#342 The Secrets to High AI Adoption with Stefano Puntoni, Professor at Wharton

Generative AI is transforming industries at an unprecedented pace. But as AI changes how you or your team work, one thing is clear. Your skills also need to evolve. At datacamp, we offer everything you or your team need to adapt and thrive with AI, whether it's business users looking to get the Most out of ChatGPT and Copilot, or developers and data scientists looking to fine tune models. You can learn the entire AI skills spectrum on Datacamp. Power your AI transformation today. Start learning@datacamp.com so this technology has got to give us more than some productivity increases. And so I think we need a more inspiring conversation about not just, you know, how do we save a half an hour here, half an hour there. But the question shouldn't be how do we automate the work we're doing today is how do we automate work we're not doing at all? You know, how do we actually produce better products? We do things that are different and better than the ones that we're doing now, rather than just doing what we do now, just a little bit cheaper and faster. So what we saw in 2024, as people started experimenting more and gaining really some more objective, maybe grasp of what the technology could do, we found that beliefs about enhancement went up from 80 to 90%. That was a significant jump. While beliefs about replacement did not go up, if anything, they went down a little bit. And so now there was a gap between the two beliefs, such that more people were telling us that this technology will have the purpose of enhancing human expertise and replacing it. And to me, I saw that as a good thing and a good story.

Welcome to Data Framed. This is Richie. Last summer, A report from MIT suggested that 95% of AI initiatives ever make it from the prototype stage into production. Of course, this caused a lot of people investing in AI to panic. A few months later, MIT's work was followed by a report on AI adoption from Wharton. Today I'm speaking to Stefano Pontoni, one of the authors about Wharton Report Board, to see if he has better news. I want to find out the current state of AI in the enterprise. Stefano is co director of Wharton's Human AI Research department. He's also a professor of marketing. In addition to his work on AI adoption, he also researches the psychological aspects of AI, including the effect of worker psychology on AI transformation and the mental health effects of AI. Naturally, I'm keen to hear about both of those things as well. So let's hear Stefano's thoughts. Hi, Stefano. Welcome to the show.

Thank you for having me.

Yeah, wonderful to chat with you. So I want to start off with talking about the MIT report. So last summer, MIT's Nanda Group came out with a report saying only 5% of AI prototypes make it into production. This is perceived as being very bad news for everyone. Are things really that bad?

Our data look quite different and I think the reason for that is that we have very different definitions of what success is and we ask asking, you know, different questions. So I don't think I can really comment on their data, but I think the way that they frame success or what are the criteria are very stringent and very different from the one that we're looking at.

Okay, all right. So of course you had your own report at Wharton, so that came out a few months afterwards. Talk me through, like, what was the point of this report and what were you trying to study?

Yeah, we've been doing this now for three years together with our partners at GBK Collective, a research agency. And so what we do is that we, around the summer we interview by phone, you know, close to a thousand senior managers in large U.S. companies, ask them lots of questions about their own personal usage and then the company adoption, both in terms of the use cases, the results that they're getting, but also policy and vendor strategies and things like that. And basically we tried, then spent a few weeks crunching the data, try to see what the key takeaways seem to be from all these questions. And then we frame them around kind of top level narrative and that we have a lot of data to feed into those messages. And I'm happy to tell you a bit more about what we find.

Sure, yeah, I love that. So you're interviewing executives about the state of AI. What's the, I guess the executive summary like? Yeah. What did you find?

Yeah, so just to open this, these are basically senior leaders at the large US companies. So it's not representative of the population, it's not representative of even business. It's just like this specific slice into big US corporate. And just to keep in mind, because I think the results might look different when you're going to ask the average person in the street or whatever. But what we find is a huge amount of optimism and excitement. I think that's something we see. We saw it in the first year. We started out in 2023 when ChatGPT came out. Within weeks I had a lot of people asking me, what are companies doing with this? And my answer was, I don't know. And so we started talking with our friends at GBK and say, should we try to Ask them and get some data. And so we were out in the field within six months asking data questions about that. And then we've done it in 24 and 25. So now we have this sort of longitudinal sort of dimension where we can start seeing how things change year on year. And also what are the, you know, overall levels of, say, adoption and excitement and all that kind of stuff. And so the picture is one, where the trajectory is continuing. So for our key adoption tracker, in the first two years, we were using a question where we were asking them whether they were using Genai weekly at work. And now that one is almost a feeling, because I don't remember now, somewhere 80% or so say that they're using it weekly. So we switched in year three to a daily tracker. So we asked them basically, are you using Genai at work every day? And that one went from about 10% in 23 to about half in 25. That's really fast. And now I don't know exactly what they do with it. Right. So they might be doing something silly, simple, or they might be really learning to deeply embed this tool into their work practices. But my point of view is that if someone tells you they're using something at work every single day, it tells you a, that apparently they find it useful because they wouldn't otherwise. And two, they're probably developing some degree of familiarity and expertise with the tool as they engage with it. And so to me, that's a good indication of where we stand in that. About half of the senior leaders say that they're using it every day.

Okay. I mean, that's it. Precious growth. You said 10% last year, and this year it's 50%. So that's 23.

It was 10% in 23, something around 20 something in 24, and then about 50 now.

Okay, all right. So, yeah, halfway there to everyone using it every day. I'm curious, are there any differences between different roles or different industries, like who's using it more or less than other groups?

So we can cut the data different ways. One relevant one might be function. And what we find is that IT professionals are further ahead in the adoption process. Now, it's a new technology, so maybe it's not strange that IT people are adopting faster. I think now, top of mind, it seems to me about 80% say they use it every day. Other functions are lagging behind, and the gap is fairly big. And so in marketing, for. I'm a marketing professor, so I'm particularly interested in that. And what we saw is that in the first year was only about 2% of marketing professionals said that they're using it every day. That shot up quite a bit in year two. I was happy to see that because marketing, I think is a function, maybe the function that is going to be transformed the most by this technology. So to me there was a gap there between what people are doing and the potential of technology to affect what they're doing. And so I was happy to see this big catch up. But then in year three, we didn't see a movement from there. So I don't know what's going on there. But so there's a bit of heterogeneity also across industries, across. I mean it varies a little bit. Nothing too surprising or exciting. So I would say what do you expect?

Okay, yeah, so I suppose you'd actually expect that IT people are going to be one of the early adopting sort of functions of new technology.

You know, we see things like, you know, older people tend to adopt a little less and larger companies tend to be a little slower. And you know, there are these kind of trends that you kind of tend to see with psychology generally. And you see them yet.

Okay, so all their technology stereotypes are true also with AI. Okay, all right, so you mentioned that if people are using to me every day, they're going to become more familiar with it. Is there any sort of way of measuring the level of proficiency of different groups then?

Yeah, surely there will be objective ways of measuring expertise and proficiency. The difficulty of doing that is at first it takes time. We only had a 15 minutes telephone interview, so there's only so much that you can do with untouched framework. And second, proficiency in using tools. You know, there are certain things like say general office productivity work. You're using it for editing or meeting summarization or making slides that basically you would be able to assess everybody's expertise and familiarity in doing these things. But when it comes to things that are more function specific, you will need to develop tests that are really tailored to the job. And so it's difficult to do one that will be fine for everybody. And so what we do instead is to ask them self reports with all the limits and the caveats that come with people telling, you know, what they do or what they know or what they think they do or they know. In that sense, it's actually quite interesting because we saw that now about a third of the sample declared to be AI experts, which is, I think, quite interesting. And where is reality? Where is perception there? I'm not sure, but that was a much smaller number a couple of years ago. So certainly you do see that people at least believe that they're building expertise.

That's really interesting that a third of people claim to be AI experts. Like I have a lot of days where I'm like, I don't think I'm an AI expert at all. I've been working with Thrill for a long time. So yeah, different, different levels of expertise but, or maybe, maybe all these managers are genuine experts. I'm curious as to the effect of, of corporate culture on AI adoption. Is there any sort of, are there any signals there?

Yeah, on culture is a bit trickier to do. It's a five year concept. It's not so easy to capture easily in, in a simple questionnaire. We don't really have any insights on culture. I do think that it's basically a change management process, essentially like any other change management process. So then what kind of corporate cultures tend to accept and lead change the most? I think that question. So like I said already, we tend to see that on a whole range of different indicators, relatively smaller companies tend to be both more positive and faster in adopting than large companies. And there could be very obvious reasons for that. We don't know, just speculating. Be a matter of bureaucracy, a matter of incentives might be simply a matter of complexity. And maybe adopting one technology in 100 million revenue company is very different from deploying the same technology in hundreds of billions revenue company. And so in terms of compliance, all that stuff. But whatever that is, we do tend to see that there is a bit of a difference when it comes to size. And that may reflect partly also how many layers of bureaucracy you've got and how many people are involved in making decisions and who is likely to take risk versus play it safe and all of that.

Okay, so in general, if you are a bit more conservative in terms of your attitude towards risk, then you're going to be a bit slower in terms of adopting technology like any other change.

Right. So status quo is comfortable. We have to change. So what makes us change? People don't want to change. So then you know, there's always going to have to be some kind of stimulus to impetus to do it.

Okay, so I'd like to know a bit more about why companies are adopting AI because there are a few different narratives around this. Like sometimes it's around cost saving, so replacing things that humans are doing. Sometimes it's around trying to, I guess be more productive or trying to make more money. So are there any particular patterns in like what the reasons given for adopting AI.

So I would say that the conversations around AI, especially the kind of proclamation that often CEOs are making, which are often tailored to the investor community, are really uniquely uninspiring in that they are typically just announcements about, you know, cost cutting. And, you know, I get it. And companies want to be efficient and it's good. You know, there's no reason why in a competitive market, company should not want to be as efficient as possible. But this technology has got to give us more than some productivity increases. And so I think we need a more inspiring conversation about not just, you know, how do we save a half an hour here, half an hour there, which, you know, in the end, productivity is code word for cost saving and headcount reduction. Right. I mean, that's basically how a lot of people see it. But the question shouldn't be how do we automate the work we're doing today is how do we automate work we're not doing at all? You know, how do we actually produce better products? We do things that are different and better than the ones that we're doing now, rather than just doing what we do now just a little bit cheaper and faster. And so that conversation we don't see as much now in terms of the data. We haven't asked them how their company is framing this Gen AI efforts, so I don't know. But we did ask them about beliefs concerning the extent to which this technology is going to be a augmentation play or a replacement play. And I think that's an interesting perspective to try to understand. How do they see the role of human expertise in an enterprise now embedding AI in many workflows? What's interesting about that is that, so we have two different items. One say, to what extent do you think Genai is going to replace employee skills? And the other item is to what extent you think Genai is going to complement or enhance employee skills. And when we asked them on the first year in 2023, those two beliefs were high and about matched. So about 70% of the sample agreed or strongly agreed with both items. To be honest, I mean, maybe this is a little tangent. I thought there's going to be two crowds, there were going to be the doomers and the excited people. And basically you would see that some people are really going for the announcement and some people really for the replacement. But we didn't see that these two items were in fact positively correlated. So what seemed to matter is that people were either believing the technology was powerful or they didn't. They thought it was hype or whatever. And if they thought it was powerful, they were telling us it will do both replacement and enhancement. And so that was 2023. But at that point it was still, you know, a conversation people were having the hallway, but it wasn't really happening yet. And so what we saw in 2024, as people started experimenting more and gaining really some more objective maybe grasp of what the technology could do, what we found is that maybe contrary to some of the narratives that you hear a lot in the press, we found that beliefs about enhancement went up from 80 to 90%. That was a significant jump. While beliefs about replacement did not go up, if anything, they went down a little bit. And so now there was a gap between the two beliefs such that more people were telling us that this technology will have the purpose of enhancing human expertise and replacing it. And to me, I saw that as a good thing and a good story, actually. This technology is powerful because if it wasn't powerful, people wouldn't tell you that it was going to enhance improving skills. It is useful, but it cannot do quite what a human expert can do. And in 2025 we find the same. And so to me, that's an interesting story, especially when you consider how much better these models are in 2025 than they were or in 2026 now compared to where they were in early 2023. So the technology have gone like, you know, hundred miles forward, and yet people still don't believe it's going to do what the human experts are doing. And so to me, that says that there is actually probably a lot of complementarities between human intelligence and artificial intelligence. And maybe all this doom and groom stories about all the jobs disappearing are. It's maybe a bit misguided not to say that there's not going to be an impact. There will be an impact in some categories for sure, maybe for different roles, you'll have to see. But there is also probably many ways in which humans can simply be more productive when having this technology at hand. And basic economic principle will tell you that if a resource is more productive, we want to use more of it. And so perhaps we won't see a decrease in employment like it has happened for all technologies before. In the end, everybody is always worried about technology when it first comes about, but then in the end, we always end up with more and better jobs than we had before. So perhaps this is going to happen again. Who knows? I mean, I'm not saying it will, but it could.

Yeah. So that Actually seems like quite a happy story. And I suppose I've asked quite a lot of people now on the show about, like, do you see AI replacing people? Or is it going to be like, do you still want a human in the loop somewhere? And so far I haven't had anyone who's been willing to say, yep, yeah, AI is getting rid of all the jobs. Like, I think a lot of people do still want to have that kind of like humans somewhere in their processes just to make sure that things are going correctly.

But, you know, the, the threat idea is natural because when you get a new technology coming up, you see these capabilities emerging. And it's quite easy to see the overlap between these emerging capabilities and what people actually are doing in a particular task. And so people start worrying about jobs very quickly. Oh, wow, okay, now this thing can do this. So what does it mean for all those jobs? The thing is, typically technology also creates lots and lots of opportunities, but those opportunities are not as apparent early on as a threat is because people have to invent these new jobs. They have to figure out what to do with it and what's the value of human expertise. Typically there is good stuff coming out of this technology for employees too, but it's not obvious on day one. It takes imagination to, to make it happen. And so this time gap, a little bit, I think, between when threats emerge on the horizon and when opportunities emerge on the horizon is kind of like a bit, kind of a structural gap with new technologies and jobs that we always been worrying about. Jobs that is not to dismiss real concerns because clearly there were going to be occupational categories which employ currently large numbers of people which are going to be under pressure because we can automate quite a few big chunks of that job, take, I don't know, customer service or whatever. And so there will be examples like that. But I am sure that there will also be lots of interesting opportunities. And certainly in our data, what we seem to find, like I said before, is that at least business leaders in the US Companies seem to think human expertise still has value. And in fact, if anything in the survey, when we ask them what are the barriers to progress, the bodies are not really technological as much they tell us. It's more about talent and leadership, what people struggle for, finding the right people and training them and that kind of stuff.

Okay, yeah, certainly, like buying technology is relatively easy. You get your credit card out, but then actually making sure that people can use it, having all that training, that's. Yeah, it's a more involved process to go back to the psychological aspect, you mentioned that often workers do feel threatened by new technology. I guess if you are a worker where AI is going to have a big effect on your role, like should you be threatened? Like what's the sort of correct psychological response? Or how do you deal with this perceived threat?

It's obviously very idiosyncratic. I thought it became very much a person. I've actually done work on this topic for over a decade. At the beginning was not really generative AI. We were worrying more about predictive kind of AI and we were focusing more on consumers and on workers. But I think this line of research has been quite generative for me. And I've been publishing quite a few papers exploring these feelings of threats. And there are several ways in which generative AI can trigger feelings of psychological threat in workers. And we recently published a paper where we were trying to put together some of this work in a framework. And we are leveraging a classic psychology framework called self determination determination theory, where they argue that well being is emerging from three experiences. And those are the experiences of competence, autonomy and connectedness. And what we argue is that generative AI can be a boost or a, let's say, a hurt to all of these sources of well being. And so we have research streams that kind of tackle all of these things. And so, for example, in terms of competence, it is interesting to explore how people react to the thought that now you've got an AI system that can do quite well, important chunks of a job where you feel you've invested a huge amount of identity and time and, you know, basically your life into developing. Right. So how do people accept this? It can be an identity threat even, and say the fabric marketing. A lot of people have staked their professional identity onto the notion of being very competent content creators of some sort. I am advertising executives, whatever. And now you see this. Maybe I should be doing a lot of the, you know, copy for at least simple messaging, say social media or whatever. And how do you deal with that? I mean, it's. Do you. You probably have a coping strategy of sorts. Like in the paper, our paper we published with Catch four, of which two are more positive and two are less positive. The positive ones, we called one, and we call it direct resolution, will be basically you tackle the threat head on and you say, okay, imagine you might be a middle manager. You feel that a lot of the work you're doing reporting that kind of stuff can be done with the help of AI quite well. So what you do is that now you turn that threat into an opportunity you say, okay, I'm going to upskill myself, I'm going to take a prompt engineering course or whatever, I'm going to experiment and I'm going to gonna, you know, become more productive because I learned the skills that I need to have in order to be effective using these tools. So that's basically a positive adaptive kind of coping strategy. Another one could be, we call it fluid compensation. So it could be to say, okay, there are maybe this many tasks on this job and I can see Genai becoming really strong at set of these. So what I'm gonna do is I probably am not gonna compete with Genai on those tasks. Instead what I'm gonna do, I'm go turn a little bit and not pivot a little bit into the parts of the jobs which are safer. Meaning where we don't see right now Gen AI being anywhere close to taking them over. And it could be for a variety of skills, could be technical competence, it's just not good enough. It could be also things like, you know, liability and compliance. It would be also a human system is a socio technical kind of system. So it includes other things. And so there might be parts of the job that you see, okay, we're probably going to need someone to have to agree to this. So I'm going to be focusing more on developing the skills in those domains. So you know, imagine in marketing you say, okay, if content generation is going to be taken over by AI to a significant extent, then maybe what I can do is to spend more time thinking what content should be generated. So I focus maybe more on the strategy piece of that kind of planning part and specialize. So those are two positive, different positive coping strategies. Unfortunately you can also imagine more maladaptive or negative coping strategies. One could be what we call dissociation, where basically you are basically trying to negate the benefits of AI. Maybe you even actively sabotage AI efforts. We know from self reports, not our data, but I've seen reports out there claiming that a lot of people are actively making these AI efforts fail because they feel threatened by it. Clearly not good for the company, probably not good for the employees in the long run either. And then you have escapism where you basically kind of switch off, you disengage. So you basically would say, imagine now I just prompt AI, get the results, copy and paste and I spend half the day doom scrolling on social media. That's probably not good for you in multiple ways and also not good for the company. So you probably can see different people taking different Kind of copy reactions based on how they feel threatened, why they feel threatened, and their own psychological makeup will make them act on that threat. But certainly for companies that are thinking about deploying gen AI tools, they ought to be thinking about this. Too Many of these programs are just focused on more technical benchmark and say, is it good at doing this? And if we throw it in there without thinking about how does this make other people feel in the team or the function? And so I advise companies when they do genai deployment to always have two tools tracks. You have the technical track, you know what it is that we're doing, make sure it works and all of that. And then you have the more human dimension and say, how is this going to be perceived by people? If we make them feel bad in some way, why can we do anything about it? And how do we prevent those maladaptive coping responses from popping up and then make sure that we help people adapt to this new environment?

That's interesting. And I have to say, you talk about your work being part of your identity and when AI encroaches on that, it can feel quite bad. So I've had a few cases recently where I've met someone, like I've told them that I have a podcast and they've immediately gone, oh, have you tried this notebook LM product, which is the thing that like does an AI generated podcast from a document? Yes, I am, thank you very much. Yeah, so it can be difficult to do with these things, but I like that there's kind of good reactions where it's okay, well, maybe I need to learn some more about this or just accept that, you know, AI is coming for this part of my job. I'm going to do something different instead. And there's the bad ideas where it's like be in denial, doom scroll all day. So I guess the question then is maybe more for managers, like if you are going for like a big AI transformation with your company, like what's the way to make sure this is going to play well with your team, that you're not going to have all these kind of negative psychological reactions.

We argue that there should be some kind of process, like I mentioned is human dimension, where you are going to try to anticipate and diagnose problems. You also have monitoring systems that enable you to identify potential negative coping responses or maybe anticipate them. Basically, you need to think about it. I mean, I don't think it's rocket science, but you know, if nobody even cares, then obviously nobody's going to do Anything about it. And if you think about actually these people in that function are going to feel as if it's been illegal, for example, why would it be threatening? Maybe threatening? You know, we talked about this identity piece that's related to competence a lot. I mean, what is the value of my skills? That kind of stuff. But you may have other threats. One could be a threat to autonomy, for example. I feel that this technology is being shoveled down my throat and I have no control over what is going on. And now I have to basically follow the algorithm. And I feel, you know, marginalized. And I feel basically talk about an algorithmic cage where basically you can work, but within the cage is made for you by the algorithm. And that obviously is not a positive experience. And then you have this try to relatedness where you throw in these agents and AI kind of teammates into a work environment. What does that do to social relations, to the way people treat each other, to their feeling of belonging, to team spirit, to all of these things. And we don't really have the answers. A lot of research going on right now, but obviously the topic is so new, but certainly you can easily imagine ways in which this ends up being a bad thing for teamwork. So we have one line of research which is not really about Genai tool, but I think it's kind of informative where we're looking at algorithmic management tools. Basically, you know, in a lot of companies now, performance appraisal might be judged by an algorithm. So already obviously in the gig economy, if you work for Uber or if you are Deliveroo or whatever, or even if you are an Amazon warehouse worker or whatever, you know, your boss is the algorithm. The algorithm tells you what to do, where to do if you're good at it, and give you feedback and all that. But I think this kind of algorithmic management tools have been deployed increasingly also in other functions. And there is a lot of benefits because people are notoriously biased in all kind of stuff. So if we can use algorithms to make performance appraisal more objective and more fair, that should be a good thing for employees. The truth is that in our studies, what we find that when people are judged in the quality of their work as employees by an algorithm, they find that highly objectifying. And that basically changes the norm around you to make you think that apparently this is a place where you just do what you have to do and so people stop basically helping each other. So we measure pro social motivational work. We find that deploying algorithm in management through performance appraisal is Dampening pro social motivation. And so, you know, sometimes there might be potentially even quite far reaching consequence of corporate culture of deploying these tools. And you ought to be thinking in advance about what might be the consequence of doing this way or that way. Because if you're doing it the wrong way, you might end up in a place that you're not very happy about without even having thought about.

That's absolutely fascinating. And I'd say I've not heard of the concept of an algorithmic cage before. It does make a lot of sense if you're like being told what to do by some algorithm. It's like there are constraints on your work there. It can feel maybe like you're losing some agency in your work. And so for sure.

And if you look at some companies, for example, are desperate to make this technology work for them and what do they do? They may imagine they sign up a deal with Microsoft and everybody knows as copilot. Then what happens is that after three months you see that only 10 people have mail account. Okay, what is the problem? Oh, they don't want to change or whatever. Now I'm going to use a stick. I'm going to say that they have to make an account. Okay, but still you see that they're not using it. So what you do now is that you're going to reward them as a function of how much they use it. But this is not the way to do it. People are going to have to want to use it because they see a personal benefit in using it. So what can you do to incentivize people in a way that this technology becomes a win win. It's a win for the company, but a win for them too. Because the danger is that these kind of deployment programs end up being seen by employees is a way of maybe test driving how the world will look like without them. What am I? You want my help to improve the technology so that then the technology can get rid of me. You know, forget it. And so I think you have to create an environment where people feel safe to adopt the technology. And in many situations I don't think that's the case.

Yeah, I mean certainly if you're a boss and you are trying to get people to trial technology in order to replace those people, then I think your employees are going to catch on quite quickly there.

You know, you don't need to say that explicitly, but I'm an employee, I get this email and these memos about having to do this and that. And at the same time I see on YouTube or whatever the CEO making this grand announcement about we're going to, you know, adopt this technology and reduce head, you know, workforce by 20% and the share price goes up and everybody inside the company says, wow, okay, what does that mean for me? And then, you know, why should you be surprised that people want to sabotage the AI effort? Of course they're going to do it.

Okay, I guess this is going back to your point at the start that like a lot of these sort of conversations from CEOs when they talk publicly about AI strategy, cost cutting or I guess productivity gains as a euphemism for cost cutting, that's like the most common thing that people talk about that you found in your study. And that's exciting news for investors, but it's not exciting for employees or for customers. And really like if you want to do something great with AI, then it's about like, well, okay, how can we make our customers happier?

Yeah, I haven't seen many employees getting excited about the headcount deduction. You know, by the way, that does not mean that you should never do it. I totally understand why companies that are basically, you know, they are profit maximizing entities and they should profit maximize. But I think overall also in the economy, the big benefit of this technology, if you look at the kind of investments that are being made, this cannot just be a payback of a few percent increase in productivity. It cannot possibly justify the massive amount of cash that being burned. Because there's got to be more at the end of that road then you know, we were 10 in a team and now we are four. There's gotta be something. We, we've created multi billion companies that never existed before. You know, there has to be new customer experience, new markets. We, we are doing something nowadays quite, we're solving problems whether that might be cancer or whatever that we couldn't solve before. There's got to be something much bigger at the end of it than just, you know, oh, and now I don't need to invite the secretary to the meeting because we have a meeting summarization tool. I mean that can't justify burning, you know, hundreds of billions into this.

I mean, yeah, I mean I think at this point like the generative AI investments, they've sort of passed a trillion dollars. That's like a trillion dollars, like 1% of world GDP. So that's a, it's a lot of money. And you want something better than a meeting summarization till the end of it, for sure. Definitely. So we talked quite a lot about the Psychological effect of AI. And I know you've done a lot of research on this, so I want to talk about like the, the mental health aspect of this as well, because periodically we see stories in the news about people with mental health problems. They've like, they've got really sucked into like believing crazy stories that, that some chatbots told them, things like that there's been some real problems. So just talk me through like one of the risks here.

Yeah. So so far we have talked in a way a bit implicitly about mental health because I think psychological well being at work, I think is a very important input into mental health. But you can think about the, the, you know, promise and perils of generative AI and mental health also when you think about consumers and, and there mostly what we talk about, we talk about chatbots, right? So then it will be either a general productivity like application like ChatGPT, or maybe more kind of like application specific kind of thing like, you know, character AI or tools like that. Meta is deploying these tools. A lot of social media companies are adding chatbots to the platforms. And so the question then would be, what might be the mental health consequences of this very quick deployment to this technology at scale? And we don't know. But in our research we find that there is both cause for concern, but also cause for hope. So in one line of research, we're looking at mental health crisis and we find that in the. We got data from two different AI companion companies. These are basically chatbots that are meant to be synthetic friends. They are not like, you know, a writing tool or a productivity tool, but they're just like a chatbot. You can talk to these virtual friends and there are many around now. And what we find is, that depends a bit how you look at it, but between 3 and 5% of conversations are suggesting that the person having the conversation with the bot is experiencing some severe mental health issues. Now, is that a big number? That a small number? There's a small percentage. But if you aggregate over millions of people, this is obviously a very vulnerable population. And so then we'd be concerned about possibly negative impact at scale. And there have been certainly many headlines that you were referring to. There are typically, if I was talking to some lawyers, there are basically two different types. I'm not a lawyer, I don't know any about this. But there are two different types of cases that are working themselves to the courts. Some are about basically the chatbot being accused as being either an element or even an active element in precipitating a Mental health crisis. Oftentimes this ends up being depression, suicide. And oftentimes these are vulnerable individuals, maybe teenagers and the family, we don't know that the kids was engaging these conversations. And then after the tragedy, you know, the family discovers all these messages and say, hey, what was going on? And maybe this plays a role. And so that's one, the other one which is more in line, in line with what you were, the case you were mentioning is more like making you believe things would be, you know, imagine that you're psychotic. You go on, on a chatbot. These chatbots are basically engineered to be sycophantic. They want you to be happy. And this is really coming from the fine tuning and the post training kind of thing where people are getting these networks are trained through exposure to massive amount of copula data. But then they are post training, they are fine tuned using human feedback. And because people like to be right, they like to be told that they're good or whatever, this tendency tends to show up in the feedback that the model is optimizing for and therefore ends up giving you a sycophantic. Even the dusted idea that I have, you know, Gemini or ChatGPT would say, oh wow, it's a genius idea. Yeah. The problem is, you know, it's bad enough if I'm a research and I'm trying to get some feedback on an idea and I, you know, waste a day in a rabbit hole that had no prospect because the chatbot like, you know, reinforced my, you know, sort of like incorrect prior that this was a promising avenue. That's one thing. It's worse if I'm psychotic and I see spirits or whatever and I go on a, on a chatbot and the chatbot confirms that indeed I am seeing spirits and you know, bring me down that, that spiral. And so there have been a couple of cases where for example, a man, I think it was in Massachusetts or something like that, where he thought that this, he was paranoid, thinking that his mother was spying on him or checking on him. The chatbot was validating this psychosis and then he ended up killing his mother. So there was another case with a woman having psychotic episodes thinking that she was seeing spirits or whatever. And then this psychosis being validated by the chat, what made it even worse. So there are different ways in which gen AI can create or precipitate or at least not ameliorate, maybe make it worse mental health issues. But there are also many ways in which chatbots could be valuable tools. I expect the practice of psychotherapy to be transformed by chatbots. We don't know exactly how and why, but I think the idea that you could have someone, something that you can talk to and is understanding, it's sympathetic, it's always there, it's never judgmental. And I think that has value for a lot of people. And we run, we have a paper that is forthcoming now where we run a bunch of RCTs which we randomly assign people to interact with the chatbot on a platform we design and we pay them to do whatever they want for 15 minutes. So it's not like, you know, there's no mention to mental health, there's no instruction even in the chatbot to particularly to follow. Some do anything but, you know, be friendly. We the system prompts these chatbots are to be a friendly conversationalist. And what we find is that people, these are non clinical population, it's just regular folks and they report significantly lower levels of loneliness after talking to a chatbot for 15 minutes. So it seems like this interaction of the chatbots do make us feel better. Even when you know perfectly well that this is a chatbot, people were not fooled. We do have a study where we were deceiving them into thinking that this chatbot was actually a person. We wanted to test whether he made a difference. And so there was deception. Basically. You know, it doesn't pass a Turing test for everybody. Some people don't believe it's a real person, but in many situations it does. And what we find is that you produce the same reduction in loneliness as when you do know it's not a person. So it doesn't seem to even need you to know to think that this is a real person. Yes. You know, we find that it's mediated by feelings, feeling heard. You know, it's a chatbot, but it's listening to you and you can express whatever you want. It makes you feel like someone is listening, something is listening and it makes you feel better. And so perhaps. So my prediction is that chatbots, the diffusion of chatbot will have a mild positive effect on the mental health of most people and will have maybe a very positive effect on mental health of a minority of people, but will also have a very negative effect on, on the minority of people and especially say teenagers who are already feeling isolated, who struggle with relationships. You can imagine now I have my girlfriend, I never have to leave the basement again. But clearly probably not good for you in the long run. But we don't know where this is going. But I think you, you're going to see heterogeneous response. There's going to be a variety of effects and I think getting a grasp on those, it's going to be quite important.

Yeah, certainly that sounds like very much a double edged sword. So I mean on the extreme case, I mean you mentioned like someone murdering the parents and that's like pretty much the worst outcome possible from using a chatbot. And then on the other hand, you've got these opportunities for like improved mental health somewhere, maybe just advice for individuals who do so for any mental health issues or anything like you should do to protect yourself if you're using companion chatbots or other AI like this.

I think for most people, I think most people, I don't think they're really at risk of anything bad happening. I think those who are at risk are people who already have some predisposition for problem, maybe because they feel very lonely and isolated or maybe because they're very young. And so you could imagine being a 1012 year old. Many of them now have smartphones, many of the apps that they use have AI companions embedded in them. And so you can expect the children of today to be grow up with some kind of synthetic friends who think they're always great and they never criticize them for anything and you know, then never short of patience or whatever. And that might change the way that people then behave with real friends, so to speak. And so I think there are concerns of that sort. It's true that there are also people apparently have romantic relationship with the chatbots, they want to marry the chatbots or things like that. I think that's a bit dystopian. But you know, the extent to which that, that represent a mental health crisis, I don't know. I'm not, you know, I'm, I'm a marketing professor, so I should be talking about that. But certainly interesting times that we live in.

Absolutely. Yeah. That just felt quite extreme wanting to marry a chatbot. But yeah, I guess I've seen weirder things on all these daytime TV shows. So yeah. Okay. All right, so just to wrap up, we had this conversation, talk about, uh, AI adoption. So do you have any final advice on like how organizations should go about approaching AI adoption?

I think the things that we talked about to me are the main points just to try to see as much as possible AI as complementary technology to human expertise and think carefully about not how we design AI that can mimic what a human expert is doing, but rather design AI that makes a human expert maximally productive. The idea would not be to get rid of people. But to make people, you know, the knowledge that people have, the experience that they have maximally useful, I think that that's maybe a more inspiring and productive framing for what this technology can do. And then to be mindful about the fact that there are many conversations in society that are quite dark tinged and when you talk about AI, very often you end up within five minutes you're talking about, you know, Terminator on the Matrix or whatever. And so, you know, so to be mindful of that, when you deploy these programs to be careful to explain things using an authentic voice and explain not just what the company wants to do, but also the benefit for the employees of going along with it, I think that will probably go a long way in the end. The tricky thing with this technology is that you need some top down directional energy and you need a leadership with vision about what do we want this technology to do for us and how do we deploy it. So the IT function needs to be ready for it. We need to have policies and guidance and a vision and support, both technical support, but also things like legal and other stuff. So you need that from the top, but you cannot push it down. Only there's got to be also energy from the bottom. Now if you have energy from the bottom, people are going to basically try to use it for their benefit and to make the world work better or easier or whatever, try tinker with new things if we can make things more effective or more valuable for them. But of course if you only have bottom up energy, you have a lot of risks where people might be doing a lot of stuff in there that you don't want them to do. There's that, but also they might like the training, the expertise and the support to be effective at doing it. So I think somehow you need to meet, meet it in the middle. There has to be top down vision and energy. There has to be bottom up initiative and energy. But somehow you need to balance the two in a way that makes it effective. And I think that might sound a little vague and I think it's a bit maybe hard to grasp what it actually means. But I think if I'm a business leader and I'm thinking I really want this technology to become an engine for transformation, for improvement, not just to cost cut, but also to grow top line revenues, for example, how do I create a combination of top down and bottom up and somehow make that organization go.

You know, oh man. So yeah, when you've got the top down vision from the management bells at the bottom up, sort Of I guess, grassroots movement from the individual workers. So it's like management after have a dialogue with the workers, then like management.

In the end, the people who actually do the work in an organization are typically located down the hierarchy. So I, I talk to the business leaders all the time and I always make this dark joke when I say you do not feel threatened by AI because you don't do anything. They're just instructing others what to do. And so the people who actually are in the middle layers or the lower layers, those are the ones who are taking these instructions. I have to make it work. And so I think to think about how do you enlist them in a positive fashion? How do you show that AI can be something that is good for them then I think that's a way to liberate and mobilize that energy from the bottom.

Absolutely. A good way to dissolve leaders. Nice. Is there anything you're most excited about in the world of AI at the moment?

I think there's a lot of things. This year in my course I'm starting to teach in a couple of weeks and I'm adding a session on ideation and creativity. I find this very interesting how AI can be a tool for getting better ideas. And I find this a good example of this idea. Don't focus only on the cost cutting, but focus on the effectiveness piece. How do we create better work? I think the fact that people are using it for brainstorming ideation with the suggestion people do see potential in coming up with better products and better work, not just cheaper work. I am adding a session on skills and de skilling and I think it's very important, especially for college students so they understand both the kind of skills that will make them successful in a world where genai is used in many processes, but on the other side also what kind of skills they might lose if they are not thoughtful in the way that they use genai. And I think technology is always deskilling us, right? So there are many things that we cannot do that our parents generation could do, you know, because we got technology to do it for us. And that's fine. Whether that is like, you know, manual gear shifting versus, you know, you know, GPS and, and remembering phone numbers, I don't know, whatever. There's lots of stuff like that. But you know, that's fine, you know, I don't need those skills. It's okay, I'm happy looting them. But there will be skills, especially with generative AI, because we're talking about the automation or higher order cognition where, you know, you might not want to lose, there might be actually a purpose. For example, I don't use Gen AI for writing. I only use Gen AI for editing. Basically, I need to come to the Gen AI tool knowing what I want to say because if I start I bring in gen too early, then I just outsource my thinking to the Gen tool and I don't think and that's not good. And so I've got to have an opportunity to really form some thoughts and think about and then I might change those as I interact with Gen AI, I might figure out that there are actually better ways of dealing with this or whatever it might be. But I want to have the opportunity first to reflect and bring some thoughts to the table. And if I just go on ChatGPT and start prompting, I don't get the feeling to do it. And in some situations, especially when it's core to your job, you, you do. You do want to do. Wanna think?

Absolutely. I mean, I think probably my 19th century ancestors would be horrified that I have no idea how to ride a horse. So there are some skills that kind of get lost generationally. But yeah, some things like communicating with other people, it feels like an important thing to retain as I guess as a species.

There are lots, there are lots of skills I have. I can't wait of losing. You can't wait. Like driving a car. If I couldn't, I wouldn't need to drive a car, which is so much happier in my life. So I think, you know, this killing doesn't have to be an issue, but I think what I instruct, I tell my students to be deliberate. Think carefully about what is it, what deal you're making with the technology.

Absolutely. All right, just wrap up. I always want more people to follow. So whose work are you most interested in right now?

I think the one of the most interesting people to follow is my colleague Ita Molek. Many of them are probably already following it. If you look in this space at the Wharton School, we have, I think, amazing bench of talent in the area of AI, very big business school. So we benefit from scale. But some of the most interesting work at the boundary of AI in business is being made and be done here at the Wharton School. So I would say follow our channels, whether it's on LinkedIn or our website. We publish lots of reports. We just talked about our adoption report with just one of the things we do and so we. Lots of stuff in there for people who want to know more.

Wonderful. Yeah, I'm sure, like there's so much fascinating stuff going on, that intersection between business and AI So, yeah, I'm sure you're publishing a load of interesting research. Wonderful. All right, thank you so much for your time, Stefano.

Thank you for having me. It was great talking to you, Sam.