Recommender Systems with Carey Morewedge

Hello and welcome to the Behavioral Design Podcast. This season we're diving into the intersection of behavioral science and AI. We want to make sense of the state of AI, from understanding how humans interact with intelligent systems to using AI to do behavioral design itself. I'm Aline Holsworth, a health tech advisor specializing in AI and product design. Over the past 15 years, I've been crafting human centered products with behavioral science

at the core. At Apple, I LED Behavioral Science for Health AI, designing and launching AI powered features to help users reach their health goals. And I'm Samuel Sultzer, your second Co host. I'm a behavioral strategist specializing in hybrid formation and designing products that drive long term baby change. I work with leading tech organizations integrating AI to scale behavioral design for good.

And I'm also the founder of Baby Bites, a dedicated community on behavioral science and AI. Quick word on Nuance Behavior where we help organizations build impactful digital products using behavioral design. We only take on a few clients at a time to ensure the highest level of quality for our tailored evidence based solutions. If you'd like to become one of our special projects, e-mail us at hello@nuancebehavior.com or we could call directly on our

website, nuancebehavior.com. Hey Aline, I'm curious if you've played this game that I call the do but not recommend game if you live with us? It sounds a little bit like you just made this game up. I I somewhat did. It reflects a little bit of my own behavior, navigating a world filled with recommendation systems. And what I mean here is like I do something, but the game is for me to ensure that the platform I'm using is not recommending me things based on what I do and so.

Because they're things that you don't want to do. They're things that you wish you didn't do. Yeah, like so on Twitter for example, I would say a lot of my time is spent kind of reading about facts or things relating to my interest in football. And so I for football club, I get some news around that and and so on. But I've made sure that in the feed I've made it so that I never recommended any football facts or any any football news and so on. So I had to act yeah, yeah.

You can use like block word, certain words, you can downvote certain things, you can like depends on the platform. You can kind of like say, don't recommend me this, don't recommend me this. So I usually have to like I have to go into the various accounts that I kind of still follow to

get the news. So I'm like playing this game of like getting into to this because I see myself as someone who's engaging with Twitter in order to stay up to date with what's happening in AI and behavioral science and so on. And not to stay on top of what's happening in football, even though that's what I often times do. So I'm I'm playing this game and I'm curious if you can relate to it or? Yeah, in some ways it sounds like a pretty nice commitment device, right?

Like you're adding all this extra friction to getting the information about your soccer. I'm just going to say soccer. And, and maybe you don't consume as much soccer as you would otherwise, which, you know, maybe it's part of the goal, maybe it's not. But honestly, when it comes to Twitter or X itself, I pretty much stopped using it when it became X because I, the recommendations that I got were, were so far off from what I was interested in.

I felt like I had for many years really sort of, I don't know, led the algorithm towards things that I was really interested in. It was purely academic. It was it just like my colleagues sharing their new findings and, you know, adding the TLDRS about it and just just really, really useful summaries and commentary about what would have been published.

And that at some point all of those disappeared and I just got political commentary and, and things that I'm just so, so, so uninterested in or lately I don't, I don't want the, the X version of that, let's say. But, but going back to your original question, do I, do I do any of that manipulation? I actually do a much lamer version of that, which is I just exhibit self-control in the moment, so when I see. Something. I do, yeah.

I actively stop myself. I say to myself, I see something that I that I don't want to be. I don't want that view to be incorporated into into my algorithm. I look at it and I say don't do it, that's going to work against you. And then I actually don't look at it as much as I want to. I I tell myself, Nope, you're going to be seeing those stupid videos for the rest of your

life. Yeah, but then you're like, aware of the consequences of like, actually watching this video will have some ramifications. But I don't miss it. I, I think that the difference is that those, the videos that I'm tempted to look at are not something that I, I really care about keeping up on. Whereas for you with your, with your silly sports, you do, you do care about keeping track of that thing.

For me, it's like parenting videos, you know, like people, people telling you it's all right, you're going to get through this and here's why parenting is so hard. Like you don't really need to see some random stranger with a million followers telling me that that's not how I want to spend my time. So yeah, that that has been my

solution. That's fair. And I guess that's the last point on this from from my, I guess I would love to say that my motivation was that it's like some form of commitment advice in a positive way. But I think it's a little bit of like some form of Dorian Gray thing in terms of, you know, how in Dorian Gray is this person who lives a life where he's beautiful, but he has like a image of himself that becomes ugly and ugly and uglier in

hiding. And I feel like what I don't want to see is when I open my feed is to be reminded of the ugliness, like in terms of my behavior. I don't want to see that I am someone interested in. Turn the mirror. Away some other things and so I think kind of ensuring some form of yeah, nice image of myself is probably why I do it. You're you're sweeping ugly Sam under the rug, hiding him so that you can believe you.

You can even fool yourself into believing that that this is that ideal Sam is Sam. Wow, interesting. All right, cool. I love talking about recommender systems and I, I also love how the algorithm has just become part of like general conversation. Like everyone understands that they're being fed recommendations based on what they've seen or what they've done. And, and that, that this is just

a part of everyday language. Like what, what you're doing to modify, to control or not control your algorithm. It's this is not even niche language anymore. No, we're way past that point where we said like, oh, isn't it creepy that what I do online impacts what I see and stuff like that. It's just that's just so 2010 or 15 or something. Yeah, pretty much.

Well, and it's funny because even though this, this is has become so familiar to all of us, what I didn't realize until a couple years ago was just how complicated these systems are. And Oh my gosh, until I, I was really like in the weeds of working with recommender systems. I just, I just didn't have an appreciation for the math and the engineering and the the complexity behind them. You had some, you had some first time experience. In, in I did. Yeah, yeah, I worked.

I worked quite a bit with Recommender Systems. I won't say where, but but yeah, I thought it might be useful to do a very high level overview of, you know, if you like take a recommender system course that that you always get the, the sort of these are the types of recommender systems. I thought I'd do that for us today for our behavioral scientists audience who you know, probably hasn't come into

into contact with this. But I think it'll be also useful for our conversation with Carrie where we really get into the types of data and how preferences are modelled.

So when we, when we think about our recommender systems, one O 1, there are two major categories and how they work. We have content based filtering, which where each user basically a profile is created for them based on what they say they're interested in. So for you, Sam, this would just be your your academic articles on Twitter. And then that information is matched to content in order to provide recommendations to to users. So there's type 1 and then type 2 is the collaborative

filtering. And this is I think often what we think of as recommender systems. And that's because it is more commonly used. And this is where you, you're really relying on people's implicit data. So what there's what they're viewing, what they're clicking on. And, and this would be if you didn't override the system, this would be your, your soccer content. So they're basically inferred preferences.

We, you know the we, we the algorithm, think that you prefer watching videos of soccer because that's what we see you secretly doing at night with your flashlight under the covers. Yeah. And I guess this is what I would say of all of the different algorithms and recognition systems, I think TikTok is probably the one that has become most kind of famous or infamous for this type of kind of not only relying on kind of what we click on, but use like the time spent looking at something.

I do feel like the the vernacular of train the algorithm actually really did come out of the early TikTok days where people realized it was just so blatant and obvious. If you watched one video that you know, the you would be served up another video just like that one right after. So it was, it's just so blatantly obvious. I think that's sort of is when it entered our vocabulary awesome content based and collaborative filtering, two

major categories. Of course most recommender systems that we have today are some combination of those and we call those hybrid models. Yeah. And so those are major types. One thing that it that I think is maybe interesting about the types as well is just that there is so much more implicit data available. And so just as a result of that surplus, you know, you, you kind of can only use the data that you have.

In some ways that ends up being what companies rely on in order to make the recommendations, but it also tends to be more predictive of time spent on the platform. So if that's what you're you're trying to optimize, which in many cases unfortunately is you, then you might be sending more gratifying content that is that's going to keep people, keep people clicking and viewing and scrolling and so on.

Yeah, I know this is such an interesting subject and I'm so happy we had a chance to to speak with with Carrie about this. So maybe, yeah, do you want to introduce Carrie? I would love to.

So we have a very special guest today, Carrie Morwich. And for me personally, Carrie is really the person who took what was purely an academic love of behavioral science and showed me that that this could actually be a career, not just a passion. And so Carrie was was really my very first real mentor, the first person who kind of picked me under his wing and showed me the way. I met him when I was an intern actually at Carnegie Mellon way back in.

And I'm not sure I should date myself, but back in 2008. And I was so excited by the work that that we did together during this internship that I actually asked him to be an advisor on my thesis. So he agreed. We, we get to work together and just the, the amount that I learned from him in that period is, is unparalleled.

And I, I remember this moment, I think we were eating lunch somewhere and we were, we were talking about my thesis work and just just totally out of thin air, he pulled out this, a full citation for some statistical analysis that we were talking about. And I was, I was just blown away. My, I'm just completely awestruck. Like, like this guy is pure genius. Like how, how is he doing that? And yeah, Carrie really taught me what it what I think it takes to become an expert at something.

So if we Fast forward to today, Carrie is a professor of marketing at Boston University. Here he he continues to strike on his students and to research decision making. Yeah, well, I I first of all, I love this idea of having in your back pocket a cool citation to to impress people with but but also I'm interested. So tell us more about what does it mean here in terms of research decision making in the context of Carrie? Well, first I'll say I don't

think it was intentional. I think that is that just comes naturally to some people. I certainly don't have my own version of that, but yeah, so does. Yeah. What does it mean to to study decision making? It's actually kind of hard to

describe Kerry's research since it spans so many different categories. If you sort of take a look at his early work on affective forecasting with Dan Gilbert and Tim Wilson, some very amusing research on consumption behavior. And and then, you know, he's really, he's also covered all of what we consider the the behavioral economics classic, right? Like so mental accounting, anchoring, confirmation bias.

He's looked at the endowment effect, actually teasing it apart from loss aversion, because some people kind of treat these as as two sides of the same coin. But Kerry will tell you they are. They're not. And today, very excitingly, he's one of the most prominent researchers looking at how humans interact with AI systems. Which, of course, is what this episode is all about. Yes, I'm giving you a quick

preview of this episode. We look at AI systems from perspective of Kerr's work and especially kind of understanding this idea around the complexity of human behavior. And that if we don't understand that, this is to fall into these three common pitfalls that can lead to misalignment within AI recommendation systems and honestly so much more. In this episode, there's so much that we cover, including strategy and carry to our quick fire round of the season.

To AI or not to AI and asking for his most controversial opinion about AI. Oh wow. Carrie, welcome. It's wonderful to have you on the show. Thank you for having me. Such a pleasure. It's been like a lifetime since we've hung out. It does feel that way. Our our beers in in Pittsburgh have gotten warm by now. At least our children's life. That's. Right. And for me, I think it's been a lifetime in terms of we haven't had a chance to meet before.

So very excited to to shout. Me too, thanks for having me. So, Carrie, this season we're really focusing on AI and what behavioral science can say about AI. And you of course, have done some very exciting work at how people think about AI systems, how we perceive them, when we trust them and when we don't trust them, but also how these systems that are supposed to be these ultimate prediction machines, right? They can actually get it very

wrong sometimes. And, and perhaps that could be because they're solving for the wrong goals or, or they don't have the, this, this understanding of human behavior built into them, or, or at least the, the machine learning

engineers who are devising them have not thought about how humans think and behave. And, and so one of the points that comes through in your research is that when, when these AI systems base their recommendations, So in a recommender system, for example, when they base those recommendations on users digital footprints. So what was viewed or clicked or bought all all of those behaviors that are tangible and and seen captured by the system.

When the systems are, are designed in this way, they they ignore what's arguably the most human thing about humans, right? And that's the gap between our intentions and our actions. So the the behaviors that we exhibit may not necessarily map to the things that we actually want to do, the ideal humans that we see ourselves as. So what I'm hoping you can dive into now is, is like what are algorithms doing wrong when they're basing their recommendations on users digital footprints? Right.

So that's it. It's an amazing lead in and maybe it's useful to begin by thinking about like what they're trying to solve and how they do that. Let's begin by thinking about it as a spreadsheet, right? So an algorithm, you're telling it here are some columns to consider in the spreadsheet, right? I want you to use the information in those columns to

make some kind of prediction. And like if you're thinking about it as regression, like those are your like X variables and the prediction is AY or outcome, right, that you're trying to predict. And then the data is really like the rows, right? So the users are the rows right in that spreadsheet. And so as algorithm designers, right, you're trying to think about, OK, well, what columns should I include or exclude in these kinds of predictions?

Do I want to just like take all the data I have and like try to model something and see what it is? Do I think that there's like a certain number of columns that does something and then I'm trying to predict that? And then you're also trying to think about like, what is the thing that you're trying to predict and what's the best representation of that, right. So if I want to think about, for example, let's say that you are Netflix, right?

And you want to show Samuel movies that he will enjoy, right? Like the outcome you're trying to predict is Samuel's enjoyment of the movies you want to recommend movies that he'll enjoy. So he stays on the platform, spends more time there, keeps his subscription. What kinds of features of films do you need to include in your prediction to like get a good prediction about movies that Samuel will like?

Is it about like the actors, the genre, the length, right, like the visual representation, the language? So there are lots of different kinds of features, the director, right? Like there's lots of different kinds of features that you can include in that kind of prediction, right? And then the. Then you're trying to think about like, OK, well, what does it mean for Samuel to like the movie? So is it that he finishes the movie? Is that he finishes and rates the movie positively?

Like, you get that like or love, right? Yeah. Is that like, or should we just use like or love? Like even if he watches the movie and he doesn't like or love it, like then should we not say that he liked the movie? I mean, I've had many experiences on Netflix where I ended up watching a whole thing and I was like, that was terrible, that. Was a crappy movie. Yeah, exactly.

So there's a lot of different kinds of assumptions that we're making when we tried to take some abstract concept like enjoyment of a movie. How much will Samuel enjoy the movie, right. The idea behind that is that we're using, we're observing his behavior. So how much of the film, which movies he chooses, which movies he watches, how much of the movie he watches, how he rates it, right? We're observing his behavior to infer how he feels about the movie.

And Economist would call these two kinds of states that we're using revealed preferences. So what he's doing, what he reveals about his behavior to infer people's normative preferences, what they actually believe. So let's take an example from Amazon. So many of us use Amazon for like as current for most of our purchases, right? Yeah, so. It's not returned. Yeah. So let's say that most people don't rate the things they buy on Amazon, but they might. Let's say that you buy diapers

on Amazon for your child there. We could say the purchase is 1 signal of liking the diaper and then repeat purchase would be a stronger signal of liking that diaper. It's harder to do something with a lamp which you buy and either return or you don't. The return is a clear signal.

I don't like this lamp. Buying the lamp and not returning it probably means that you like it, but we don't know how much you like it. Do you buy, I think, other things from that brand, or do you go back and look at other lamps after you buy it? Right. So it's a really difficult problem because even with big data, we don't have a direct conduit into people's mental states. There's a lot of missing data. Yeah, there's a lot of missing data there.

And there's a lot of assumptions that we're making about like what people are doing and what it means. And so the what people are doing are the revealed preferences and the what it means. Like, do we actually enjoy watching, you know, Love Island, right? Is that something that I would enjoy watching? Yes. Is that something that I would be happy about watching tomorrow? Those are different questions, right?

Oh, and do you see yourself as the kind of person who watches Love Island and then the, you know, sort of the, the enjoyment while you're watching it versus the regret you may feel later after, you know, however many seasons of binging it? Yeah. Are are we still talking about me as a as an example? So here's an example. And it's not always that people are trying to pursue pleasure, right? So reading the New York Times is like a factory of sadness, right?

It makes me so sad to read the New York Times, but I feel obligated to do it as a citizen, right? But it's something that like, I feel afterward, I feel good that I know what's happening in the world. Although if you like looked at how much, like if you were to measure my affective state while I was reading New York Times, I feel very unhappy most of the time. You know, So like, there are different kinds of goals that

people have. There's a really amazing data set from the Neely Center at USC and they look at like people's emotional states while they're engaging in different kinds of social media. And like YouTube is one of the most positive kinds of experiences. And you could think about it like it's not necessarily a platform's design, but like why people go to YouTube is to learn, right? Like you go to learn about strollers or cars or like how to fix your refrigerator, right?

Like that's like you're, you're learning things and it's kind of interesting, right? Like Twitter and like next door are like the least positive platforms, but like there you're kind of going there for like things that will make you unhappy, Like, and so there's a lot of kinds of assumptions that we're making when we're looking at people's behavior about like what the purpose of what they're doing is like what are they

trying to optimize on? And so is the platform optimizing on what the user wants to optimize on in that situation. So that's like a long winded version of like what the difference between like normative preferences, like what I'm trying to optimize on as a user and then reveal preferences the platform trying to infer or what am I trying to do? And how well does my behavior predict what they think I'm trying to do?

And yeah, and, and there are different ways of getting at these different versions of people's preferences, right? Do you think there there's a way to have your cake and eat it too? And, and thinking about, you know, how if you just take the Netflix example, you have multiple shelves, right? And you can have your like, you know, classics in one, which maybe is the version of you that you want to be. And then you have your Love Island shelf of, you know, all the trashy reality shows.

Do you think that these multiple shelves can coexist or that, you know, will they just be overtaken by the algorithm learning over time when you continue watching Love Island that the classics kind of get pushed lower and lower into your recommendations? Yeah. I mean, I think that's a huge problem because it's hard for the platforms to observe like the heterogeneity of our

preferences. And some platforms do do that, like Netflix does try to cluster different kinds of preferences that we have and they're building up like. But a lot of that is called from like collaborative filtering, which is it's showing us things that people who like the particular movie also. Yeah, Someone Like You who watched similar things would all. Here's something else that they watched and enjoyed.

I, I think the a big problem is that our preferences are much like a, a Senate right or a, a, a group of people like a, a collective right. We have different preferences and sometimes they conflict. And which preference is dominant in a particular situation often depends on the goal we're trying to pursue or where we are in some particular context, right? You could think about at a family level, we need different Netflix accounts to make reasonable predictions.

I have a friend who shares an account with his children and his wife. And it's like there's no coherence to what it's because like it could include like Coco Melon and like, you know. Spotify has that problem as well. Yeah, some like, you know, Scorsese film and like you wouldn't suggest like, you know, Scorsese film to a child. So, but like in our in, in my family, we have 3 accounts, right? One for my wife, one for me and one for our children. There's more, more coherence there.

They, they, there's still some heterogeneity too. Yeah, but. Even within individual preferences, there's a lot of variety, right? Yeah. And if I recall, there was a point maybe early on in the Netflix days where Netflix asked people what kind of films would you like to see? And so it really directly got at their stated preferences. But that went so poorly because people didn't actually want to watch the films that they said they wanted to watch.

And so they, you know, they lost users and so on. So 1 is this issue right? So Katie Milkman and Todd Rogers and Max Bazerman have a lovely paper called, I think Highbrow Films Gather Dust. And the basic idea there is they're looking at what movies people put into their watch lists. And the canonical example is like, Schindler's List, right? Most people feel like obligated to watch Schindler's List or think they should want to watch

Schindler's List, right? Or movies like that that are like historical documentaries that are kind of sad, right? And then what do people actually watch, right? I think like the top movie in Netflix was like Red, which is like an action movie, right? So, you know, our extraction is also very popular in Netflix. When we're thinking about our choices, like what do we think are the movies that we should be watching? We're putting those things in in

our watch list. And they're probably like classic movies or movies that are emotionally difficult to watch, things that require attention, things that will make us feel like better people and that we would be be proud of the next day having watched or spent that time that way, right? But when we look at when people log on and like you're exhausted from work and you're trying to pick something, you don't want

to do that, right? Like. Exactly. That's that's like a eating your like you've already eaten your vegetables for the day, like you're, you're ready for dessert, right? So, you know, so in, in those kinds of cases, then people are watching more lowbrow films, right? So they're watching like comedies and action films and whatever is going to be more like emotionally palatable to them. And so this, there's this really big problem about platforms. This isn't just a Netflix

problem. Platforms more generally are observing people's choices while they're making the choice. And that sounds very reasonable, right? Like I want to see what people choose when they choose and I'm going to show them more things that they choose. The the problem there is that like a lot of platforms then end up recording the weakest point in our self-control conflicts.

Instead of reminding me that I should be watching these other, I want to watch these other things or I should feel like I should be watching these other things. And it's not measuring like how I felt about the movie I watched last night. It's like, how did you enjoy this movie right now, right? And like, do you go back and watch that movie again, right? Or do you watch other movies like that? So, you know, if you keep feeding people, you know dessert, right?

And you're like, people eat dessert. Like if I put dessert in front of my kids and they eat dessert, right? And then you're like, oh, my kids like dessert, I'll give them more dessert, right? So like you keep doing that, like they'll keep eating it. But like, that's not to mean that like they wouldn't eat an apple if you put it in front of them, right? Or like they don't just want to eat dessert. They might like eat some apples and some strawberries or

whatever, right? So the platforms even in like say, like, let's go to like an example, like Twitter, right? Or you know, and there or like TikTok or something like that, right? So they're looking at what you're consuming then. And like in Twitter, like you might be interested in like a salacious partisan news story, like JD Vance being like attracted to furniture, right? Like that might be funny, but that's not like, and you might like, look at it while you're scrolling, right.

But like, that's not like what you. If the platform is all that, then people are done. Oh, yeah. And so the yeah, like, so like the the real problem with these kinds of platforms. And I think you can see it illustrated with like, sort of like the problem that Twitter's algorithm has had since it was like, acquired and became acts, right? Is that like, if you get too much into the want space, your platform is basically becoming like, like a flight from Pornhub, Right? Yeah.

And like, no, seriously. And like, it's like people, people will eventually exit the platform if they don't feel good about like the next day about having how they spent their time. They might spend a lot of time on your platform when they're there because it's like shocking and salacious or like it's like, it's like emotionally evocative. And maybe they, maybe it's like moral outrage that they're

feeling right. Or maybe it's like pleasure, or like some kind of hedonistic experience, which is fine to have, but you need a balance. Yeah.

And I think that that is really the most critical piece of the puzzle, right? Is this impact on long term outcomes, both for the person's well-being and how they, you know, the relationship they have with the the product, but also, you know, for the for the company that's trying to get

users to stay. If they're only looking at these short term outcomes and behaviors of what people watch, then they're not understanding the fact that in the long term this is actually bad for business because people are not happy with their behaviors long term. Yeah, and that's like one solution is looking at like let's just track behavior more long term, right rather than like what are people clicking on like what's what's the immediate

kind of response, right. So what how long is it till people come back to the platform?

Yeah, right. What is their like frequency on on time look like over a month? Also like can you look at when people are making choices, but like in an interval of time before they consume something like the watch list, right, right. Is there a gap between what they're putting in that watch list and what they end up consuming? And if there is, then there may be ways to try to think about how do we reconfigure the recommendation system So it's reflecting both kinds of

desires. Because if you make a platform all shoulds, people won't log into it because there's no demand. People will not want to do that. But if you make a platform all wants, then people don't feel good about the experiences that they're having later on and whether or not it's detrimental for their emotional or physical well-being. We can put that aside. People just may not feel like that represents their

preferences. A broad problem too, like whether or not we're actually modeling wants or shoulds, is that a lot of the kinds of decisions that people are making on these platforms are like habitual or they're done very intuitively. People are not like deliberating a long time before they're making these kinds of choices, right? And so my decision to click on a news story is like, you know, does the picture and like, however many characters look interesting.

Does it look interesting from that point, Right. I'm not like, who is the poster? How many likes are there? How many retweets, right? How many comments are there? People do that when they're thinking very carefully about like, is this true or not? Right, but that. Sometimes. And like, and also like that. So there's like one level of people are not like fully thinking about all of the information. They're making more of an intuitive judgment than a

deliberative judgment. And so I get picking more up on like how people think quickly than how people think carefully.

And the other sort of thing is that like, you know, platforms run a lot of AB tests, but they don't make like major structural differences in the tests that they present to users. So we don't know like how much context is there. So we're then like, what we end up with is we're inferring people's preferences from like limited attention, right, from habits, and then like within a particular decision context that

could skew their behavior. So let me give you an example of like why each of these is a problem, right? Habitual behavior. If we just observed smokers and like saw how they were smoking, we would think that like, most people who are smoking enjoy smoking and like, don't want to quit, right? If you survey most smokers, most smokers say that they do want to quit, right? So like, if we're looking at some kind of behavior that's habitual, like most people, I, I would get gather spend.

Like the data suggests that most people are spending about an hour on like a platform like Instagram. I've been in a where I saw someone from Meta show a slide where they went from like the average user in India went from like 20 minutes a day to like 40 minutes a day. And they were like excited about that, right? That's great for Meta, but they thought the user was having a

great time on Facebook too. I don't think anyone consciously wants to spend an hour a day on Instagram or an hour a day on Facebook and think like that's a good use of their time. I think if you ask people like how much time would you like to spend on these platforms, they would be like something like maybe like half an hour Max. But most people would probably say like, I want to spend like 5 to 10 minutes, right?

I definitely want to spend, I don't want to spend more time on this platform than I spend with my children, right? Or like more time on this platform than I spend, like eating with my family, right? Yeah. So I guess we might want to jump to another topic, but before then I give you a chance to summarize the, you know, when we talk now quite a lot about the various kind of aspects of recommendations engine and algorithms to support us to make some more better decisions in some ways.

What are the big problems that you've covered so far? Yeah. So there's sort of three, right. We're looking at thinking fast, not thinking slow. And that would be inattention habits and, you know, context effects on preferences. Algorithms are observing wants, not shoulds, right? What we want now, right now and not what we think we should have chosen or should choose in the

future. And then algorithms are prioritizing like popular options and the status quo, like what we have data on, what most people like, right? And those kinds of three features are not necessarily inherently bad, but like they may lead us to make inferences about what users, users normative preferences, what they think, what they actually want, right, That are inaccurate because we're looking at the revealed behavior.

So if I just look at what you're choosing intuitively, I may pick up on habits not like, you know, I see that you smoke and think that you like smoking, right? If I look at what you want, I might think that you don't like more highbrow films. You really like just lowbrow films. If I look at what you choose from like a recommendation system, right, that tends to prioritize popular options, I might think that, you know, you like Taylor Swift more than this, like other indie artists, right?

Then, you know, you mentioned looking at long term behavior. That's a big kind of thing. Another thing that we didn't talk about is there are users who are modeling the behaviors that you would like users to model on that platform. Like there are people on Twitter who are reading interesting articles that are true, right? Or carrying like comedians that are telling funny jokes that everyone likes, right?

You can teach or train the platform on that subset of users who are behaving the way that you would like users to behave. The analogy is let's take like we were way MO and we're making self driving cars and the way that you train a self driving car is it observes how a human would drive. Now, let's say that you wanted to train driving cars and you train them in Boston. Boston is full of terrible drivers, like self included, right?

Like we do all kinds of terrible things like you have to break the law to drive in Boston. So like it's just impossible otherwise. So should we train self driving cars on the population? No, we want self driving cars to be trained on people who'd have a zero accident history, right? Canadians. You know, yeah, like nice drivers. We want like more self driving cars to be better versions of ourselves, right?

And so like we kind of like an idea like why don't we train platforms on those better versions of ourselves or the

users that we hope the platform was best served right? So one maybe quick thing that I want to touch upon before we jump into our quick fire round of the season is something that you've explored quite a bit, which is somewhat related to what we covered already, which has to do with algorithm aversion. So it's interesting to see how people react to algorithms and and what levels people feel comfortable with that and not comfortable and why and so on.

Sure. I mean, I think a lot of like, you know, for the most part, algorithms are doing an amazing job in much of our lives. Like we've been spending the last however long talking about some of the problems with these algorithms. But like on the whole, algorithms are like dramatically improving our lives in many dimensions, right? So for example, you know, being able to, I can find the best route through Boston traffic at

any time of day or night, right? And I know all the routes from like my office to my child's, like school or from my office to my home. But like an algorithm can guide me through the optimal route at any time of day or night. Like that's amazing. It can shave like 10 minutes off of my commute. That's, that's fantastic, right?

Like, I think algorithms are also amazing at, like, solving like, the LA problem, which is when my cousins and my grandparents lived in Los Angeles and like, when I used to go there in college, like in the 1990s, which is, I shouldn't admit that. But like, like LA is a city where, like, the most amazing sushi place you could ever find is in, like, a strip mall on, like, in some weird place, right. Like, you need to know.

You need to know that, like, that sushi place is in this place or you would drive by it and forever never see it. Never. Yeah. And like all the best like bars and clubs or like, you know, different kinds of social activities are all like, there's like, it's very hard to infer like what is good from just like casual observation, right? You need to have local knowledge

and these algorithms. Now I can like go to LA and I just pull up Google Maps and I'm like, OK, find me find the best Mexican restaurant like in Westwood, right? Like or in North Hollywood or whatever. And it's gonna like, it would give me, like it gives me that note local knowledge immediately, right? So it's like aggregating all this information that like we didn't have before. So algorithms are amazing on all

these kinds of dimensions. So, and we use algorithms too, Like, so you know, if Sandra, like if you came to Boston and Alien came to Boston and you were, you know, at a hotel and you were like, I'd like to go see Fenway Park because I'm really curious about where the Red Sox play, right? And you go downstairs and there's a hotel concierge and that person is an expert on Boston, right? They know where Fenway is. They like have maps, they can tell you how to get there, right?

My guess is you walk right by that concierge and like check out Google Maps, right? So even if the concierge told you how to get there, you'd probably still open Google Maps, right? So we have in many kinds of cases, preferences for algorithms over people. Like the. So Google Maps example is the one that rings true all the time, right? I would rather use Google Maps than any human direction, right? Always. But there are also domains in which we don't want algorithms in lieu of people.

For example, I would have a strong aversion to my son first romantic relationship be with an algorithm or a chat bot rather than a human being, right? I would feel despondent as a parent if like yes, his first love was a chat bot, right? Like that was feel terrible, right? And so that doesn't seem like a good substitute for like that human capacity, right? And I think a lot of people more immediately worry about algorithms like doing their work, right. And in some cases, like we don't

necessarily like care. So people used to do statistics by hand. Now people use statistical packages, which are algorithms. They'll do their statistics. And I don't feel threatened as a professor that there's a computer program that does statistics. That's amazing. I don't have to do this part of the work, right? But those are the things that people necessarily don't identify with as their

competency. And So what I think is when we really feel uncomfortable about algorithms doing something, it's some kind of competency that is important to who we are, our identity, right? And so I might not care about using a statistical package because for me, it's like the theory or the idea that's important. But for a statistician, they might be like, wait a minute, I don't want ChatGPT doing all of your statistics, right? Because I have important kinds of ideas or whatever, right?

So, yeah. And. I wonder how much of that boils down to the fear of losing your job, right. You might be happy to use, you know, AI tools to accomplish your tasks as long as you're the primary one being paid and not losing your job. That's one of the biggest threats that people face, but they're also resistance to automation, which is fueled by algorithms that people have where it's more about things that they enjoy doing that they don't want to outsource.

So Stefano Puentoni has this amazing paper where like, I don't enjoy baking bread, right? So if I had to make a lot of bread, I get a bread maker. There are a ton of people during the pandemic who really enjoyed making sourdough, right? Those people would not want to buy an automated bread maker. They want to knead the dough and bake it and observe what happens, right? I love cars. I have a manual transmission car cuz I enjoy having that kind of

control right? To me an automatic is not as fun to drive cuz I have less control over it and the self driving car sounds terrible. Whereas there are many people who like driving is just a chore and they want an appliance that gets from one point to another and they don't really enjoy that kind of experience and so or having any control over it, they just don't want it to crash right? And so for them, an automated car would be delightful.

And so in these kinds of spaces, it's really sort of how much do you identify with the thing that the algorithm would be doing? And is that a compliment to like what you're doing? Does it enable you to do something more easily and like more fluidly, like the statistics package? Or is it replacing like a core competency that you identify with, whether it's like making bread or like driving or writing

screenplays or writing the news? Well, that seems like the perfect segue. Since our quick fire round is all about good or bad uses of AI. We are now going to share a a number of activities and ask you, Carrie, to tell us to AI or not to AI. Well, do it. Predict the outcome of a

presidential election. Should but not there yet cuz there's so much change in historical context and there's not enough data that I I think that's an area where AI doesn't do well now but sure in the future why not? How about predict outcome on the football game? To AI or not to AI? Totally, yes, definitely not. Early season maybe, but later in the season it's gonna get more accurate. That's right, organize a summer internship program. What kind of summer internship

program? Behavioral decision research, for example. Sure. I mean, if there is, if there are lots of good models to emulate, why not? What about the sign A better motorcycle? Sure, this is my favorite. Devise a diet plan using imagined satiety. Why not? Why not? This is when you came out with this paper, Thought for Food, where you show that people can habituate through imagining eating foods and desire those foods less often.

I used to always think in my head about how I could create my own diet of just imagining that I'm eating any time I was inclined to lose a couple pounds. Like no need. Like take your reminder, like think about eating this. Exactly. Are you happy? No, yeah, that that could work. I, I, I would be excited to see a test of that. Well the human version of me never came up with the actual diet plan, so I I guess we'll have to rely on AI to do it.

Yeah, but I mean, AI is amazing for things like putting a meal plan together, you know, devising an exercise routine. You know, I think it all depends on like the quality of the data out there for it's great. If it. Saves us from like if it saves us from the blog post Intro to the Imaginary Diet plan or like every recipe now gives you like

the history of bread. Nice. So I think you mentioned maybe it was when we started the the podcast that you have two children and so this is a relevant one for, for you read a bedtime story to a child to AI or not AI. As a parent I feel I should do that, but there are times when I can't so I would say sure. Better than nothing. Yeah. All right. Act as human participants in experiments. I don't think we're there yet.

But you think we will be. I think, you know, we do simulations like, there's lots of like game theoretic simulations. I think it can get there, and it's probably useful for like highlighting a concept, But you know, there's a lot of work suggesting that it's just recovering data that's out there so you can get basic kinds of people's response or demand to like price curves and things like that, that resemble what humans would do. But I think a lot of that's like scraping prices.

So I think we'll get there and it can be a reasonable test run of like let's say you want to run a big field study. You can do a simulation with your program, but I don't think we're there yet. OK, final one, diagnose skin cancer. Yes, it's not as good as a dermatologist yet, but I think it'll get there. Maybe AI plus dermatologist is the right combo. I think that would be better than just a dermatologist.

Yeah. And that brings us to maybe the final question, which kind of maybe indicates as well kind of your stance on AI. What is your most controversial opinion about AII? Think my biggest concern with AI. I'm not sure I have a controversial opinion about AI, but I would say my biggest concern is like it replacing human. Like large scale replacement of human interaction. Just people at home having relationships with a IS.

And not that there's anything wrong with that from a philosophical or logical perspective, but it seems like a way to escape fundamental aspects of the human experience. You're really concerned about your son's the first relationship? No, I mean in general, like it's like a is are going to put up with our personality quirks and like grumpy moods, like more open, they're gonna be able to tailor a conversation to what we would find exciting.

And like there's part of the human experience that is being validated and like feeling like your emotions and ideas and opinions are valid. And that's an important aspect. And I think that they will play to that. But I'm not sure that maybe they'll get there. But like I'm I'm worried that they will also include the kinds of challenging conversations or differences of ideas or opinions that interacting with other humans you know involves. Are you familiar with the company replica?

Yes. I think one of the interesting

subreddits out there that I randomly came into it at some point because I was interested with Replica being kind of a little bit of like before the very recent AI hype, they were already somewhat established as some form of AI friend platform. And they have a pretty big Replica subreddit, for example. And he's reading about people's kind of experiences from losing their account or talking about celebrating anniversaries and all of these things with these. I guess replicas has created

mixed feelings. So yeah, I'm just curious if you had any thoughts. Yeah, I mean, I think that that there's a use for those kinds of proxies or parasocial relationships. One could be teaching people how to have different kinds of conversations. Like let's say you want to have a conversation with your child's teacher about an issue, right? Or your family member's sick. And like you want to think about how can I convey how I feel in a way that's sympathetic and empathetic to them?

Or maybe you and your spouse are trying to negotiate what to do in the backyard, like what kind of landscaping to do, right? You could think about these kinds of platforms as mediation tools or teaching tools or simulation tools. And so for example, as department chair, I have to send a ton of different kinds of emails, right?

And if I have an e-mail that I think could be particularly important to a variety of people and needs to be particularly sensitive in the way that I'm expressing information, I'll run it through like ChatGPT or Google Gemini or a Claude just to see, you know, is there a more positive or optimal tone for this e-mail or can I improve it?

Right. So I think just as we can do that kind of simulation run, we're copy editing on the things that we're doing, having that kind of chance to simulate new or challenging or particularly important kinds of experiences and get training through that kind of simulation, I think would be amazing. What I worry about is that like we will end up substituting the simulation for reality.

And maybe that's a, you know, bloodite kind of opinion, but like, that's that to me, like feels like what I would not want to do personally, Yeah. Yeah, and I think it's a it's a the memorable thing I remember from just kind of going on the replica subreddit was just someone one of the most liked posts or upvoted was someone that had basically found a way to hack the algorithm to be a good friend, which in this community meant basically like downvoting times.

The the AI friend would say things that are like critical or not positive and uploading everything basically that when it said things that were friendly so that it would ensure that it had a friend, virtual friend in this case, that was always supportive, that was always super, super positive and so on. And it kind of becomes like a little bit of dystopian in terms of what you were describing here in terms of a bit Huxley. And it becomes very unfortunate,

but also to tie things up even more. So a need to have your kind of work when it really comes to understanding how this interchange happens between algorithms, how they're built, how they're recommending things for us, but also how we're kind of using them in turn. And understanding that kind of interchange is, is so important right now. And I'm really happy that we get a chance to explore some of it today.

I feel like we could talk much more about much of this, but thanks so much for joining and taking the time and for all of your great work on on this topic. And. Thank you for your interest in like, giving a platform to these kinds of ideas and conversations. It's great to think about these things together. And that's a wrap. You've been listening to the Behavioral design podcast brought to you by Habit Weekly and Nuanced Behavior.

Sam and Eileen tell me. This season is packed with incredible insights about behavioral design and AI, so be sure to subscribe and share the podcast with your friends, though you might want to keep it away from your enemies. In case you haven't noticed, I'm an AI voice. Yep, pretty crazy. Quite the improvement since last season's AI outro, don't you think?

If you'd like to collaborate with us at Nuance Behavior, where we use behavioral design to craft digital products with Nuance, e-mail us at hello@nuancebehavior.com or book a call directly on our website, nuancebehavior.com. A special thanks to the amazing Dave Pizarro for our show music, and to Mei Chen Yap and April English for their help in producing and publishing this episode.

Thanks again for tuning in. We'll be back soon with another exciting conversation where behavioral design and AI intersect. Oh. And do you see yourself as the kind of person who watches Love Island and enjoys watching Love Island? And then the enjoyment while you're watching it versus the regret you may feel later after, you know, however many seasons of binging it? Yeah. Are are we still talking about me as a as an example?