we talked about earlier. 00:37:34.480 --> 00:37:47.599 That's a fallacy, that oftentimes we need to embed friction into these systems just to keep ourselves in the mix, which feels very silly, like of adding friction into something intentionally, but that turns out to make all the difference. 00:37:47.920 --> 00:38:19.519 So I wonder, Eric, I was in a meeting a couple of months ago where there was a high-level person involved in AI in what is now Department of War, was formerly Department of Defense, and the

person was making the case for why there shouldn't be human involvement or human in the middle, in part because of speed of decision making and speed of action. 00:38:19.760 --> 00:38:35.920 And as I'm sure you're also aware, that's where Anthropic got in trouble with Department of War, Department of Defense, where they wanted a human in the middle for certain decisions, and Department of War didn't want that. 00:38:36.400 --> 00:39:00.159 So would love to get your thoughts on that because the argument the person was making, and the same argument could be made in a lot of organizational decision making, is that the speed and accuracy of decision making of your adversaries can be drastically higher without the human in the loop. 00:39:00.320 --> 00:39:02.159 So what do you think about that? 00:39:02.639 --> 00:39:09.679 I think it really speaks to the broader AI gravity and the competitive landscape that we face. 00:39:09.920 --> 00:39:14.639 If you listen to AI leaders to government leaders, they'll talk about this arms race. 00:39:14.800 --> 00:39:31.280 You may have seen at the 2026 World Economic Forum, Dennis Sasabas was on stage with Dario Amadei, and they said, gosh, I wish we could slow down model development because it's happening so fast, but we can't because China is developing the next model where adversaries are developing. 00:39:31.519 --> 00:39:38.400 And it really speaks to this idea that there's pressure to develop and use these systems just as a means to keep afloat. 00:39:38.639 --> 00:39:50.639 I think in the context of decision making, in the context of war, gosh, that's such a difficult ethical issue to tackle and one that I don't claim to have easy answers. 00:39:50.800 --> 00:40:07.679 But I think there are so many instances in which sort of human ethics and judgment are necessary to make the right decision that I would be very leery of removing AI unless it was like removing humans from that process, unless that was really the only possibility. 00:40:08.079 --> 00:40:09.440 It is very hard. 00:40:09.679 --> 00:40:21.440 You also give example, this is pre-AI, but still it's automated systems, USS Vincent's and issues there where they shot down a passenger airliner. 00:40:21.679 --> 00:40:38.000 There are cases where F-16s with AI pilots outperform the best fighter pilots because the AI can more quickly adjust, adapt, and come up with decisions in a dog fight. 00:40:38.159 --> 00:40:53.039 So there are a lot of challenges around this, but it is an issue that is a struggle where on one end you have people that say others are going full force ahead, whether it's with military or it's organizationally. 00:40:53.519 --> 00:40:55.440 Therefore, we have no choice. 00:40:55.599 --> 00:41:00.559 And some say that we do need to have that human judgment involved. 00:41:00.880 --> 00:41:09.280 Now, you spend much of your book, Eric, focused on how AI impacts an individual's mind. 00:41:09.679 --> 00:41:18.239 Most of my audience, they

are executives, they are individuals, as well as leaders of organizations. 00:41:18.800 --> 00:41:27.440 What happens in your view when this kind of cognitive outsourcing occurs across an entire organization? 00:41:27.840 --> 00:41:28.000 Yeah. 00:41:28.400 --> 00:41:29.360 Thank you for raising this. 00:41:29.519 --> 00:41:33.039 I think that I would say a number of things to organizations. 00:41:33.360 --> 00:41:40.880 One is that it really threats to hollow out many aspects of continuity within organizations. 00:41:41.039 --> 00:41:46.800 So, for example, some people are turning to AI instead of hiring junior employees. 00:41:47.039 --> 00:41:55.599 And the problem with that, of course, is that if you don't bring in junior employees to do the quote unquote club work, struggle through problems, that's how they develop expertise. 00:41:55.679 --> 00:42:03.280 But that's also how they become the next person, the manager or executives who has the experience to understand the particular context. 00:42:03.440 --> 00:42:07.440 And so I really worry about this being a short-term solution. 00:42:07.679 --> 00:42:18.480 It's true, you can save money potentially by not hiring as many junior employees, but at the threat of preserving a lot of information which is not actually codified or written down. 00:42:18.719 --> 00:42:25.840 A lot of an organization is like this living memory, and it's not necessarily in manuals or PDFs or spreadsheets. 00:42:26.000 --> 00:42:30.400 It's in the dialogue, it's in the conversation, it's in the minds of employees. 00:42:30.639 --> 00:42:38.400 And the problem is if people are not developing that expertise, they don't have the ability to carry it forward or to pass it down to the next generation. 00:42:38.719 --> 00:42:55.920 It's also the case that when you introduce AI systems, there's this really subtle thing that happens, which is previously, if I were your coworker, I would bring to the table my set of expertise and you would bring your set of expertise, and we would essentially engage in a trade. 00:42:56.079 --> 00:42:58.719 I would help you with some aspects and you would help me with others. 00:42:58.880 --> 00:43:04.239 And we develop a sort of a shared bond, a trust within an organization as a team. 00:43:04.480 --> 00:43:08.079 But increasingly, I don't turn to you to fill my sort of gap. 00:43:08.239 --> 00:43:09.360 I turn to the AI. 00:43:09.519 --> 00:43:19.840 And what that means is we lose a lot of that connection that forms trust as the basis for partnership and loses a lot of that communication that happens along with it. 00:43:20.000 --> 00:43:24.719 And so it reminds me a little bit of the evidence of what happens from work from home. 00:43:24.880 --> 00:43:30.880 A lot of times people they can get a lot of work done because they don't have to commute, but they miss that interaction with other people. 00:43:31.039 --> 00:43:36.880 They miss the information sharing, the connection, the trust that I

think is can be very important for organizations. 00:43:37.360 --> 00:43:43.519 So I think what I would say is that you're right, we shouldn't just be thinking about AI on an individual level. 00:43:43.679 --> 00:43:54.800 We should be thinking about what it does to us in terms of our teamwork, our trust, our culture, but also in terms of not just where we are today, but where the organization is going in the future. 00:43:54.960 --> 00:43:59.760 And I think smart leaders will recognize that and take actions now rather than just take. 00:44:00.159 --> 00:44:03.840 Taking the short-term benefit of saving costs of AI. 00:44:04.480 --> 00:44:18.079 I repeatedly say, Eric, and mention it to my clients as well that without that human judgment, if you purely rely on the AI, that can be duplicated instantaneously. 00:44:18.320 --> 00:44:21.360 There is no competitive advantage to that. 00:44:21.599 --> 00:44:27.440 For a month or two, you could be ahead, but that's it at best, at most. 00:44:28.320 --> 00:44:43.679 Therefore, you do need to focus a lot of effort on the parts where human judgment adds value and you need to maintain, and sure, on the parts that AI can effectively do separately. 00:44:43.840 --> 00:44:54.800 So that is part of the unique challenge that we have as we go through this transformation, which I believe is very different than what we have experienced before. 00:44:55.119 --> 00:45:02.480 Now, I want to find out a little bit more about your practices and the way you approach AI, Eric. 00:45:02.719 --> 00:45:11.119 First, out of your four rules that you went through yourself, which ones do you find hardest to follow? 00:45:11.840 --> 00:45:22.239 I think it's probably the rule four to use AI as your cognitive trainer in the sense that there are a lot of topics that I would like to engage and learn more about. 00:45:22.639 --> 00:45:27.519 I think, though, AI is so tempting as an answer machine, as a quick shortcut. 00:45:27.920 --> 00:45:40.880 And so rather than me asking the AI to be in tutor mode to help me understand something like quantum physics, the temptation is so strong to just say, explain to me quantum physics in two sentences or something along those lines. 00:45:41.199 --> 00:45:51.440 In doing so, I'm being unfaithful to this value, the struggle point, that you really have to engage in friction of the process of learning to encode it, to develop that type of mastery. 00:45:51.679 --> 00:45:55.119 So that's definitely a part that I catch myself doing a lot. 00:45:55.280 --> 00:46:02.559 And what I've noticed is that most of the leading frontier labs, like Anthropic and ChatGPT, they offer a tutor mode. 00:46:02.639 --> 00:46:06.159 You can enable them where they don't and give you instant answers. 00:46:06.320 --> 00:46:10.639 But it requires you to actually use it and to turn it on. 00:46:10.880 --> 00:46:27.199 And more for parents and for educators, what I've noticed is that if you give people a system where they have tutor mode built in, the temptation then is to reach for their phone and then open up the web browser to go to ChatGPT, wherever that doesn't have the tutor mode in place. 00:46:27.440 --> 00:46:31.760 And so I think this is in part why I think we need a societal reframe. 00:46:31.840 --> 00:46:35.440 We need to understand sort of ethics and value in using these systems. 00:46:35.679 --> 00:46:42.159 And we need to have and teach people about this in the same way that we've taught people about responding to past revolutions. 00:46:42.480 --> 00:46:44.159 It does take effort, Eric. 00:46:44.239 --> 00:46:57.679 I've used Khan Academy with my daughters, and the tutor mode, as you're saying, becomes very frustrating for someone who is used to typical AI answers. 00:46:57.920 --> 00:47:03.119 Because the tutor mode is brilliant in that it wants to get you to think. 00:47:03.519 --> 00:47:06.000 But as you've mentioned, we want to skip that. 00:47:06.159 --> 00:47:07.039 Give me the answer. 00:47:07.199 --> 00:47:08.480 Give me the answer. 00:47:08.800 --> 00:47:13.119 Assuming if we get the answer, we will learn, but that's not how we learn. 00:47:13.360 --> 00:47:22.320 So that's a great practice for us all to use, including as you mentioned, whether with GPT or Claude. 00:47:22.719 --> 00:47:38.960 Now I wonder if we're five, 10 years from now, Eric, what do you hope that we as humans are better at because of AI rather than merely better at doing with AI? 00:47:40.159 --> 00:47:57.519 What I really hope is that AI is a big unlock for us, that it removes frictions and impediments that stopped us from pursuing our intellectual interest, our artistic pursuits, that it reduces barriers to allow us to reach for more. 00:47:57.679 --> 00:48:01.679 I think that's where we've seen a lot of the transformative change with past technologies. 00:48:01.840 --> 00:48:07.280 And so it's not simply about producing more of the same, it's about reaching for higher aspiration. 00:48:07.519 --> 00:48:16.239 And clearly, the examples that the IA companies often talk about is curing disease or preventing mass outbreaks or something along those lines. 00:48:16.400 --> 00:48:19.840 And I think that's really where I see a lot of the promise. 00:48:20.000 --> 00:48:22.880 That's not necessarily how I see people using it now. 00:48:23.039 --> 00:48:35.039 But my hope is that by reducing some of those barriers, it can just unlock what is so special about humanity, our creativity, our sort of understanding of parts, our connections to others. 00:48:35.199 --> 00:48:39.360 And that, of course, requires more than one thing to get us there. 00:48:39.679 --> 00:48:41.280 It's going to take a lot of effort. 00:48:41.360 --> 00:48:59.199 And again, connecting it to one of my passions, which is, I mentioned nutrition, is that processed foods go through the system a lot faster and cause a lot of issues that now we have become a lot more aware of. 00:48:59.440 --> 00:49:06.719 The

processing of food doesn't help the human digestive system at all and causes issues. 00:49:06.960 --> 00:49:13.760 This is the same way in that we can have AI process the thoughts for us and give us the answer. 00:49:14.000 --> 00:49:24.239 That does not produce the kind of results that we want, whether in terms of our cognitive capabilities or the outcomes. 00:49:24.480 --> 00:49:38.800 So that's why I appreciate you raising the issue, Eric, because I do see all around me, including in organizations, just a couple of months ago, I had a CEO of a major telecom saying that you know what? 00:49:39.119 --> 00:49:47.519 I now have executive VPs present beautiful PowerPoint decks on strategy to me. 00:49:47.760 --> 00:49:54.960 When I ask them a few questions, they have very little idea on the point that they're making on the PowerPoint. 00:49:55.199 --> 00:49:58.800 So this is not just an issue that some students are having. 00:49:58.960 --> 00:50:06.480 This is an issue all across society, including some of the most senior executives in organizations. 00:50:06.880 --> 00:50:13.599 So for my audience to find out more about your book and follow your work, where would you send them to, Eric? 00:50:14.079 --> 00:50:17.599 My book is available on all the major retailers like Amazon. 00:50:17.679 --> 00:50:20.159 It's the collision, what AI Does To Us. 00:50:20.400 --> 00:50:27.280 I would also recommend that people check out my website, which is ericsow.pro, where pro is short for professor. 00:50:27.519 --> 00:50:35.599 And on my website, you'll find a workbook where I provide people with exercises to try to make the rules that I talk about in the book tangible. 00:50:35.840 --> 00:50:46.880 I've led a number of organizations through this exercise, and it's meant to really force people to design how they work with AI in a way that will make them better off, not just faster. 00:50:47.679 --> 00:50:49.199 I agree with that, Eric. 00:50:49.280 --> 00:50:51.840 I am a big advocate for AI use. 00:50:52.639 --> 00:50:55.840 But we need to use it properly. 00:50:56.239 --> 00:51:01.280 So that's why your insights are of tremendous value. 00:51:01.440 --> 00:51:04.559 And I so appreciate your book as well. 00:51:04.719 --> 00:51:10.320 Thank you so much for the conversation on the collision, what AI does to us. 00:51:10.559 --> 00:51:12.239 Thank you for joining me, Eric So. 00:51:13.679 --> 00:51:14.880 Thank you so much for having me. 00:51:14.960 --> 00:51:15.840 It's been a real pleasure. 00:51:15.920 --> 00:51:18.239 I really appreciate you engaging with the topic. 00:51:18.320 --> 00:51:19.679 And you've been a wonderful host. 00:51:19.920 --> 00:51:20.719 Thank you. 00:51:21.840 --> 00:51:26.159 You have been listening to Partnering Leadership with your host, Mahan Tavakoli. 00:51:26.719 --> 00:51:42.239 If you enjoyed this episode, please leave a rating and review of the podcast on your favorite podcasting app and forward the conversation to a friend or colleague so you can help more people discover their purpose, grow professionally with meaning, and have a greater impact. 00:51:42.480 --> 00:51:48.639 For additional leadership insights and bonus content, visit us at partneringleadership.com.