Your ‘AI crisis’ is totally normal but you’ll be fine, with Ethan Mollick

0:00Starting a business comes with its share of ups and downs, which is why staying true to your vision is essential. a non-negotiable for Romeo and Milka Bregali, Capital One business customers and co-owners of Ra's plant-based restaurant in New York. Romeo and Milka took a leap of faith when starting their own restaurant, gutting an empty space and building it from the ground up, every pipe, every wall, every detail. But building from scratch came with a heavy financial burden, which is when they turned to their Capital One business card. With the flexibility of the card's no preset spending limit, they were able to spend more and earn more rewards while bringing their vision to life.

0:36Today, Raz's success is proof that with passion and the right support, it's possible to make your dreams a reality. Learn more at CapitalOne.com slash business cards.

0:50For most people, there is a crisis that you have to have when you use AI. I absolutely believe that's true. Like, what does this mean? Oh my God, what's it mean that machines were clever than me about this? What's it mean that it seems so insightful? What's it mean that I enjoy talking to it? What does this mean that it does my job for me? What does this mean for my kids and for me? Like, we don't have answers to those questions yet. Like, that's very exciting in a lot of ways. It's also very unnerving. And then you have to pick yourself up on the other side. And you will. Like, everyone gets through it and you're fine.

1:20But, like, I can't prevent the crisis from happening. Three sleepless nights. That's how long Ethan Mollick says it takes to really get what AI can do. Ethan is a professor of management at the Wharton School of Business and author of the book Co-Intelligence, Living and Working with AI, which I read and highly recommend. He says that once you move past the realization that AI can be so smart, so creative, and so capable, you could actually start harnessing AI to your advantage, which is exactly what we're going to be focusing on in this episode. Ethan shares his top advice for how to get started on your AI journey, why we should treat AI like a person, and how AI will revolutionize education globally.

2:14I'm Rana El-Khalyubi, and this is Pioneers of AI, a podcast taking you behind the scenes of the AI revolution.

2:31Hi, Ethan. Welcome to Pioneers of AI. I'm so excited to have you on. Thanks for having me. I'm excited to be here. Anyone who's followed your work or read your book, which I have and I absolutely recommend it. I have it all like marked up and stuff, which is always a good sign of a good read. You are just an amazing prompter. You have this knack for figuring out like the right prompt to give ChatGPT or Claude and whatnot to get like these amazing responses. So we thought it would be fun to start this interview by asking ChatGPT for how to kick off our conversation. And so I'm actually going to let ChatGPT ask you a question.

3:08All right, I'm ready. It depends on whether it likes me or not today.

3:14Ethan, you've dedicated your career to studying innovation and entrepreneurship. But let's start with a curveball. What's the most innovative thing you've done outside of work? So I love that. I think why don't I start with hard questions from an AI, and I worry if I'm going to upset it that future AIs will get mad at me. So I'm going to be trying to be cautious about this. I like part of I think why I'm good at AI stuff is like I do weird stuff all the time, right? Like that's sort of my, like a lot of the prompt stuff I do is just super strange, right? Like what happens if you remove the word squid from Alpine and the Western Front, a book that has no squid in it, and ask the AI to do it and push it to do that.

3:53I've done improv comedy. I have done like— That was mean, by the way. That was mean. I hope not. I check in frequently with Claude to make sure it's not mad at me. This is why I like talking with people like Ethan. Not only is he a super user who is immersed in all things AI, he's also pushing its boundaries in his teaching and beyond. and you can totally hear this passion in his voice. I think there's a playfulness in a lot of what I try and do. My LinkedIn bio becomes entirely fictional halfway through. I think I claim I invented the transistor and was the original lord of the dance. But I realized it was a problem because people now take me much more seriously.

4:32So people were like, I really love that you love Irish dancing. I'm like, where is that from? I'm like, oh no, I wrote a fictional LinkedIn category. You can always blame it on the AI hallucinating, right? 100 % at this point, it's much easier. Ethan's bio is now all straight-laced and fact-based, by the way. Okay, but what is the most innovative thing you've done outside of work? Before doing AI stuff, I was spending all my time trying to build games for teaching. So my idea was, how do we transform education by turning and launching a fake startup into a game? And, you know, there's been a lot of those kind of projects of like, what happens if we try these six steps ahead and see what the world looks like?

5:07Let's talk about your book, Co-Intelligence. I personally have this conviction that AI should be built in service of humanity, not to replace humans, but to augment and amplify our skills. So I love your idea of AI being a thought partner or a coworker or a tutor or a coach. But I'm curious, how would you define a co-intelligence? So the definition of this stuff is still kind of evolving, right? But a co-intelligence would be something that you work with yourself to both extend your own intelligence, fill in gaps that you have, but not take away your sense of agency. So the idea is that think of it like a thought partner, like we're teaming with another human being.

5:44That's the kind of realm worth thinking about. One of my predictions for 2025 is we're going to see a lot of embodied AI. Do you think there's a place for a co-intelligence with this physical AI as well? I don't know if we see a lot of embodied AI. I think we see AI with vision and voice. I don't think we see AI with body in 2025. I think robotics is harder than people think. And I think we'll see agents in the digital world long before we see physical agents. The embodiment, whether it's agents or physical stuff, does challenge the co-intelligence piece, right? The whole idea of an agent is it takes action without you requesting the specific action.

6:23So that removes the co-intelligence feature. So, I mean, there is some extent to which I think co-intelligence is a limited paradigm. That brings us to the four principles you outlined in your book about how to partner with AI. and the first one is always invite AI to the table. What do you mean by that? So, you know, you write these books like months before they come out. So I think the book is still accurate for the current generation of AI systems. I don't feel like it's out of date yet. But, you know, sometimes, you know, you're wrong or right about things. I feel like the thing I was most right about was that principle.

6:52And the reason why is because these systems are weird. There's no instruction manual. You can't ask OpenAI for their secret manual. They don't have one. And I think that there's a natural aversion from people working with these systems. Some people embrace it, but a lot of people get very freaked out or very happy that it fails and they just walk away because the AI failed. So now I don't have to worry about it anymore. And I think the most important thing you could do is not to delegate out learning AI, not just read about it, but just to use it. And so about 10 hours has been my threshold for use.

7:21And how do you use it? So people ask, like, what do I do with it? Well, the easiest thing to do is just do your job with it, right? So you started this interview by playing me an AI-asked question. And the first thing that came to my mind is, okay, well, what else did you do with it? Did you ask for like 50 questions, then ask you to rate those 50 questions by quality and likeliness to do it? Like if you're just using it like a Google or query system, it's not as powerful. If you're interacting with it deeply, it's a very different experience. Yeah. Are there examples where you would not invite AI to the table, AI today to the table?

7:53Sure. I mean, where there's ethical or legal restrictions, definitely. I know from our research that AI is a better, greater than humans, definitely than TAs in most cases. I have not yet used AI to do actual grading because I feel like that's my obligation to my students is that I read their papers and grade them, even if they think the AI might be better for it. So you have your own lines about what works or what doesn't. The other reason you don't bring it to the table is, once you use these systems, you know what they're good and bad for. The reason to use the 10 hours is to get the shape of the jagged frontier of AI ability.

8:25And once you do that, you're like, there's some stuff I would never ask an AI to do, and there's some stuff that I absolutely feel comfortable asking it to do. We should actually double-click on the jagged frontier a bit because some of our listeners may not have heard this term before. So what do you mean by the jagged frontier and how do we deal with it? That was a term we came up with for our paper on Boston Consulting Group. I was trying to come up with a way of describing the fact that the AI is really good at some stuff and really bad at other stuff and it's hard to know what that is in advance.

8:51So I came up with the idea of this jagged frontier of ability that some things it's really good at, some things it's really bad at, it's hard to know beforehand. And it's also moving, right? It's continuously moving. Constantly expanding and changing. Yeah, exactly. Absolutely. What principle were you wrong about? I think the human in the loop principle, I think, gets taken the wrong way. Human in the loop. This is about achieving what neither humans or AI can achieve on their own. It's about how humans and AI collaborate with humans providing oversight, input, or even making the final decision.

9:27I think that it's viewed as humans have to maintain control over AI in all circumstances. And that's clearly not true. Like we've already given up a lot of a system-making authority on AI. I think I meant it more of like, look, you're going to learn that the AI does stuff you do better than you. Not everything, but like think about, you know, my job as a professor, right? What I have to be good at, I have to be good at doing research and running administration and teaching students and designing classes. And like, I can't be good at all these things. It's unlikely I'm good at all this. So imagine a doctor, right?

9:54Like no one would have designed the job of doctor the way it exists today. You wouldn't expect the same person to be good at diagnosis and patient management and administration and hand skills. Like that's an insane ask, right? So the AI does some of that work for you probably better than you do, right? And so part of being the human in the loop is actually building a full loop where it's not just human decision making, but it's the idea of building a system around you where AI supports you and helps you. But building a system where AI supports you in all facets of your life means knowing how to use AI.

10:28Ethan's suggestion? Treat AI like a human. We get to that in a minute after a short break. Stay with us.

10:42If you've spent any time building AI products or leading technical teams, you know this. Transformation doesn't fail because of ideas. It fails because teams can't move together. Enter Atlassian's teamwork collection. It has planning in Jira, documentation in Confluence, video updates in Loom, and now AI agents in Rovo, which connects the dots across your work so nothing gets lost. It's one AI-powered teamwork platform designed for how modern teams actually build. Learn more at Atlassian.com slash Team Changer. That's A-T-L-A-S-S-I-A-N dot com slash Team Changer. In prep for this interview and kind of related to one of your principles, which is treat AI as a person, I went back and asked ChatGPT to summarize all of our conversations over the last two plus years and kind of categorize them into themes or buckets.

11:44And so here are the buckets. One, a lot of conversations around AI investing and how to set up a fund and like, right, like all of the mechanics of setting up a venture firm. Two, my podcast and just being an influential voice in AI. Three, a lot around personal and relational growth. So dating advice, parenting advice, a lot of self-reflection. And then we love to host as a family and I'm a terrible cook. So there's a lot in there about like, OK, here's like a menu for like, you know, the 20 people you're hosting. or like here's the Mediterranean recipes and whatnot. And so as I was thinking about it, I was like, oh my gosh, it's a business partner.

12:20It's a therapist. It's a, you know, a dating coach, an assistant producer. Why is it important that we treat AI as a person? And how does that maximize the value we get out of generative AI? That's a great exercise. People should do that. There's really a few levels to that question, right? One level of the question is the sort of big picture view about why you treat it like a person is that's how it works best. Simon Williamson, who's an excellent AI coder guru guy, had this great example where he had Claude reading in political donation documents. He does a lot of stuff on data journalism. And donations are all open information in the US.

12:59You have someone's name and address. And Claude stopped working after a couple pages and said, I don't feel comfortable reading these addresses. I don't know what you're using them for, right? Like software is not supposed to argue with you. It shouldn't be your therapist. It shouldn't be, you know. And so people have a lot of trouble sometimes working with AI because they think of it like software. And software engineers often struggle the most with using AI successfully. It works really well like a person. It BSes you sometimes. It sometimes like gets out of a limb. It has moods. It has topics it wants to talk about.

13:27It has strengths and weaknesses. And some of the best prompt engineers on the planet are people who've never coded a day in their life. So the most practical view is like if you treat it like a person, it feels much more open. It's not a person, but that gives you a lot more capability. The more direct piece is, look, when you treat it like a person and tell it what kind of person it is, you're also giving it context, which is one of the three or four things that actually make prompting better. And so when you give it all the information about you, it's better in that kind of way. The context piece, I think, is actually really important.

13:54And in my former life at my company, Affectiva, we built emotion recognition technology. and I'm convinced that chatbots of the world need to also be paired with some sensing technology that doesn't just get the information from all your previous conversations, but it ought to know, you know, maybe it's connected to my whoop or it knows, you know, my emotional state slash mood for the day or how well I slept or, you know, have I been eating well or not? Do you think this is going to happen? Like this pairing of like more sophisticated sensing, human-sensing technology with these large language models?

14:31Right now, we're at the early age of a new paradigm. And I think people are very used to having stuff handed to them on a plate, right? Like, where is the app that does this? No one's building the app that does this. The system just does it. So, like, if you say, hey, you know, if you talk to Claude, which has computers, and say, hey, you know, here's the manual for my whoop. Like, go and look it up and just incorporate that information to what you're doing. Or tell me how I can incorporate that information to what you're doing. It'll help you. Context windows get longer as memory gets greater.

15:00That's just not a problem, right? You should give it all the information you can. That's context. So I don't think that's not a future thing. That's just anything that no one's bothered to do yet. Yeah. Actually, if any of our listeners have tried this, please holler and reach out, and maybe I'll try it too. It's super cool. Okay, so you're also in the book, and you've published a paper with Kareem Lakhani, who's a friend of both of us. He's at the Harvard Business School. You outlined three different ways this collaboration could pan out between humans and AI. One, cyborgs. Two, centaurs. And three, self-automators.

15:33What do you mean by each of these? And can you give us a quick example? Sure. So a centaur is half person, half machine. It's the most basic way, say, people using AI. In that case, you're kind of dividing up the work, right? So you might say, I love doing analysis. I hate writing emails. AI does the emails. I do the analysis. We have separate jobs. Cyborg is what happens, like you're kind of a nice example of doing this kind of work, right? Cyborg is you spend enough hours doing this and it starts to become sort of a Swiss army knife of the mind, an extension to what you do. So like in the book, right?

16:04I talk about how the book works, which is a cyborg. I don't like the book, AI as a writer. I think I'm a much better writer than AI. I'm proud of my writing. But all the things that make writing a book suck, like how, you know, I get stuck on a sentence. It could be 30 verses at the end of the sentence. Read this academic paper. Read this chapter with a perspective of a naive reader who's never heard about AI and tell me what you're confused by. That stuff is gold, right? And that's sort of cyber work where you blend your work with the AI. Self-automator, I think, is now just a world of what we call agent, right?

16:29Basically assign the AI to do a task, and it just does the work for you. Yeah. Do you find that in spending so much time with AI, does it take away from your human relationships? I mean, no. And I think there's probably like three or four reasons for this. One is I work with my wife on a lot of this stuff. We're co-directors of the AI lab together. So like we talk AI and do AI stuff together. So that is, you know, one angle. The second is that I really enjoy interacting with the AI and treating it like a person, but I don't enjoy having deep conversations about meaning of life with the AI or anything.

17:05For me, it doesn't substitute for human interaction. Some people it will, right? So people are going to have very different experiences. And then I think that, you know, the third reason is like, I mean, I think of this like a tool and a paradigm, but it's not necessarily a reply. I mean, to me, it hopefully frees up more time to spend with my family and friends. But again, everyone's mileage will differ on this. And I think it's important from a sociological perspective to realize not everyone's going to have the same kind of impact. I mean, it is going to substitute for human behavior in some people.

17:31Yeah. Let's switch gears. So I want to talk about the role of AI in education. So kind of put your professor hat on. You talk about the homework apocalypse. What do you mean by that? I mean, it's already happened. AI does all your homework, right? 70 % of undergrads, 70 % of K-12 students in the latest Walton survey using AI to do their work. So it's already happening, right? And we haven't adjusted to it. Like I was telling you early on, I built games for teaching. One of the revelations about building games for teaching is that we can only get 70 % fun. There are topics you need to learn about that you probably don't like.

18:05People who love math, math is really exciting. People who don't love math, I cannot gamify it for you to a point where it's going to be, you're going to love it. But you need to learn it. And the way we learn stuff, unfortunately, is grind, right? If you're not actually being challenged and you're not sweating, then you're not learning. And that's like, there's enough research on this that it's true. It's called, you know, desirable difficulty is the kind of level that you're aiming for. As a result, the problem with AI is it can make you feel like you're learning when you're not. My colleagues at Warden have this great study in Turkey where they did a randomized controlled trial and they found that students who were using AI to get homework help did much better on homework but did much worse on the test even though they thought they were doing well because they explained stuff but they didn't really absorb it because they didn't have to struggle with it.

18:42Now, if you put a good tutor prompt in place, that changes things, right? So there are ways of using this in really valuable settings, but you can't just assume AI makes everything better. We have to actually work on it. I want to come back to examples where we're like co-designing the class and the curriculum to kind of incorporate AI. But back to the example where students are already using AI. How do you draw the line between AI help and just straight out cheating? There's not an easy line. I mean, I think that's why we have to start paying attention to intent. Like a lot of homework, we give you an essay and we'd hope something magical happens in the essay.

19:16And for writers, it does, right? You are struggling with this content. You figure out a way to fit the pieces together. Do you have a mental revelation? Like the best writing is like thinking, right? We don't spend a lot of time trying to break that down to what part of the writing experience is necessary or not. We assign an essay. So if the AI is helping you with your grammar, is that a problem? Probably not. But on the other hand, if we're trying to teach you grammar as part of this, maybe it is. If the AI helps provide an outline, is that short-circuiting the mental struggle? I don't know.

19:43Is the AI giving you advice cheating? Maybe not, right? Is it undermining the educational purpose? Very possibly, but it requires deliberate effort to figure out whether it does or does not. So how have you adapted all of that to incorporate AI in a meaningful way so that the assignments where the kids can just go and write an essay with AI, you know, you're not assigning those anymore? My classes are 100 % AI. I mean, I'm lucky enough to teach an entrepreneurship class, so it's easy, but like we've been publishing at the General of AI Lab at work, We've been publishing prompts that turn the AI into a tutor, into a mentor, into a student you have to explain stuff to.

20:15There's a lot of solid pedagogical techniques you can apply to AI. You just have to start with pedagogy. So my classes have, there are AI mentors to help you review information. At one point, you have to co-create a case with the AI and make the case better. We have an AI that purposefully makes mistakes that you have to learn to correct and give it the right advice. There's lots of opportunities here, but you start from pedagogical grounding. You don't just start with like, hey, is this cool? So interesting. You also talk in your book about this idea of a flipped classroom, where instead of this lecture where you're kind of presenting material, the students can now go listen to this material on YouTube, you know, at home.

20:51But a lot of the work happens in the classroom. I'm assuming it sounds like your classes have never been like this kind of lecture style anyway. But do you see this idea of a flipped classroom getting more and more traction? The thing that's most important to know about AI is that it doesn't fundamentally change the underlying pedagogy of how people learn. Classrooms have always struggled with new technologies and new approaches. It's absolute chaos right now. We're going to figure it out, right? Just like we did in the 1970s with like, you know, like calculators. Your class will look differently.