(bright music)
The goal for this paper was to really help people understand this literature about when people wanna use algorithms to make predictions and decisions and when they don’t. There’s so many papers and there’s so much of this research that I think if you read through hundreds of these papers, you’d probably just come out confused because a lot of it seems contradictory and hard to reconcile. And so, because I’ve been in this space, I decided to write a paper that really shares a framework to help people understand when might people be more or less likely to use an algorithm to make a prediction or decision, like how long it takes to drive somewhere, who’s gonna win an election or a sports game, or anything else you could think about predicting with an algorithm.
This framework starts with the notion that fundamentally, algorithms are just tools that are designed to help us or supplement human decision making, right? And from this perspective, the way that we’re thinking about this is that an algorithm is really something that people use if they think it’s going to be helpful and help them accomplish some task or goal. And they might not use the algorithm if they don’t think it’s helpful.
So given this perspective, what we really need to understand to think about will people use an algorithm, is how do they define what’s helpful or beneficial, or what is their goal for their predictions or decisions that they’re making? So I turned to some of my recent research where I learned a lot about how people might define prediction performance and what goals people might pursue when they’re making predictions.
And it turns out when laypeople are making predictions, what they really want is they want to achieve frequent, near-perfect predictions, right? So the way they define prediction performance is when they’re using an algorithm to make predictions or decisions, they simply want that algorithm to give them pretty much a perfect answer a lot of the time. Once we understand that that’s people’s goal, then we can kind of think through, well, when will people have this belief that an algorithm might be perfect a lot of the time, and when might they not believe that?
So we kinda discuss this in two different branches. One branch is what will people do when they don’t have experience with the algorithm or haven’t seen it perform before? What we find is that when people choose algorithms without having any experience with that particular algorithm, their choices are really based on their beliefs of how it might perform. People often think that an algorithm might perform really well when they haven’t had any experience with it, but there are some exceptions. So there’s some domains where people’s intuition or beliefs are just that algorithms might not be able to perform well.
Some example of those domains are really subjective domains, right? So one would be, for example, moral domains. If we’re asking an algorithm to make some moral trade-off where we’re having to decide something like who should get an organ that’s been donated, or who should be allocated some resource or some benefit or some harm, people often feel that that’s an inherently subjective decision because we need to use our moral values to think about what’s right and what’s wrong. So in other words, in really morally relevant domains where we’re having to rely on moral values and make very subjective decisions, people don’t have the intuition that an algorithm can really reflect their preferences and understand their nuanced moral values.
Another domain where people might not think that algorithms perform pretty well from the literature are really hedonic context, right, that are just about pleasure and enjoyment. So thinking about what you find funny or what kind of vacation destination you might like, people just might not think that algorithms can really capture their preferences or predict what they really like.
Really, if we’re thinking about when people have a belief that an algorithm will perform well, if they have the intuition that it can be broken down into a math problem, and that can make a pretty good prediction, people tend to think that algorithms can perform really well.
So other times people have actually experienced the algorithm’s performance. And our framework makes really concrete predictions for once people have experienced how an algorithm performs, how they might decide whether or not they’ll use it in the future.
So what we found is that people’s objective is to frequently make perfect or near-perfect predictions. And so pretty simply, people might be very willing to use algorithms when they can experience them making frequent perfect predictions. And when algorithms are unlikely to make frequent perfect predictions, people might just abandon them or not use them as much.
This really hinges on how deterministic the domain is. So what a deterministic domain is is one where random chance or random error doesn’t really play a role in outcomes. So one really good example of a deterministic domain would be solving math problems, right? If we’re talking about really basic math problems, if we do the math the right way, we can always get the right answer, right? And this means that we can build algorithms that are essentially always perfect. That’s what a deterministic domain is.
And we see that people do adopt algorithms in many deterministic domains, right? People really like using calculators. People use computers to just carry out operations on data that are deterministic. People really like, you know, using the internet to look up established facts. And so in these domains where it’s theoretically possible to always be perfect, if you’ve built an algorithm that works really well, people are happy to use it because now they can adopt a tool that’s always gonna give them the right answer.
There’s other domains where there’s just a little bit of randomness, meaning you can’t always be perfect, but you can still be pretty close to perfect a lot of the time. So one example of this would be weather forecasts, right? We all know that there could always be an unexpected rain, or the temperature could be a little bit different than we expected, but largely, weather forecasts can pretty reliably give us predictions that are very close to what ends up happening, right? And people’s experience tells them that if they rely on these weather forecasts, they’re often going to be getting predictions that give them very good pictures of what’s gonna happen in the weather, right? And people are happy to use weather forecasts or things like driving time predictions, because even though they’re not always perfect, and sometime there are errors, very frequently, they recognize that these things will give them predictions that are just so close to the truth that it’s very clearly better than they could do in terms of frequent near-perfect predictions.
On the other extreme, there’s more random domains that are really not deterministic. This means that random chance plays a large role in what happens. In these domains, the literature shows that algorithms still have big advantages over humans, but people are reluctant to use them because they’re unlikely to provide frequent, near-perfect predictions. So my favorite example of this would be blackjack. It turns out there’s an algorithm that that you can use that tells you exactly how to play blackjack optimally. And in Vegas, you can actually go buy a little card and bring it to a blackjack table, and it tells you under any scenario what the optimal decision is. The problem is, even if you’re playing blackjack by the book this way, following the card, you’re gonna lose about as many hands as you win because it’s just a random game and you can’t make perfect predictions. So often you can’t predict what the card is coming next, or there’s no decision you could make that’s going to lead to you winning. And so people really quickly decide, “I’m not gonna play blackjack by the book,” or they abandon this algorithm because it’s just not achieving the performance that they would like.
So in these domains where performance is really random and uncertain, even when algorithms offer a big performance advantage, they just can’t give people frequent near-perfect predictions because that’s just a level of performance that’s not possible given how unpredictable and random the domain is. So in these domains, people often will deviate from algorithms or choose not to use them because they just don’t offer the performance that people value and that people are seeking. So these are domains where, right, a lot of experts would look and say, “Maybe people should be using an algorithm, but they’re not.” And it’s just because the algorithm can’t achieve the level of performance that they want or expect. (bright music)
