AI For Hearing-Impaired Workers: A Good-News Story 

- There's no scarcity of dire predictions about the impact of AI on jobs and wages. What is scarce? Hard evidence on how AI is actually affecting labor markets on the ground. In today's episode, I will speak to two economists who developed some evidence on this score. If you like good news from the dismal science, stick around. Welcome. I'm Steven Davis, Senior Fellow and Director of Research at the Hoover Institution. With me are two card-carrying members of the Dismal Science. Yan Yu Chen is an assistant professor of economics at the University of Toronto. He holds a BA in economics and mathematics from the University of Hong Kong, and a PhD in economics from Duke University. Mitchell Hoffman is chaired professor of economics at UC Santa Barbara, director of the NBR Working Group on Personnel Economics, and an editor of the Journal of Labor- Labor Economics. A welcome to both of you. Thanks for having us, Steve. Thank you. You know, as I was looking over your CVs and getting ready for this talk, I noticed both of you are associated with beautiful places. Santa Barbara is a lovely place. I I gave a job market talk on there as a rookie economist, and they had the good judgment to make me an offer. Maybe I lack the judgment to take it, but. You seem to have done pretty well with your choices though, Steve, so yeah. I've done, I've done okay, but I, you know, I gave up a few things by not going to Santa Barbara on, in the amenities. And then, you know, the University of Hong Kong, I taught many cohorts of executive MBA students in Hong Kong for the University of Chicago's Booth School of Business. I've also visited the University of Hong Kong. I, it's another place, another university in a fabulous setting. Hong Kong's one of my favorite cities. And unfortunately, I, I haven't been back since Beijing imposed the national security law in 2020. And, you know, but that's a topic for another day. But, but really both beautiful cities and universities. So, so let me, let me turn to the topic of today's show. You, you guys have a brand new NBR working paper titled Empowering Inclusive Work. And what you do in this study is you, is you look at the impact of introducing an AI tool, which we'll get into, on the performance of hearing impaired workers in the food delivery business. Okay? And you've got detailed data for individual workers over time which lets you study the effects on productivity, earnings, hourly wages, customer satisfaction, profitability to the employer, and so on over time for these individual workers and in comparison to workers who, who don't have any disabilities. Okay? So great data. And, and, and this, this introduction happened long ago enough back that we can do initiating before and after comparison. So that's kind of your, that's kind of basically what you're doing. So I want to, I want to ask you now to just tell us a little bit more about the setting of your study and, and, and the nature of the job that these workers are perform- performing, and then we can get into the particular AI tool. So who wants You to - Yeah, do you want to go for it? You want to go ahead, Yanyu? Totally. So this is actually, so in China, there are two main food delivery platforms. So you can think about this in StoreDash versus Uber Eats. So we collaborate with one of them and they have millions of workers for the platform. And we get to observe very detailed productivity measures, like their speed in finishing every stage of work, and then how many orders they finish and a weekly level. And then we can compare their performance first pre-AI, like how the hearing impaired, which we refer to as disabled, like difficult in hearing or hard to hear, like DHH workers. So we can compare their performance to the non-disabled workers. And then we can examine, like, how basically the performance change after this AI implementation. Okay. And you, and you, you can look at each element of performance, including customer satisfaction. It's not just - Yes. Speed of process in the orders, but also the customer satisfaction piece, and that'll, that'll turn out to be quite important. So you said there are millions. This is a big sector. I just want to make that clear. In China, which is the context of your study, but in many countries around the world now there's a lot of people working in the food delivery business in one way or another. So millions of people in China, two big platforms that orchestrate most of this food delivery activity in China, if I understand correctly. Exactly. And you've got data from one of them. That's right. Okay. So, so just tell us a little bit more about the jobs that these people do. You know, a little bit more detail about the tasks so we can understand how and where the introduction of this AI tool might have an impact. So let's say I'm, I'm a delivery order guy and I, what do I do? I go on this platform and what do I do? Mitch, do, do you want to take the. Sure. Yeah. So it's mostly restaurant delivery. And I get orders and I deliver them to customers, but there's other types of work too I could deliver from a pharmacy or a grocery store, but it's primarily restaurant deliveries. So I have to read the order, put my, say, "I'll do this order." Then it comes back and says, "Okay, here's what you're supposed to deliver to where?" And then it follows each step of the process, including some customer service rating at the end, I guess. Is that, is that what happens?. Okay. That feeds back the. So the worker gets some, he gets paid something for this activity, job by job, I take it. Yes. And the the, the restaurant or pharmacy is paying something to the platform. The platform pays the worker, keeps some, keeps the rest, that's kind of gross profit. Is that, is that the, have I got it right? Yeah, I think so. Yeah. Okay. Okay. So, so the job itself, you know, there's multiple elements to it, but in some sense, at the level of the individual delivery person, it's a reasonably straightforward job. All of the logistical coordination's being handled by the platform. Taking in the requests from the restaurants, pharmacies, et cetera, grocery stores maybe, and then making those requests available to the gig workers. They pick up the request and, and execute. Okay. So tell us why might there be challenges for hearing impaired workers in, in carrying this out? It sounds straightforward. Where do the challenges arise? Well, so Yanyo actually worked alongside some of the workers, so he, he can give a much more detailed answer. Yanyo, do you want to jump in on that or? Would've been more impressive if he's, if he said he, he did that on the side while he was a PhD student. Or an undergrad student. But anyway, but anyway, okay, so. Totally. Go ahead, Yanyu. Totally. Because we are studying those hearing impaired workers, like, we actually want to know, like, okay, how they actually interact with restaurant and with customers. So we spend a couple of days working with them full-time to observe the interaction and where the bottleneck may come from. So there are mainly two stages after they're receiving the order. They first need to go to the restaurant. And then after they pick up the food, they need to deliver it to the customer. For the first stage, like interacting with restaurant, this is actually very pretty straightforward. And it's very standardized. All the restaurant, they know how to handle the orders. So I don't observe a lot of difficulty in that step. But in the second step, when they go to deliver to customers, there could be occasions where they have communication issues. Because in China, everyone pretty much leaves in those condo complex. You need to first find which gate to enter into the condo complex. And sometimes you need the buzzcode to. So you need to make phone calls with customer and to have some communication of navigation and all that. So this turns out to be actually pretty challenging for the - Okay. So you arrive at the front of the building, you can't get into the building. You need to call the customer. And of course, if you're hearing impaired, there's already a challenge. Totally. Because you have to have a short conversation with the customer to let you up in the building or for the customer to come down and accept delivery. So that's an issue. Okay? Exactly. And then do the customers sometimes give feedback? They tell the agent, "Oh, no, this isn't the what I ordered and I need something else." Does that arise? They not this particular type of complaint, but customers sometimes complain about like, "I don't want the worker to knock on my door." So they will give specific instructions. Maybe they're in a meeting or whatsoever. Yeah, and they complain about the customer not calling them to notify them in advance like, "Okay, the order has arrived." Right. And for hearing impaired workers, of course, they cannot make the phone call without any help. So they sometimes get bad ratings from customer just because not being able to make the phone call. Okay. And so this is getting into the outcomes data already, but give us a little bit since you raised it, what's the instance, the incidence or the frequency of bad ratings from customers for hearing impaired workers as compared to other workers before the introduction of this AI tool that we'll get into? So bad ratings are rare in our data that they're not a ton of bad ratings, but the rate pre-AI is much higher for deaf and hard of hearing workers than for non-disabled people. So in our data, it's a pre-AI gap of about more than 30% where DHH workers receive substantially more bad ratings. But the rate overall is not very high. It's an infrequent event, but it's a very consequential event because the driver has to pay a fine. Oh, okay. Well, that's a big deal. Well, it makes sense. If everything goes according to plan, it's completely routinized. There's no need for any special communication between the delivery person and the customer. It's in those instances where something doesn't quite go right and you need to have some kind of communication and adjustment. And there, it sounds like that's where the disadvantage to the. I call them hearing impaired. You call them DHH, which means deaf and hard of hearing. Okay. So that's a big difference. So then there was the introduction of this tool. Tell us about the tool and how it works and how it aids the hearing impaired workers. Yeah, I knew you want to step in. Yeah. Totally. So the platform calls it intelligent outbound call. So it's like calling mechanism which the drivers can directly initiate from their own own phone and associate it with each other. So in the intelligent outbound calling, there are some default prompts and also customized prompts the workers can enter. And once they initiate this calling, then there will be basically a voice message or phone call directly made to the customers. And this does a couple of things. First, this will notify the customer, like this is a hearing impaired worker. And then what specific requests the workers choose to do. It could be customized prompts like saying, "I have arrived, but I cannot get access to the elevator. So please get me access to that." Or it could be asking about a buzzcode for the building. So the worker themselves can choose what prompts they want to put into that message. And one way that it uses. Oh, sorry. No, go ahead. That it uses AI to make the call sound natural. It's a big improvement over the robocalls of your that you may have had experience with during elections where you just immediately hang up because it's such an unpleasant voice. And there's research in the human computer interactions field that people vastly prefer AI calls to robocalls. I have a question here because my. Are they female voices within a certain vocal range? I understand that those are the most pleasant to the human ear. Is that what the. I'm just curious what the AI is using here. I doubt it's some gruff masculine voice. Very good point. I, I don't know if they have tested different versions of the vocal. Okay. But I believe the default is a female. Okay. Okay. So, okay, that's a little digression for my personal interest. But, okay. So from your description of this tool, it sounds like it would also be extremely useful to speech-impaired people. Okay, maybe that, that's less common. And by the way, I should have asked you this at the outset. I forgot. What's the incidence of hearing impairments in China and in other countries around the world? So there's different ways of defining that. So of course, you know, as people get older, their rates of hearing impairment go up. So we report in China that 8% of Chinese adults have moderate to severe hearing loss. There's a kind of a global prevalence rate of about 5%. But of course, as people age, the rate of disabling hearing loss is going up. So the WHO actually estimates that by 2050, about 10% of people will have some disab - yeah. Yeah. About 10% of people will have disabling hearing loss. So just, so these are big numbers. Yeah. So this means like something like a hundred million on out order hearing impaired people in China around the world. Half a billion or more. And given the aging of the population, the incidents that that's likely to rise. So this is not most people, but this is millions and millions of people around the world that have some degree of hearing impairment that could, of course, interfere with their ability to engage in everyday human interactions. And this type of AI tool, not your, this particular tool in this particular setting, but this type of AI tool that could aid hearing impaired and speech-impaired people in a whole variety of work and social settings, this is really strikes me as a big deal. This could be an improvement. It may be a modest improvement, but a modest improvement multiplied by hundreds of millions of people is a big deal. Is that, is that how you guys think about it? Yes. And, and, but we, we don't necessarily think the effects would only be, be modest. In our setting - Okay. They're very substantial effects and very big closing of the wage gap between DHH and non-disabled people. So, so we think that in some context, it could have very sizable effects. Okay. Well, let's, let's get into that. So tell us first about what you measure and what the differences are between the hearing impaired people and the people who don't have disabilities before the introduction of this tool. And then we'll get to the introduction of the tool and its parent effects. Yeah. So I mean, pre-AI, you know January has experience with this, but as, as, and you know, directly what he said in the data, on average DHH people are, are slightly slower, but the differences are concentrated in the last stage, the customer delivery part. The biggest differences you see are in the so-called left tail adverse outcomes. You know, much more likely to have significantly late deliveries, much more likely to have bad ratings, unfortunately. But kind of interestingly, the, the DHH people have higher labor supply. They work more, and they're much less likely to quit. So in terms of income, they actually don't fare that, that, that badly, Priya, but because they work more hours there, there is an hourly wage gap. So on average - So yeah. So what is that gap? What's the hourly wage gap before the introduction of a tool? 10%. 10%. Okay. So they're, they're basically getting compensated 10% less per hour. Yeah. And this is after you've taken out the effect, you've, you've accounted for the effect of these fines you mentioned earlier? Yes. Of, of the 10%, about 1%. So about 10% of the gap is, is due to the bad rating fines. Okay. So they get Fines and they're, that's one. And they're just slower, especially in this last stage of the process. Yeah. Okay. And you mentioned, but they work longer to make up really, you know, so that's not. It's, it's great that they can make up the difference and, but they have to work longer to, to make the same amount of money, is what you're saying. The attrition thing, you know, they don't, they don't quit as often, but that, that could easily be just because they have fewer outside options. So that, that kind of ha - that's a double-edged sword as well. So they do this job because they can make it work even though they're not on average earning as much per hour as the as the workers who don't have an impairment. Okay. So that's, that's, and also you should say something about profitability to the, to the firm. We actually estimate that they're more. Oh, sorry, go ahead, please. Go ahead. Yeah. We, we estimate that the DHH workers pre-AI are actually more profitable, and it's these differences in hours and retention that are really important. Okay. Because they're actually doing more orders on average pre-AI because, you know, such a substantial difference in, in labor supply. Okay. And is there a substantial onboarding cost to the firm, to the platform? Is how is, is turnover expensive? Is attrition expensive for the, for the firm that runs the platform? That's what I'm trying to get at. Not relative to other industries, but there's just very high turnover. So if you reduce that turnover - Okay. And there's a huge pre-AI attrition gap. So we have a pre-AI attrition gap of 41%. So if, you know, 41% magnified over a very large number of people is a big difference even if the onboarding costs are not that big. Yeah. Okay. So this is an interesting setting in the, in the following sense. These hearing impaired people are actually more profitable for the platform even though they're getting paid less per hour. It's not that they're getting paid less per task completed. They're not. That's everybody gets paid the same. Everybody has the same pay schedule, I take it, per task. Okay? So it's not that the company is doing anything, is treating these hearing impaired workers poorly relative to the other workers. It's just that because they have less attrition, that's beneficial to the company because they work more hours. So whatever cost that the company has incurred to employ these people to bring them on board is kind of spread out over a larger number of hours worked for the hearing impaired people. Yes. That's the source of the extra profitability, correct? Yes, that's right. Oh, okay. So that's the setup. And then you introduced this tool. When, when was the tool introduced? In mid - 2022. Mid - 2022. Okay. So that puts us two years past the onset of the pandemic, more than two years past the onset of the pandemic. I, I mentioned this because I, because certainly in many countries, and I, I'm guessing in China too, the pandemic itself was a spur to food delivery. Is that, is that, did it play out that way in China? A lot, a lot of people stopped going to restaurants for a while and there was some kind of stuck. There's more, there's more delivery now than there was before the pandemic. Maybe that, I don't know whether that was true in China as well. Mostly accurate, I think. Yeah, because my, yeah. Okay. So you, so this tool's introduced, and then what happens? What happens to these performance and profitability measures that we were talking about before? How do, how do each of them change? Daniel, you want to take that or do you want me to? No, please. Yeah. I mean, so, so you see very substantial effects. And they're all concentrated in terms of speed on the last stage. I mean, there, there's a small improvement in the speed at which they accept orders, which we think is also making them more confident. But most of the improvements are in the final stage where it closes about three-quarters of the pre-AI gap in the customer delivery speed. You see it entirely closes the gap in substantially late deliveries that was present. It closes about two-thirds of the gap in bad ratings. And it also tends to accentuate the areas where the DHH workers were doing well pre-AI. So they further expand their hours advantage relative to non-disabled. They further expand their retention difference. And because both their orders and their hours are going up, the effect on the hourly wage is a little bit more muted, but it's still very substantial. So it closes about one-third of the hourly pay gap of the initial 10%, which is still very substantial, we think. Yeah, it is substantial. I just want to restate that in other words. This is a, this is a clear, concrete example of where the introduction of this AI tool is pulling up the bottom towards the middle in terms of hourly earnings. Well, we don't. Sorry, what do you mean by bottom, Steve? Well, these guys were earning 10% less per hour - Yes. On average before the introduction of this tool. You said it closes one-third of that gap. So these people, in terms of, if you think of an hourly earnings distribution, the hearing impaired people are disproportionately occupying the lower rungs of that distribution. And that's being brought up towards the middle. That's how I'm. The reason I'm framing it this way is there are, you know, I, back to my opening remarks about dire predictions. Some of the dire predictions are not that jobs will disappear, but that all the advantages of new AI technologies are going to go to people at the top and leave everybody else behind. And what I'm stressing here is in this context, what's happening is just the opposite. You're pulling up a group of people, hearing impaired workers, who are disproportionately at the bottom parts of the hourly earnings distribution, and you're pulling them up. Absolutely. Yes. We - That's, That's a good news story from two members of the dismal science here. So that, that's my framing that I, I'm putting it that way to push back against all the doom and gloom about what AI will mean for jobs in the labor market. Yes, absolutely. Yes. Okay. And we. Oh, sorry, please. Go ahead. No, please. Just one important thing, and perhaps you were going to allude to is we, we do look at heterogeneity by our AI effects by different pre-AI measures. So we look at pre-AI productivity, pre-AI tenure, and you see essentially no significant heterogeneity by those things. So it's not in our setting that the AI tool is disproportionately benefiting lower productivity disabled workers or higher productivity disabled workers. And we think that's actually a contribution of our study that disability might be something separate from productivity. You know, if I'm lower productivity, I can work at it. I can read a book. I can - You mean separate, Separate from tenure and experience? Is that what you meant to say? I, I lost you. Sorry for going too fast. That, that disability might be a different dimension for the AI debate than productivity or tenure. That if we're thinking about, as you said, Steve, whether AI is going to lift up the lower tail or accentuate the upper tail, it might be useful to kind of dive further into these dimensions that in our setting, it's not that it's lifting up higher versus low pro - lower productivity disabled people, but rather people who are more disabled benefit more. And we think this is a - Yeah, That, that, okay. Elaborate on that because you haven't done that yet. Yeah. So there is, aside from just bringing up disabled hearing impaired workers on average in terms of hourly earnings, if I understand correctly, you find bigger positive effects of the AI tool on people who have more severe hearing impairments. Is that correct? Yeah. That's right. It's a an advantageous feature of our data that we observe the exact level of hearing impairment. There are four levels in the Chinese system above moderate. So we have moderate, moderate to severe, severe, and profound. And about half of our sample of DHH workers are in the profound hearing loss. So that's hearing loss of more than 90 decibels can't hear shouted speech that the effects are strongly concentrated among the profound people with profound hearing loss. Yeah. Okay. So that, that's kind of another piece of good news. The people who are most disadvantaged in terms of their hearing capacity are the ones that are having the most positive impact of access to this tool, this AI tool. Yes. Okay. But that's not true across productivity or tenure. It's not like lower productivity. I get your point now. You're trying to say, look, what you're finding is really about the impact of this tool on the earnings and performance of hearing impaired people. It's not somehow confounded by how experienced they are in this job, where exactly they sit in the productivity distribution. You're saying it's, it's basically operating on the, on the impairment dimension. If I want to put it that way. It's kind of an awkward way to put it, but that, that's what I understand you to be saying. Absolutely. Yes. Okay. Now, there was another interesting aspect of your study that I wanted to make sure we bring to the fore, but you can correct me if I'm wrong. You, you pointed out earlier that these hearing impaired workers were, even before the introduction of this AI tool, more profitable on average than workers with no, no impairments. But that became even more true after the introduction of this tool. Now they're even more profitable, right? That's right. In relative terms. So their, their profitability is, was already high relative to those, the profitability of workers with no impairments. And it improved their, the hearing impaired profitability improved even more. So that's the reason I draw attention to that is it's not that this tool makes these workers actually especially desirable for the platform. They were already profitable to the platform. Now they've become more profitable. So the plat - this is a, the platform says, this is great. These, these are the kinds of workers I want. And, and that's quite different in, in than the response to other policy interventions, which are designed to help people with disabilities. And we'll, we'll get to that, but I, I just want to lay. Be down a marker for that at the moment. Are there any other performance outcomes that you want to bring to the attention that we haven't already covered? Daniel, do you want to say anything on that or? Totally. I think, as we mentioned, fat ratings is definitely one of the most important - Fat ratings, yeah. Yeah, and also the speed, but we can also measure late deliveries. Yeah, if late delivery exceeds some expected arrival time of the order. And we also see this introduction of the tool significantly reduce the late delivery rate for the DHH workers. Okay. Okay. Yeah, that makes sense. So one more thing. I don't know if you talked about this in the study. If you did, I didn't see it. So you said at the outset of our conversation, there are two major platforms in the food delivery order delivery business to residential customers in China. So this sounds like a big competitive advantage for the firm that introduced it, the platform that introduced it. Did the other platform respond and say, wait, our socks are getting cleaned? We got to do the same. What happened? Do you know? Yeah. Shortly after our platform introduces it, the other platform introduces it, and we provide a robustness in our paper showing that everything is robust if you exclude the period where only our platform has it. So we don't think our effects are a short-term effect due to just being on that platform as opposed to a used in equilibrium. Yeah. Okay. So this may not be the focus of your study, but there's a really important point about economics here, which is one of the way that innovations diffuse is somebody in the marketplace tries it out for their own, this is Adam Smith in action in some sense, for their own. The company says, "This might help our profitability." So they try it. It works. If it doesn't work, well, then they stop doing it and nobody else responds. But in this case, it worked, worked quite well. Their, their rivals see, you know, we got to do this. It's kind of like Target does something and Walmart's losing customers as a result, Walmart responds. The reason I emphasize this, it's it's such a, it's such a kind of powerful way in which a decentralized markets both create the incentives to try things out to see if they work in terms of a profitability sense. And if they work, they get copied. And, and it's very hard for some kind of centralized authority, government, policymakers, whatever, to create that same kind of dynamic in nearly as effective a way. This is my reading. You, you're free to disagree with that if you, if you don't share that view, but that's my reading. And I think it's a, it's a larger message that I take away from your study and what you just told me about the, the imitation. Still on the imitation point. So does, is this technology spread to delivery workers and platforms that orchestrate delivery services in other countries, or do you know? We, that is on our list of reaching out now that we finished the study. We, we would love to engage more with delivery platforms in other countries. I mean, I, from my personal experience on, not on food delivery, but on ride hailing in the US, you know, they often notify you that your driver is deaf or hard of hearing, but the, I don't believe I ever got an AI call from an Uber. Yeah. Well, Okay. Listen, you, you, you know, there's a famous study in economics decades ago, even, even before I was in the profession by Sve Grilliches on the diffusion of hybrid corn technology in the United States. And I may not do it full justice, but my recollection is one of the central findings in that study is that hybrid corn, hybrid corn technologies were introduced in those parts of the United States where first where it was especially profitable to do so. Yeah. It seems to me you have a, you have an opportunity to do kind of a, a modern version of, of the Grillaces study here. If you can collect enough data, and it doesn't have to be the detailed data you have here, just kind of high level data on the timing of the introduction of this kind of technology into various delivery service businesses where you need interaction between the delivery person and the customer, at least some of the time, and see whether you find, just, just documenting the basic facts about that would be quite interesting. But then do you see the same kind of pattern you did in, in, in the Grillichis study? That's, that's a very interesting comment. Yeah. We, we, we can definitely think more about that. Yeah. So anyway, it's, but the larger point is the one we were talking about earlier, that the way in which markets create incentives to propagate beneficial innovations. I mean, that, that, that is a, a really important point. So let's step back from your particular study now and put it in a broader context of policies that aim to assist people who have some kind of impairment that, that might negatively affect their performance or their opportunities in the labor market. And you, you talk a little bit about this in the introduction to your study. That's kind of the way you set it up to some extent. So you point out that, look, in, in the United States and in other rich countries and, and sometimes in middle income countries, there are many policy interventions that are designed to lead to better, that are at least in intent, not always in effect. In intent, seek to provide better labor market outcomes for disabled workers. So in the United States, for example, I guess the American Disabilities Act says that employers must make reasonable accommodations for disabled workers to be able to complete their tasks. There are various laws that penalize or, or discourage or make unlawful in any event discrimination against disabled workers under certain circumstances. There are sometimes, some quote, in some countries, there are quotas. You must hire a certain fraction. A certain fraction of your workforce must be disabled. So there are many of these studies and, and many of these policies in effect, there are many studies of their effects. We're not going to dive into the details there, but I, but there is one important difference between your setting and the effects of many of these studies that's worth pointing out, which is in many of these other studies, they kind of make they don't do anything to make disabled workers more attractive to the employer. Yes. In fact, They often do just the opposite. So you must make reasonable accommodations for disabled workers in the workplace. That may be a perfectly legitimate societal goal, but making reasonable accommodations all typically involves some kind of costs, you know, including just making things accessible to people who are mobility impaired, but also to provide special tools for people who are hearing impaired. So it's kind of like you're kind of pr - kind of pressing down from on high through policies or other interventions. You're kind of trying to push employers to make the work environment more accommodating for people with disabilities. Of course, then employers have incentives to try to figure out ways to circumvent the policies and so on. So, so the intent doesn't always get realized in terms of its actual impact. One beauty of your, of your particular technology advance that you study is you don't need any of that. Nobody needs to push the employers. As we discussed before, the employers themselves say, "Wow, this is good for me. I want to do it. Yeah, it's good for my workers. It's good for the hearing impaired workers, but it's good for my profitability. So I want to do it." So there, there's a big distinction to be drawn there between kind of top-down policies that try to impose these desirable outcomes and innovations like in your setting, where once the innovation becomes available, if it works in practice, it naturally spreads through the market mechanisms in ways that benefit the employers and the workers. I just want to give you an opportunity to comment on that broad, broad perspective. So I guess from a policymaker perspective, what you want to do is you want to create the conditions as much as possible where this kind of technology, when it's available, can be deployed widely, because that will achieve the kinds of objectives that policymakers often have in mind anyway with respect to workers who have various impairments. Absolutely. I mean, I'll, I'll, I'll let Yanio chime in here too on to, to, to give his, his view. But I, I agree. In our, in our particular context, it was privately optimal for the platforms to introduce the AI tool, at least our platform. We, we, we estimate that it was highly pri- privately optimal so that the, the optimal policy in our setting might be for a government to do relatively little. And for, as you point out, Steve, that we, we study a disability, which is very, very common. One reason why it's so profitable for the platform is because there are many workers with hearing impairment. If we were studying a rarer type of disability, that might be a context where there's a wedge between the private benefit to the employer and the social benefit from having less usage of disability insurance and just kind of general benefit from having disabled people working. So for rarer disabilities akin to, you know, rarer diseases more, more, more broadly, people making arguments about pharmaceuticals, that might be the context where government intervention might be more warranted. So I take that point. What I'm more worried about is the kind of almost the mirror image of that, which is because of these generalized anxieties about what AI might mean in the labor market in other domains, there's high potential for political backlash that says, "Let's just slow down this whole process. We want to, we want to put a client, we want to study every possible AI innovation and see if it has any potential for harm before we let it out in the marketplace." The point I'm trying to drive home now is that policy stance inhibits the creation and diffusion of the type of AI advances that you're talking about now. These things, these things don't happen in a vacuum, these kind of innovations. Somebody develops this tool because the, the tool maker, we didn't talk much about the, the firm or the firms that developed the tool, but they're also making profits here. They, they have, they do, they're providing something which is beneficial to the platform and to the hearing impaired workers, but why are they doing it? Because they, they, they themselves want to make a profit. And if the government comes along and says, "You know what? We have to, we have to really carefully evaluate every possible AI innovation before we let it roll out in the, in the market," that undermines the commercial incentive to advance AI technology. That's the point I'm making, which it's different than the point you are making. And I'm more worried about I recognize the point you're making. I'm more worried about the one I'm making, that we're actually going to slow this whole potential wave of innovation and its diffusion in the, in the marketplace because of policies that are, that are basically driven by fear of what AI might do. Yes. No, I mean, I, I, I, I agree with that. And that I think one contribution or something that we learned while our study is that this is an, that AI even, that, that the effects of AI may be complex. That there can be effects on job loss. There can be many effects, but one, one benefit might be this benefit to workers with disabilities, which we haven't seen very much in the conversation about AI. Yanyo, what do you think? Sorry for, for monopolizing conversation. No, no, not at all. Yeah, I, I, I think yeah, I agree too. And I do think this cost of implementation or innovation is important here because we are talking about a tech firm, a food delivery platform who already have this technology or advantage of developing new AI tools, right? So when we think about maybe for other types of jobs with people with disability, maybe they need the tool, but they don't have the technology or they don't have the advantage of developing such tools. Yeah, so I think the opportunity for them to roll out those technology is indeed important, not only for people in this particular industry, but maybe people with disability, but in other industries as well. Yeah, that's right. Okay. Thanks so much for the conversation and for telling us about your study. If you do try to figure out what's happening to this kind of AI tool and its introduction more broadly around the world, I'd like to hear about that. So send me that study when it comes along. And any other studies in this space. Okay. Thank you so much for having us, and we'll definitely keep you posted.