Algorithm Aversion: Fear of the Flawed Machine

Using AI Effectively: Pairing the Machine With Your Own Brainpower

In this episode, Bryan Foltice explores how AI can be used really effectively only when it’s coupled with our own brain power. We discuss how conversations about AI often lead back to this core idea, and why it helps to name and formalize a continuum for where we and others fall in our attitudes toward AI. We explore the value of understanding prior research on these perspectives, building awareness of where others might be—like blind rejection—and how we can help move toward a more skeptical, exploratory stance that still involves dipping a toe into using the technology.

00:00 AI Needs Brainpower
00:39 Mapping Adoption Spectrum
00:46 From Rejection to Skepticism
00:59 Putting AI to Work
01:01 Closing Thoughts

Transcript

Copy

SPEAKER_00 0:04

Welcome to the Brian Fultis Behavioral Finance Podcast, where we unravel the mysteries of behavioral finance and unlock the secrets to making smarter, more informed decisions with your money. Now, here's your host, Dr. Brian Fultis.

SPEAKER_01 0:24

Welcome back, everybody. We're going to continue our conversation on using technology and how much should we trust it? Or should we never trust it at all? And that's where we want to wind up here having a good answer and having an actual framework on where we fit on this continuum by the end of this episode. And this is part two of the series on automation bias, which is putting too much trust in technology, to today's topic of algorithm inversion, which means when an algorithm gives you bad information, you have a very strong negative reaction to that, much stronger than if a human were to give you bad information. And so that's what we're going to be talking about today. So remember, as we said, automation bias, we were looking at the proceedings and this summary of automation bias using pilots back in the 90s. And we found that the two types of errors from the automation bias, omission, so not getting information from the cockpit or the computer, and also commission, where it gives you the wrong bit of information and you file or follow that blindly without checking your own information. Those are the two types of errors that lead to this automation bias. Same authors that I referred to last week, Kathleen Mosier and Linda Skitka. They both uh additionally, I made some extra notes that if you add a second person, that that does not necessarily help you alleviate it. It just basically gives you another excuse not to be vigilant and to just trust what the computer tells you, even though you would think another set of eyes would reduce that quite significantly. They also found that uh teaching about this automation bias actually does reduce people's bias and they actually will question and be more vigilant. So it's actually very encouraging for us because as we are learning about this, we're also learning that with awareness comes that ability to say that, hey, look, I'm using my type one thinking, my automatic way of thinking, maybe a little bit too much. Maybe it's time for me to really critically think or me be more critical in this particular area. And it's through that awareness that we can actually get to that spot where we can step in, be vigilant, or double check that the information that we are receiving from technology is actually correct. And so when we're talking about the messaging here, we have to realize that it is not about AI. Um it's a lot about our ego. And so we are way more willing to forgive other humans than we are machines. And I had this happen in my research work this summer, where I was writing, I'm writing this paper on uh withdrawal rate rules. And you're gonna hear a whole series on what I found this summer because I'm just on the back end of it. But I remember one of the days I was actually hadn't read this paper for a little while and just wanted to go back in and was asking ChatGPT to give me that summary of their methodology and what they were doing, because I just needed a refresher. I didn't want to go and find the paper and dig in and go through it all myself. So I was kind of leaning on them. And ChatGPT came back with a very complete, non-complete answer saying that this study actually went and did a a Monte Carlo Carlo simulation and that it was different from all the other papers. And it's through this the findings and then just knocked out like these bullet points and was fairly believable. You go, okay, well, that makes sense that um it may that somebody went to a Monte Carlo simulation with this data and that they would use it to basically recheck these findings. Until I had to step in and go, I don't think that's true. That sounds way off from the paper that I read a while back. Well, sure enough, I look at that paper and go, that's not even close to what they use previous empirical data, like like historical returns in their analysis. And not even one single mention, let alone their main finding around this Monte Carlo simulation. And I go, I felt so betrayed by AI, and I was really really angry about it, going, my reaction was way stronger than somebody saying, Hey, look, I think it's this, or or just basically lying, or whatever. If a human were to tell me, oh wait, I had that confused with something else. I'm sorry. You, but the way that AI presented that to me and and just gave it as this gospel, and I was like, I felt really betrayed by this. So what we're talking about here, that's what started. I go, man, I really had a strong negative response to this, saying, I'm never gonna have ChatGPT help me with any of these summaries. I gotta do it all myself, just like the old days. And it's like, okay, are you are you overreacting maybe a little bit? And and then I found out this is called algorithm aversion. Uh so the first paper to cover this, uh, what we would kind of consider the more seminal or or popular paper around this is a paper by um what do we have? Berkeley Diet Wurst, uh, Joseph Simmons, and Cade Massey. So I know Kay, uh recognize Cade Massey. They're all from uh uh Penn, so University of Pennsylvania. And um they wrote a paper called Algorithm Inversion. People erroneously avoid algorithms after seeing them. And so I go, that sounds like me. People erroneously avoid algorithms after seeing them error. Uh, this was a paper in 2015. This looks like Journal of Experimental Psychology, JEP again, what we call that in academic ranks. So a nice journal, and uh basically they showed within five different settings, they had an algorithm that would help predict future um MBA salary, student performance, future outcomes of graduates getting admitted into the school. And they compared that to people's judgment. And the way that they calibrated all five of their studies in this paper is they objectively had the algorithm as more accurate than any human prediction performance could give. And they also put in that error and they waited to see how quickly that response shifted away from algorithmic preferences to human preferences once they were betrayed by that error. And again, they found out that um one mistake, people immediately abandoned it. No second chances, no strike three, it's one and done, and it's a very aggressive move to the opposite way. And that's where we found this is algorithm aversion. And so we go, well, why is that happening? I've also had to think about that myself. But if you think you go back and go, you know, with humans, what do we expect from humans? And we're flawed, we're gonna make mistakes. When I think about ChatGPT, we think, well, they're beyond that, they're computers. And so my expectation level is they're gonna be perfect. And I always say I I love low expectations, I thrive in a low expectation environment. I hope you have low expectations for this podcast because then I can outperform and go, holy cow, I learned a lot. Um, once you get high expectations, like, oh, this guy's master expert, then you're gonna go, that's disappointing. He didn't really give me that much. Um, and so we think about these reference points here as well, and go, oh, well, that actually makes a big deal. It makes a difference. Because of human errors, you go, all right, well, it happens. Human nature, um, algorithm misses, that's below your threshold of perfection. That's negative reaction. We're gonna remember that very significantly. So it's uh a bit asymmetric here, whether information comes from a human or from the uh algorithm. So some theories that explain this. Number one, perfection expectations. This is uh also kind of tied into our reference points, what we consider gains and losses. That's called perfection expectations. Loss of control. There's another component here where we have this illusion of control where we want to be in control of our own decision making, our own destiny. This is why we do weird shit during our football games, because I think we're gonna control this by wearing the jersey or standing in a certain place while it's third and long because that's where they got first down on third and long before. Um, and so this loss of control when we're we're giving this uh away, we're we're losing our control, and and so that potentially, even if the computer is better, um, could explain why we dislike surrendering our decisions um fully to these algorithms. What about our egos? If pr the the computer is right and I'm not as right, that's a damage to egos. And the research has also found that more experts, so if you're an expert in your field and you're proven wrong or the computer outperforms you, then there's a damage to ego as well. So you've lost control, you've lost some ego. And this is where if you are in finance or if you're helping people um with their financial decisions, then all of a sudden you are in conflict now with your expertise and what AI might be telling the same clients and individuals. And this is where we're gonna have various various levels of humility around this. And there is a way to try to compliment them both, but anybody who's in the financial industry will tell you that their clients are most likely checking AI along with the responses that they are getting. Or maybe they're just blindly following AI, and the advisor has to somehow get them out and reroute them in the places that AI might be misleading them. Um, and then another one, this is the final one we'll talk about, is this uh betrayal aversion. And this is what I felt. I can't, I just have to explain it. People feel betrayed when technology disappoints them. And I don't have that same reaction when humans were to not give me the right information. Uh, and so we had some follow-up research on this. And uh, if we can actually have some sort of influence, I'm adjusting the algorithm, this can reduce our algorithm inversion. And this is where I found interesting with Chat GPT is okay, I've had this negative experience, I had a really strong and negative reaction to this, but now is there a way for me to go back in and teach this to the learning model where they don't get that wrong again? And so I use that as an opportunity to go back. I don't, I don't think I was very um tactful with how I told Chat GB team, but I told them it was total bullshit what they told me, and that next time, if they don't know the answer, tell me they don't know instead of giving me the lie. And that led to a real improvement not only of what the output was, it was also how I prompt, an improvement in how I prompt. And so garbage in, garbage out, if you give it better prompts, it'll give you way more insightful information that is going to be helpful. And so that's what I found. And this is what is helping reduce our algorithm aversion is being able to go in and make those adjustments. And that's what research has found as well. So in our world of finance, um, we now have robo advisors, um, AI is doing our portfolio management, our stock picking. So where do we fit in? Where do the humans fit into this? And I say our real value add is obviously from an emotional standpoint. Um, don't think AI is there when we're having that trouble of, hey, look, I'm ready to sell my entire portfolio. What do I do? What I'll have to check that with AI because now I like wonder what AI would tell me after all the different prompts it's told me to challenge me. You know, it always seems to revert back to confirmation bias of making me feel, hey, I think you're on the right. Well, that's a great idea. This could be one of your best episodes yet. Um I wonder how much it would push back if I say, hey, I'm gonna sell everything. Would it how far would it push me compared to a financial advisor if I were to call them on Monday and say, hey, I'm looking to go all cash, give me out, I'm ready to go. Uh so again, I think that's where the advantage lies with the humans. Uh at the same time, we can what else? Zillow. Zillow estimates. That's another one that makes you feel real good until you actually put it into the market and you realize that this estimate that Zillow gave you is not the same thing. Alright, so here's the thing. Automation bias, trust automation too much, algorithm inversion, we don't trust it at all. We don't trust it enough. Um, both are wrong. Now, where do we find ourselves in the middle here? And this is where we want to to to calibrate this. Now, the way I think about this is through framing, and I've had to take myself through this after my bad experience. Literally, it was like it was one time uh that it really, really betrayed me. But you can see it's still sticking here. Um we say, yeah, the machine will make mistakes, then we're gonna be averse to the algorithm. But at the end of the day, can we come back and say, look, which one is statistically going to be more reliable? And so the way we're setting this up, or where we're framing this question, is which one is gonna be more statistically reliable? And which one can I use, even though I can't have that expectation of perfection. So that gets me into this sweet spot of not 100% relying on it, but I've got to screen just enough where it's helping me, and I've got to reprompt my prompts to reconfirm. We talked about this uh on the last episode. Sometimes the AI prompts are like your teenaged boys, where you thought you told them once, and you realize that you have to tell them over and over again because they will quickly forget. And so you have to remind them about the prompts. Remember, you're gonna be challenging me, you're gonna be pushing back. I don't need to feel good all the time about what I'm trying to ask you. These kind of things. And um finding that sweet spot. So here's where we're gonna tie this all on. We already know. Judge algorithms by averages, not anecdotes. Um where humans still add value. This is where we talked about the emotional component. Um if you're not giving it good prompts, context a lot of that emotional decision making still here in 2026 favors itself in in favor of humans over AI. But again, AI's catching up, and it's all about how good or bad are are the prompts that you're giving it. Um and then calibrate your trust instead of making it binary. That's that's pretty good. And so as we wrap this up, the way I like to think about this is is within frameworks here. So we know we have the two posts, yes and no, total trust, no trust at all. Let's look at it as a continuum. We're gonna add a couple more pieces in the middle here. So blind rejection on one end, blind acceptance on the other. Way more than beat that one up. But in the middle, healthy skepticism or appropriate reliance. And you can see in between healthy skepticism is still closer to not trusting. Healthy reliance is over by the blind trust. That's where I would really, my ideal spot is in that area of that healthy reliance where you can still lean on it to really help you think better and to enhance your life, to make you more productive without putting your blind faith in it. So I had my um project that I was working on. I tell you all about the bad stuff, but I had a this project where I actually ran the analysis and was coming up with um Monte Carlo simulation on my own for this paper using 50,000 samples. And I had everything calibrated, all of the returns and deviations and correlations were all matching from different things, um, which are all green lights that you are set this up correctly and you have have done this right, except for one correlation. Oh god, why is this correlation off? So I go in. I was just like, I finally threw it into ChatGPT, and I was like, why is this correlation missing? And it found cell and one of the tabs, phase six, where I had put in the wrong like the the parentheses. So it didn't follow the mathematical um hierarchy, and it was coding it wrong. So I had one cell wrong, and it was all based on where I was putting my parentheses, and that was causing this error. And I was really kidding me. Found in within literally two seconds, I did it, re-ran it, everything all back in place, spot on. And so I'm going, okay, yes, you can use this really, really effectively. Um, but it has to be coupled with your own brain power. I think all conversations around AI eventually lead us to this. But at least I've really, in my own, for my own edification, I wanted to know what this actually is called, who's done the research on this before, and formalize the continue continuum of where do we fall ourselves on this space? And where do others fall? So then we can at least be able to have awareness on others where they fall. Oh, yeah, they are in blind rejection mode. And so maybe just trying to get them to a little skepticism uh around this, but still dipping a toe into using this machine. We can now use it in that way. So with that, I'm going to get off my perch here and move on with life. I'm going to forgive AI about its mistake. We'll continue to teach it. Um, but if you're liking what you hear, if you made it this far, thank you very much for doing so. Uh, we'll urge you to uh or encourage you to like, subscribe, comment. This gets the algorithm motivated to get more people involved in these hopefully really interesting conversations for you. Also, make sure right below in the show notes, you're gonna see a way to ask me a question or write a comment straight to me. You can leave a voicemail. It's real simple. And I'm starting to make my episodes around QA's, which have become very, very interesting to me. Newsletter also there on the moneystrong.net. You can go right there, get involved in a newsletter, stay up to date on what we're talking about, get it in written form, and don't miss anything as we're rolling out new behavioral finance concepts, topics every single week. So thanks again. Hope you have a wonderful day. We'll see you on the next episode. Take care, everyone. Bye-bye.

SPEAKER_00 22:54

Thank you for tuning in to another episode of the Brian Fultis Behavioral Finance Podcast. We hope you found our exploration into the fascinating world of human behavior and finance, both enlightening and thought provoking. Be sure to subscribe for future episodes. And until next time, stay curious and financially savvy.