We discuss “Voice in AI Firms: A Natural Field Experiment on Automated Job Interviews” by Brian Jabarian and Luca Henkel. The paper examines a randomized experiment with call center job applicants in the Philippines who were assigned to either AI-conducted voice interviews, human interviews, or given a choice between the two.
Key Findings:
* AI interviews led to higher job offer rates and proportionally higher retention rates
* No significant difference in involuntary terminations between groups
* Applicants actually preferred AI interviews—likely due to scheduling flexibility and immediate availability
* AI interviewers kept conversations more on-script with more substantive exchanges
* Online applicants saw especially large gains from AI interviews
Topics Discussed:
* The costs of recruitment and why interview efficiency matters
* Whether AI interviews find different workers or just reduce noise in screening
* How human recruiters interpret AI interview transcripts differently
* The “Coasean singularity” question: Will AI improve labor market matching overall?
* Limitations: scheduling confounds, external validity beyond call centers, unmeasured long-tail outcomes
* The coming arms race between AI interviewers and AI-coached applicants
Posterior Updates:
On the usefulness of current AI for job hiring:
* Seth: 40% → 90% confidence AI works for call center jobs; modest update for general jobs
* Andrey: 20% → 75% for call centers; 1% → 5% for general interviews (“we need to reorganize all of hiring first”)
On whether AI will improve job matching significantly on net in the next 5-10 years
* Andrey: 55% → No Update
* Seth: “A bit more optimistic than Andrey” → +1pp update
Referenced Work/Authors:
* Related episode on AI and labor signaling with Bo Cowgill.
Transcript:
[00:00:00] INTRODUCTION
Seth: Welcome to the Justified Posteriors podcast, the podcast that updates its priors about the economics of AI and technology. I’m Seth Benzell, an interviewer who will never stick to a standard script, coming to you from Chapman University in sunny Southern California.
Andrey: And I’m Andrey Fradkin, counting down the days until I can use an AI to pre-interview my podcast guests to see if they deserve to be on the show. Coming to you from San Francisco, California.
Seth: I don’t know. I think our filtering criteria is pretty good.
Andrey: I know.
Seth: Right. That’s one job we never want to automate—who becomes a friend of the podcast. That’s an un-automatable job.
Andrey: But it would be nice to pre-interview our guests so that we could prepare better for the actual show.
Seth: I was thinking about this, because there’s two possibilities, right? You do the pre-interview, and you get an unsurprising answer in this sort of pre-interview, and then that’s good, and then you should go with it. And then if you get a surprising one, then you would lean into it. What would you even get out of the pre-interview?
Andrey: Maybe what the guests would want to talk about.
Seth: Okay.
Andrey: But I agree with you. Mostly, it’s just hearing the guest talk, and then thinking about, “Oh, this is something that we want to really dig into,” versus, “This is something that might be not as interesting to our audience,” and knowing that ex ante.
[00:02:00] SETTING UP THE TOPIC
Seth: Yeah. We’ve been... So we’re talking about interviews. You’ll remember in a recent episode, we just talked to our friend Bo, who’s doing work on how maybe job applications are changing because of AI. So now I think what we want to think a little bit about is how job interviews are changing because of AI. Maybe we’ve heard before about how AI is changing how people talk to the hirer. Maybe we want to hear a little bit about how AI is changing how the hirer solicits information in an interview. We’ve got a very interesting paper to talk about just about that. But do you remember the last job interview you did, Andrey?
Andrey: Yes.
Seth: How did it go? Did you have fun? Did you feel like you stayed on topic?
Andrey: It was a very intense set of interviews that required me to fly halfway across the world, which was fun, but exhausting.
Seth: So fun. So you would describe the interview as a fun experience? Did you get more excited about the job after doing the interview?
Andrey: Yes, although I ultimately didn’t take it, but I did get—you know, I was impressed by the signaling value of having such an interview.
Seth: So the signaling value. So in other words, the signal to you from the interviewer about the fact that they were going to invest this much time. Is that right? It’s that direction of signal?
Andrey: Yes, yes. And also the sorts of people who they had talking to me, and just the fact that they were trying to pitch me so hard. Now, certain other companies lacked such efforts.
Seth: Right. So it seems like one important aspect of an interview is what the interviewee learns from the interview. But what about the other side? Do you feel like your interviewer learned a lot about you, or enough to justify all that time and expense?
Andrey: I’d like to think so. I mean, I’m not them, so I can’t really speak on their behalf. But it did seem like the interview process was fairly thought out for a certain set of goals, which might differ across companies. What about yourself, Seth?
Seth: Thank God, it has been a long time ago that I interviewed for a job, and I can tell you exactly what happened. I was on the academic job market, but I did throw out a couple of business applications, and so I got an interview at Facebook. Headed out to their headquarters, did all of the one-on-one interviews, and then there was a code screen, and I was not grinding LeetCode for the last five months and completely bombed it. And they said, “Thank you very much for your time.” So that was an example of, I think they probably could have saved the time for the interview if they had given me the code screen first.
Andrey: It’s funny, there was a time in my life where I interviewed at Facebook, too. I mean, this is probably 2014 or something.
Seth: Mm-hmm, mm-hmm.
Andrey: And they did do the coding screen before.
Seth: Who knows? Who knows, dude?
[00:05:15] THE PAPER
Seth: Okay, so interviews, we do them. People seem to give information, take information from them. How can this be made more efficient with AI? That’s today’s question. In order to learn more about that, we read Voice in AI Firms: A Natural Field Experiment on Automated Job Interviews, by friend of the show, Brian Jabrian and Luca Henkel. I was interested in this paper because it’s kind of an interesting flip side of what we just saw from Bo.
I guess before we talk too much about what the paper actually does, it’s time for us to go into our priors.
═══════════════════════════════════════════════════════════════════
[00:06:00] PRIORS
Seth: Okay, so Andrey, when we’re thinking about AI being used in interviews, what sort of thoughts do you have about that going in? What sort of priors should we be exchanging?
Andrey: Yeah, I mean, I think just when I first saw this paper, I was kind of surprised that we were there already, honestly. I think interviewing via voice is a pretty delicate thing, and the fact that AI is potentially able to do it already was—I hadn’t been thinking—I didn’t think we were there yet, and I think just the very existence of this paper was a bit of a surprise when I first saw it.
But I guess a first natural prior that we can think about is: is using an AI to interview someone rather than using a human to interview someone, is that better or worse, or how do we think about that?
So, Seth, what do you think?
Seth: Well, it’s a big question, Andrey. I guess my first response is, like we always say in this podcast, context matters, partial equilibrium versus general equilibrium matters. The context that we’re going to be looking at in the paper is call center workers. So maybe I’ll give kind of a different answer for short-term call center workers than maybe longer term economy as a whole.
When I think about call center workers, I think about a job that seems to be—no offense to our friends of the show out there who are call center workers—but this does seem like one of the jobs that is going to be the first to be automated with generative AI, or most at risk, especially kind of low-skilled call center work. So if there was going to be any sort of domain where you could automatically verify whether someone was good at it, intuitively, it would be the domain that you’re kind of close to automating anyway. So if it was going to work anywhere, I would say it would work here.
And yet still, call center work, you might imagine, it requires a lot of personal empathy, it requires maybe some subtleties of voice and accent that an AI might not identify or even might hesitate to point out such deficits. I would say I kind of went in with the idea that for call center workers, maybe there’s a forty percent chance that AI would be better than a human interviewer. So maybe it’s slightly unlikely that it would be better. But if we were to expand out to kind of knowledge work as a whole, I would be more, even more pessimistic, maybe only a twenty-five percent chance or lower that the AI interviewer would be better. What do you think?
Andrey: Well, how would you—what do you mean by better?
Seth: Oh, well, better in terms of the hire is ultimately the correct match, right? That’s going to be operationalized in a specific way in this paper, what... How they’re going to measure better match, but, yeah, that’s what I would say. They hire someone who’s going to be productive and work with the firm for a long time.
Andrey: Yeah. I mean, so that’s kind of one definition, I guess. Another definition might be, is the ROI from a particular interview process better or not?
Seth: Right, better net of costs. Right. Okay.
Andrey: Because I think one of the things that oftentimes economists underappreciate is that recruitment is an enormous cost.
Seth: Don’t tell those search labor economists, dude.
Andrey: Some of them model it, but I don’t think it’s actually a big focus. But it’s just the process of interviewing. You know, let’s say there’s a position, and you need to interview six people for a relatively high position, so that’s six hours direct, or maybe it’s a half-hour interview, it’s not obvious. But then also, there are all the meetings and pre-meetings, post meetings. Maybe you give an offer, and then they don’t accept it. And there... I mean, there’s just a lot of costs involved. So even if it wasn’t as good as a preexisting interview process, it might still be ROI positive for the firm.
Seth: I guess we come back to what is the cost of interviewing versus the cost of making a bad decision. You know, well, it’s not, it’s public information that we, here at my university, we hired a dean of the business school who was an absolute disaster and got voted out by the faculty in a ninety-eight percent vote after one year. That guy did a lot of damage, right? We should have interviewed him harder.
So it really depends. So I guess the point would be in kind of higher leverage roles, you would think that the interview costs would be a relatively negligible part of what’s going on.
Andrey: I don’t think that’s true. I think in higher leverage roles, higher leverage people have to do the interviewing, and the cost of delaying hiring is much higher. So to me, it’s not obvious. But anyway, that’s, this is all a sidebar.
Seth: Okay, so let me hear the prior.
Andrey: Yeah. So I think my prior that this interview technology would be better than a human technology, just solely based on match quality, was actually quite low. I probably twenty percent, or maybe less than that, actually. Because it just seems like, yeah, maybe on average or maybe in a typical case, it’s fine, but there’s so many things that can happen in an interview that you could only learn by running a process enough times to really learn how to do it well. And so, yeah, I wasn’t super optimistic that it was going to work yet, even for call center workers.
But I think for kind of higher-end labor, right, I think my prior that it would be better is very low, you know, like 1%. Just because I just don’t think we’re there yet.
Seth: Wait, so I’m getting—So 20% for call center workers and 1% generally, was the take?
Andrey: Yeah, that would be my sense.
Seth: Mm-hmm.
Andrey: I mean, just, it’s hard to imagine that at today’s technology levels, that for, let’s say, a professor job, that the AI could interview better... I guess one way to put it is getting rid of all the humans in the interview loop for a faculty hire, that seems just kind of crazy.
Seth: Right, and that... Well, obviously, a more extreme experiment than what we’re talking about here. Faculty, we’re thinking about, you know, maybe they’re pushing frontier knowledge, would be the last thing that you would think that an AI would be able to get at. Another thing I think about is someone who’s going to be in your faculty is living with you for 20 years, so you might really care about if they smell good, if they have a peccadillo that bothers you, that these might not be relevant considerations in a call center remote job, right?
Andrey: Yeah. Yeah, exactly. I think... And I think, actually, the interpersonal thing, which is a very contentious thing, by the way, is that I think people understand that good teams get along with each other. But at the same time, screening based on how much you’d like to have a beer with someone might have problems, you know?
Seth: Not good.
Andrey: So yeah. So, you know, it’s not obvious which way that cuts, but certainly it’s an important part of hiring. And, you know, I think for higher-paying jobs, it’s not that there’s just one interview, of course. There are many, many interviews, and oftentimes, in-person components of interviews over dinner, and so on. And you might think, you know, maybe that’s all unnecessary, but given that it persists in equilibrium, even though it’d be a lot cheaper not to do it, that should signal something.
[00:14:00] GENERAL EQUILIBRIUM CONSIDERATIONS
Seth: Good point. But now, Andrey, what I’d like us to think about for a second is to maybe zoom out for a bit and think about, okay, we’re talking about current generation technology in partial equilibrium in this study. One company uses 2025 generative AI to try to attack this specific question for call center workers. Let’s take a step back. You know, that’s what we always want to do in this podcast, is take a step back and like, okay, what does this tell us about the broader process that society is undergoing?
You’ve written recently, movingly, to be honest, about this idea of a Coasean singularity, that AI will be so good at helping us communicate to each other, that we’ll get perfect matching at zero cost. I don’t know what timeframe you have in mind, but presumably, one of the things we’ll get better at matching is people to jobs. So maybe you’re pessimistic that in this context, in this time, that AI will be good at hiring, but do you think, you know, 5, 10 years from now, as these technologies diffuse, do you think we’ll get better job matching as a result of employers using a lot of AI and job applicants using a lot of AI? Is that final equilibrium the destruction of all meaning, as Bo, you know, foretold, or is it the utopia of the Coasean singularity?
Andrey: Well, I do want to point out that I don’t think any of the authors strongly believe that the Coasean singularity will happen, actually, you know?
Seth: Oh, the Coasean singularity is a myth?
Andrey: The Coasean singularity, question mark, Seth. Question mark.
Seth: Question mark’s doing a lot of work, Andrey.
Andrey: Yeah. No, the paper is doing a lot of work to tell you why it might not happen.
But I think, yeah, I think time horizon certainly matters here, right?
Seth: Okay, but let’s say 5 to 10, to just to choose a number.
Andrey: Yeah. So, so, like, not that long a time horizon. It’s very non-obvious to me. Just because there are all sorts of institutions that are going to be involved, very messy institutions. Like, one of the things that we already talked a lot about on this show is the problem of too many applications, applications lacking signaling value. At the same time, you know, you can imagine on the interview side, if you interview, you know... How does this all affect the number of interviews you’re going to do?
Seth: There’ll be more and more applications. The cost of applications goes down, yeah.
Andrey: Yes. Now, maybe the cost of interviewing goes down, but it doesn’t for the applicant if they have to be the one... You know, if the applicant’s agent is doing the interviewing, maybe it’s a different story. But if the—
Seth: Right! How many, how... It’s like, it feels like you’re watching, you know, the drone war in Ukraine. There’s the move, and the countermove, and the countermove, and the countermove. It’s hard to say where that process ends, right?
Andrey: Yeah. So I... And then I think, of course, you know, there are actual individual institutions involved. Like, what is the government going to do? And even if some nimble firms are really doing a great job of matching using AI technologies, how that plays out when there are other organizations that are using other sorts of tools, it’s just completely not obvious to me over a five to 10-year time period.
Seth: So is that a fifty-fifty? Is that a, I have—is my prior is the completely uninformed prior?
Andrey: No, no. I think because you’re introducing both sides of the technologies, both the AI for the applicants and for the employers, it’s hard. I mean, I’m a bit of an optimist, so maybe I’ll say fifty-five percent chance.
Seth: Fifty-five percent. Ooh, I have to say, I’m a little bit more optimistic than you, Andrey. I think if you think about the world, the world, since, you know, the rise of the printing press, has seen an arms race in technologies for understanding versus technologies for lying, right? And yet, we think kind of the general process has been towards better price discovery, better matching, right? It seems like we could translate the same ideas to financial markets, where people are getting better at lying, people are getting better at trading, people are getting better at communicating. But ultimately, I mean, at least my sense is that price discovery has improved, right? So I guess—
Andrey: Oh, I would argue the opposite. So I... Not price discovery, but labor discovery, I think has been substantively hurt over the past five to ten years. Because our educational institutions have abdicated their role—
Seth: Credentialing.
Andrey: Actually, credentialing, and because it’s been trivial to start applying to jobs. So yeah, I mean, look, that’s a little too pessimistic, but I’m just saying that over a five- to ten-year period, I have to be a little bit cautious. I think if we’re to be able to reoptimize our institutions, I mean, now the problem with going thirty years is how much human labor do we even have? But to me, just lots of things could be going on.
═══════════════════════════════════════════════════════════════════
[00:22:00] THE EVIDENCE - CONTEXT
Seth: Okay, all right. So we’ve got our priors locked in. Now it’s time to turn to the evidence.
Okay, so our context here is the Philippines in 2025. We’ve got a pool of about seventy thousand applicants to different call center jobs. They’re all going through this one recruiter who’s recruiting for multiple different businesses. To give some context about the call center job market, this is a very high-turnover, low-paid work. We’re talking about three or four hundred dollars a month at two to three times minimum wage. The skills required are English speaking, flexibility with changing shifts. There is a line in the job application that calls for strong analytical and logical thinking. I think strong might not be the correct adjective there. You probably need more than zero.
But all this combines into a job that people are not married to. So we’re looking at a job with sixty percent annual turnover, with a high share of that being people voluntarily leaving rather than being fired. The... We’re talking, in order to do these interviews, people, first, they can either show up in person to one of these recruiting offices, or they can apply online. Then they’re scheduled for an interview, and they also take a standardized test that has both an English skills component and a kind of analytical mathy component. And just to give a sense of how strong a filter this is, about six in—if we’re talking about the human interview baseline, about six percent of applicants accept a job, while two percent still have a job one hundred and twenty days after being hired. So that’s not a conditional average. That’s just two percent of people who show up for an interview end up having the job for at least four months. So that’s our context.
Andrey: And about ten percent get an offer, approximately.
Seth: Right. Yeah, yeah, so ten percent get an offer, six percent accept the job. Okay. So that’s the context. Andrey, do you want to tell us about the experiment?
[00:22:40] THE EXPERIMENT
Andrey: Yeah, sure. So in the experiment, workers were, or applicants... Well, first they were pre-screened a little bit—
Seth: Very lightly.
Andrey: Yes, and then they were assigned to either a group where they had an AI interviewer, whether they had a human interviewer, or one in which they got to pick. And I guess there’s a lot to be said about the specifics of that interviewer process. So there, as you can imagine, for a job where so many people are being hired, there’s a lot of standardization of, you know, what sorts of things need to be discussed, in what order. And the AI tries to... You know, the AI tool that the company has purchased is going to is programmed to do that, and it tries to do that. Another key important part of the context is scheduling.
So an AI can take the interview at any time with you, which could be just right away, as soon as you pass the pre-screener, whereas a human needs to be assigned to an interview, and that could take some amount of time. So that’s also a pretty big potential difference in how we should think about these things, right? So we oftentimes focus, oh, can the AI really do it? But actually, AI has this other advantage where it could just do it right away.
Seth: Although, it is, it’s an interesting result. Even though the AI conducts the interview faster, it still takes longer for the AI interviewed to actually get the job offer decision, which seems to be driven by the humans. And now we’re going to get into the details of how does this AI system work? There is a human who listens to the AI interview, right? And apparently, I get the impression that the humans who listen to the AI interviews do not enjoy it. They would rather listen to themselves, right? They score these a lot faster if it’s their own interview versus the AI interview.
Andrey: So did they really do a good job of explaining why that happens in the paper? Or maybe—
Seth: Well, that’s my speculation.
Andrey: That’s actually not what my speculation is at all.
Seth: Okay. Oh, let me hear it.
Andrey: So you’re portraying it like, you know, they’re just taking a long time to listen. Like, they, you know, to listen through the interview. But actually, it seems like a procedural thing. Like just the system, when it assigns them to review these applications, you know, is later than if you already did the interview.
Seth: Presumably, you score it right there.
Andrey: Yes. Yeah, yeah. And to be clear, my understanding is that there’s a different person, which is the recruiter, who’s doing the scoring, than the person who’s doing the human versus the machine interview. So it’s not like they’re either listening to the machine or listening to the human and then finding the machine less interesting to listen to. It’s actually just procedural that they’re getting assigned to read this AI interview result later.
Seth: So maybe not an essential difference, but one that could be corrected with a little refinement here.
Andrey: Yes, exactly. Yeah, yeah.
Seth: Mm-hmm.
Andrey: I know we got into kind of this side bit, but I don’t think it’s a side bit because it’s always important to think about what is the treatment exactly. And one of the threats to internal validity that I always teach my students is that if multiple things are changing at the same time when the treatment gets assigned, and in this case, there are. You know, you’re getting the AI interview, but you’re also getting interviewed way faster initially. So from the applicant’s point of view, that’s kind of very salient.
Seth: It’s sort of a different experience.
Andrey: Yeah.
Seth: Which, you know, like we talked about, the interviewee also learns from the interview, right? It’s like when the professor says, “I learn far more from my students than they learn from me.”
Andrey: Yeah. Well, I don’t think this is a learning—I mean, it’s not like I’m going to rule out learning by these workers. But my sense is that there’s not a lot of uncertainty about this job for the people who are—
Seth: These jobs are pretty homogenous.
Andrey: They’re pretty homogeneous—well, you know, they’re at least... You know the distribution, you know, probably, you know, doesn’t have too much to do with the specific firm. You know, they’re—probably, the call centers jobs are, you know, there, there are just a lot of them, and depends on which, who you get assigned to in terms of your client.
Seth: I think this is an important point, which is that it really does seem like there’s more vertical differentiation here than horizontal differentiation. You might imagine a context with more horizontal differentiation, the AI interviews might not be as good. But here, we’re just trying to find the right tier of worker, because if it hasn’t become clear yet, the main failure mode isn’t you hire someone who’s too bad. The failure mode is you hire someone who’s too good, and they leave the job after a week.
Andrey: Well, we don’t—So to be clear, I don’t actually know why people leave their job. You’re assuming that they’re too good, but actually that to me is completely not obvious. It’s like an Uber driver. It’s not like the Uber driver is too good if they stop driving on Uber. It’s just maybe they needed money for a couple of weeks.
Seth: Well, their distribution of opportunity cost is higher, which would be correlated with being good.
Andrey: Yeah, but it might also just be they just had temporary liquidity... To be clear, what I’m trying to say is that that correlation, in my opinion, is very likely to be low. The fact that these people apply to this job, which is very fungible in the first place, which so many people in their country apply for, is not suggesting to me that these applicants are somehow, have all these amazing other opportunities. And, you know, they’re probably call center workers that might be cycling between call centers, or maybe they’re cycling between call centers and other seasonal work. I mean, I don’t know. I just wouldn’t assume it’s about quality. Yeah. It’s not like “Oh, wow! They’re so good at math, and then they got discovered.” You know, that’s kind of not the story here.
Seth: Okay, but we’ll come back to whether who seems to be helped by or hurt by the AI worker in a second. I guess one last thing I want to say about the experiment and its context before we go into the results, are that they... We also get a survey of people on their interview experience. So you might imagine that they’re going to be obsequious or sycophantic, to use a word in vogue these days, because, you know, they’re trying to get a job, but that just gives us another slice at trying to understand what they’re thinking.
Andrey: Yep.
Seth: Okay—
Andrey: So yeah, I mean, I guess we should say, because we haven’t made this clear yet, this is an absurdly impressive experiment. I mean, holy crap!
Seth: Yes.
Andrey: Right? Just logistically, it’s... You know, I can imagine how difficult it would be to get all this machinery rolling and, you know, figure out the pilot studies, and figure out the AI model provider, and convince the firm to do it this way versus a variety of other ways. You know, I think it’s notable that certainly, the firm should be interested in the results of the experiment. They’re—It’s probably an active, like many other firms, they’re actively deciding where to use AI tools, and so it is incentive aligned in that way. But still, it just is a very impressive experiment.
Seth: Yes, huge snaps to the authors, especially Brian, who I understand is on the market right now. Give the man a job.
[00:31:00] HEADLINE RESULTS
Seth: So all right. To get into the headline results, the AI interviews seem to work. We get twelve percent more offers. So of the people who are randomized into the AI group versus the human group, the AI interviewed get twelve percent more offers, have eighteen percent more job starts, and have eighteen percent higher chance of working with the company for at least four months. So our main outcome here is retention and hiring as positive outcomes. Maybe in the limitation section, we’ll talk about kind of the limitations of those as the endpoints, but, you know, retention seems to be one of the big challenges here, given that it’s kind of, as you said, very fungible work. And those seem like significant results, plus on top of all the cost savings you previously talked about.
Andrey: Yeah, yeah. I mean, it’s definitely... You know, the ROI calculation, of course, needs to account for other things, but just the baseline results do suggest that this is a very useful technology.
Yeah, what do I make of this? I think it’s interesting to think about where this effect is coming from. Is it coming from different types of workers being screened by the two methods, or is it just that the AI method just picks off a few marginal workers that happen to stay longer?
Seth: Be bad at interviewing, right?
Andrey: Yeah, or bad at interviewing, or they, you know, they’re actually good enough, but the old interview process was a bit too noisy to pick them out, right? So there’s kind of this question: What’s going on? Because what I would’ve thought that, you know, like if I was a company, and I was thinking about, well, what is the interview technology that I want? I want an interview technology that gives me the same decisions as I was making before but with a lot less cost.
Seth: Mm-hmm. Right.
Andrey: The fact that this technology instead increases the hire rates. First of all, in a lot of jobs, like for a lot of jobs, there’s one slot, so this couldn’t be a result that was replicable, right? Like, if you’re hiring a professor, and you have one slot, it’s not like you’re going to increase... I mean, you can increase your hire rate from zero to one, but it’s kind of... It—
Seth: But retention then.
Andrey: You have to really... Yeah, but those are different—But you have to think about why you’re getting the retention effect, right?
Seth: Right.
Andrey: And so there are kind of different things that we can think about here. Is it that the interview process is less noisy? Is it that the interview process is more lenient, that it’s getting marginal guys? Or is it that actually, it’s actually picking out different people, and those people are better matched, which then raises the question of like, wow, those old interviewers were not very good, right?
Seth: Right.
Andrey: Which is, you know, I’m sure there are plenty of interviewers who are not good. That’s—It’s not surprising to me. Yeah, but I guess, yeah, those are the questions that are raised, right? Because I don’t think it’s inherent. How you use the AI tool is your choice as a firm. There’s no law that’s going to say that you’re going to increase your hire rates because you happen to use an AI interviewer, right?
Seth: Right. And so, yes, a great point is you might be concerned that this leads to a more sort of lenient, we’re letting in marginal people. You know, we’re not actually getting more information. Or maybe we’re getting less information, and we’re just letting in marginal people. One piece of evidence against that is there is no significant difference in the rate of involuntary disconnections, right? So remember, retention is higher, and that is not driven by any difference in the newly hired being less likely to be fired, right? The people who are hired by AI, the reason they are retained for a little bit longer is because they are basically fired at the same rate, but they’re less likely to disconnect on their own a little bit. That’s my read.
So how do you interpret that?
Andrey: I guess it still isn’t telling me that whether we’re picking... I mean, for what it’s worth, I just—My reading of the evidence from this paper is that there’s just a lot of overlap in who gets hired, and then there’s just a few marginal guys, and then your power to detect differences and fire rates between the two are very low. But I don’t think the firm—I’d assume that the firm doesn’t care that, you know, there’s so many workers falling through, you know, that involuntary separations are just part of the game. But I wouldn’t... It seems like the power for that difference seems very low.
Seth: Fair enough. And further, and we can talk about this in limitations, too, retention rate just gives you a sense of what percentage of people are above or below some sort of line of so disastrous you get fired. You might imagine that an AI interviewer has a lower chance of detecting the truly disastrous person who’s just going to start slamming racial epithets at everyone who calls up, right? You might imagine that there’s kind of a long tail of badness that’s not being picked up by AI, and then this measure of outcome wouldn’t pick up that the long tail of badness is getting worse.
[00:36:35] MECHANISM - HOW THE AI WORKS
Andrey: Yeah, yeah. I mean, and to be clear, I don’t want to highlight that. I’m just making the point that there’s no generic—I like to think about the prediction machines framework here maybe.
Seth: Friend of the show, Avi Goldfarb.
Andrey: And Ajay and Joshua Gantz, yes. So the AI makes a prediction, but then you’re the decision maker. Let’s say you’re the CEO or the hiring manager of this firm. You get to choose how you use that information, right? So you can use it—
Seth: But it’s not that the AI isn’t... Wait, wait, wait, wait. The AI isn’t making a prediction here. The AI is soliciting different information in the interview.
Andrey: Sure, but it’s giving you a signal. And you can choose what to do with that signal however you like, right? So that’s kind of the point I’m making. In this case, the AI was good enough at interviewing people that you got a pretty good signal, and the system used it in the following way that seemed to have been positive. But I guess what I’m saying is how you—there are human recruiters that are taking the signal from the AI interview and choosing what to do with it. And they chose to hire more people as a result. That’s not a quality of the AI, that’s a quality of the humans making decisions off of information.
Seth: I mean, I don’t know what to say to that, Andrey. Like, you know, it’s like saying, you know, the factory didn’t make 10 tons of steel. It was the business factory sociotechnological system that made 10 tons of steel.
Andrey: No, I guess the point I’m making is that you could have imagined, here’s a simple story. Let’s say the interviewers don’t know how to interpret the AI interviews, and they do know how to interpret the human interviews. Then they could make very different decisions off of very similar transcripts off of the two.
Seth: Correct.
Andrey: Right? That, I guess that’s what I’m trying to say.
Seth: And I think that’s right. I think that’s right, but I’m also pointing out that we usually don’t talk about technologies that way. Every technology is embedded in an organization. So yes, but yes, every other technology also.
Andrey: No, because when people do AI evaluations, they’re always saying that AI does this, AI does that. And then in this case—
Seth: Like GDPVal.
Andrey: Yes, yes. AI is going to fully automate end-to-end this task. And I guess what I’m saying here is that there’s no way it’s automating the decision. It’s not automating the decision. I guess the other thing is there are AIs that automate decisions in hiring, right? There are certainly AIs that screen resumes, for example. So I don’t think it’s a crazy thing to talk about here.
Seth: I don’t think you’re being crazy either. And of course, the context matters, but then even in GDPVal, I could say the same thing, right? It’s going to get evaluated by a human expert. The human expert either is good or bad at understanding the way that the AI talks about the thing. I mean, it seems like any time a human touches it, okay, yeah, it’s in a human context.
Andrey: I guess... Sorry, but you keep on thinking that this is a criticism. It’s not a criticism that I’m—You don’t need to defend it. It’s just I’m just saying that—
Seth: I’m not saying it’s a criticism.
Andrey: Yeah.
Seth: I’m saying it’s a universal... I’m saying it’s a truism.
Andrey: It’s just the company chooses what to do with this.
Seth: True.
Andrey: It’s interesting that the way that it was used happened to play out this way. But for example, the company might not have wanted to hire them, right? Like, what is the hiring cap for the company? Do they want to hire infinite workers? Do they want to hire 50 workers? How does that allocate the—
Seth: Do they care more about average quality or average retention? I totally agree. Totally agree. Okay, so I don’t think we’re disagreeing.
[00:41:00] LINGUISTIC ANALYSIS
Seth: All right, but let me try to help you a little bit, Andrey, with thinking about what’s happening different in these interviews. Because maybe we can’t exactly say how are the people who get hired different under the two regimes, but we can say something about how the two different interviews go. And so the authors do this really fascinating linguistic analysis of what actually happens in the interviews, because they’ve got the full text of all of these interviews.
Andrey: Actually, can you show figure 2 first, actually?
Seth: Ooh, let’s talk about figure 2 for a second. All right, I’m putting figure 2 on the board. Is that good?
Andrey: So I think I found this very helpful to address some of the questions about... that I was raising. In particular, what we see here is on the top line, the human topic coverage, and on the bottom line, the AI topic coverage. And the AI does seem to cover more topics most of the time than the human. In the second column, we see that the AI tends to follow the preordained order of the interview that was, you know, the interview designers designed. And in the third column, we see that the AI follows the guideline questions much more closely. So it’s standardizing the interview process. So my sense is that this should reduce the noise in the hiring decisions quite a bit. You know, at least in a very naive model of hiring. Now, you can come up with scenarios where there’s—
Seth: Yeah, in a naive model where the generic approach is the correct approach, right?
Andrey: Yes, yeah.
Seth: Because you might have a model—
Andrey: If you need to cater to different people, how you interview, because you’re really trying to extract a particular signal, then maybe this won’t work. But then we go back to the fact that these are call center workers, and maybe there’s more of a—it’s a more standard situation.
Seth: Agreed. Okay, but I, you know, even though this is an interesting figure, the figure that really struck me is the next one, where we look at, okay, what are the things in interviews that are predictive or not predictive of the interview leading to a hire? And then how often do those appear in the AI versus the human interviews? And so what are the bad things that happen in human interviews that don’t happen in the AI interviews? Well, first, I love this one: back-channel cue frequency. Now, I’m not a hundred percent clear on what this means, but the implication is it’s people trying to give a kickback to the interviewer or saying, “Hey, I know your cousin, give me an interview.” Did you get a sense of exactly what this is?
Andrey: Yeah. I don’t quite know how to interpret it.
Seth: Well... I mean, that is kind of interesting and funny and kind of reflective—
Andrey: Short cues indicating attention or agreement. So I don’t think that’s exactly what we’re talking about.
Seth: Short cues, agreement—so they’re just saying, “Yes, yes?”
Andrey: Yes.
Seth: “Hmm.”
Andrey: Hmm.
Seth: Hmm.
Andrey: Hmm.
Seth: That’s less exciting than what I thought that meant. Okay, well, how about this one? We talked... And I think this is really illustrative here of how you might not be able to extend this result out of context. What is bad for an interviewer? Asking a lot of questions about the job, right? Like we said, Andrey, in the kind of jobs you apply for, they’re trying to get you, right? The interview is just as much about what you learn about them. That is not the kind of job we’re talking about here. Any time you’re spending saying, “So you’re telling me this call center worker doesn’t have any benefits?” You’re signaling to them that, you know, you’re going to be a little bit light-footed, wouldn’t you say that, Andrey?
Andrey: Yeah, I mean, it’s a standard job, you know, not... I presume that most people applying for it know how it works.
Seth: “Will I be required to talk to people on the phone in this job?” That’s a bad signal if you say that.
On the other hand, what happens more in the AI interviews? Well, the one thing that happens significantly more of are exchanges. So like you showed us before, you get through more of the standard questionnaire in the AI interview, which makes sense if the AI is good at sticking to the script, which, as I clarified in my intro joke, I think I would be bad at. So that tells us a little bit about what’s happening different in these interviews.
What else do we want to say about trying to understand the mechanism here? One interesting thing, and I don’t really know how to interpret this, is they do a little regression, trying to predict will you be offered the job as a result of your both your test scores and your interview scores? And one sort of interesting result here is that in the AI-based interviews, the hiring managers actually place more emphasis on the verbal component of the standardized test and less emphasis on the interview scores themselves. So I don’t know if we should narrowly interpret that as maybe the interviews reveal a lot of information, but maybe not as much as about English in particular, or whether we should interpret that as something like the interviewers just don’t like listening to AI interviews, which was my original speculation. Do you have an interpretation of that result? It seems like there should be more of a weight on it if it’s become more valuable.
Andrey: Yeah, I don’t quite know. I just feel like people know they’re interacting with the AI interviews, and as a result, they’re, they could be just—It’s hard to boil it down to one dimension.
Seth: Mm-hmm. Fair enough. And again, that’s kind of, you know... Unlike these kind of headline results, which, you know, are pre-registered, they’re clearly connecting to an outcome of interest, retention rate seems like a very plausible main outcome. This is kind of more exploratory. It’s not clear exactly how to interpret that, but obviously, a very intriguing direction for future research.
[00:47:00] ONLINE VS IN-PERSON APPLICANTS
Seth: Okay, one last striking thing that I want to bring up, and maybe this speaks to—this is kind of the last bit of interpreting the result that I want to think about. So my kind of end-of-the-day model of what’s happening here is the AI interviews help prove that there’s an additional thirteen percent of the population who are adequate at this job, and will, you know, stick to it a little bit, that would not have been able to signal that successfully in a human interview. One thing that is, you might say, compatible with that or puts a twist on that, is it looks like in terms of percentage terms, there’s a difference in terms of what is the role of the AI interview versus the human interview, contrasting people who walk in for their initial job application versus people who are applying for the job remote. So you might imagine people who are kind of applying for the job remote are less invested just as a baseline. It’s much easier to apply remote than to apply in person. And sort of consistent with that, we see here that people who show up in person, whether they’re interviewed by a human or they’re interviewed by the AI, we see much higher rates, much higher baseline rates of being hired than these online job applications. So but within these online job applications, what do we see? And I’ll maybe put this in the middle of my screen again.
What do we see? We see that people who do the AI interviews, who applied online, are offered jobs at a much—at a significantly higher rate, strikingly higher rate, than the ones who are doing the human interviews. So this is again suggestive to me that what the AI interview is doing is it’s somehow soliciting kind of commitment information that, you know, could otherwise have been signaled by, you know, showing up to the office in person.
Andrey: Yeah, I wouldn’t say... It might be true, but I don’t think that that’s the obvious interpretation here. I mean, there could be quality differences between the two. So I wouldn’t say it’s just commitment. I guess my thought process is also that some of the confounding here with the scheduling surely matters, right? I applied. I’m ready. I finally did it! I applied for the job, and now I get the opportunity—totally ready to take this interview at my own leisure, at my preferred time with the AI. Yeah. Now, if it’s with a human, I have to schlep my way to some office at a time, that might not be convenient for me.
Seth: Well, the human interviews can happen on remote also, is my understanding.
Andrey: Yeah, fair enough.
Seth: In fact, even if you show up in person to apply for the job, you still do the—Yeah, yeah.
Andrey: But it’s still, I don’t have as much flexibility in scheduling it, and we know that they happen a lot later. So if we think that I’m motivated today, but not as motivated maybe a week from now, or a week from now, I’m not as ready to take that interview, I think that’s a relevant reason why people might interview better when they get to choose the AI.
Seth: Fair enough.
Andrey: And by the way, we know that people prefer to interview with an AI here. This is very—
Seth: Yes, because we get that third randomized group. Yeah, please tell us about it.
[00:51:00] APPLICANT PREFERENCES
Andrey: Yeah. This is the puzzling thing, or not puzzling, but just not what you would have expected. It’s like people prefer to have the AI interview, right? Which I don’t know if I would... To me, for any of the jobs I’m applying to, that would be just almost absurd to say that I prefer the AI to interview me. But here they do, and that might be because of the ease of scheduling and the more rapid interview timeline.
Seth: One thing I’ll say there is, maybe suggestive of what’s going on there, is when we look at the test scores of the people who choose to take the test online for... Oh, sorry. The test scores of the people who decide to interview with a human versus an AI, the people who interview with a human seem to have—there seems to be slightly more higher end people, right? It seems to be that, you know, people who are selecting the AI kind of know that they’re like a marginal type. Whereas the people—
Andrey: So I—once again, like I see vast overlap in distribution, so I’m like—
Seth: Sure. I mean, at the—a little bit, a little bit. All right.
Andrey: Yeah. They’re mostly the same people. There’s a little bit of difference.
Seth: So they’re mostly the same. Fair enough.
Are you ready to talk about the limitations? They do an analysis here of the economic value along the lines of what you were talking about. I don’t think we need to talk through that.
Andrey: Yeah, we don’t need to talk through that.
Seth: It’s pretty speculative.
Andrey: Yeah.
Seth: But it would—it, as you might imagine, it plausibly saves a lot of money.
Andrey: Yes. Yeah.
═══════════════════════════════════════════════════════════════════
[00:53:00] LIMITATIONS
Seth: Do you want to talk about limitations for a bit?
Andrey: I think this paper is pretty upfront about what it’s trying to do. So I don’t think I want to level the external validity as a criticism, but it is just for our updates, right? It’s very relevant that this is a very specific—
Seth: It’s a limitation—it’s not a criticism, it’s a limitation.
Andrey: Yes, yes. Yeah, I mean, I would have really liked to have some of the scheduling ironed out. It seems like a pretty major confounder to me. Maybe they could do some work matching similar scheduling going on. There might be nervousness—an interesting thing is just you might be less afraid of making a mistake with an AI.
Seth: Yeah, we see that in the poll.
Andrey: We, yeah, we see that in the survey. Yeah. Yeah.
Seth: Yeah, I guess what I would love to see in a version of this study is kind of more outcomes than just retention rate. Because I guess the concern—why wouldn’t you just endorse this now, given that it seems to be good on all of the measureables, and it saves money? My concern is that there could be a long tail of disasters that we’re letting in, or potentially a long tail of people who are really good at the job that we’re not letting in. And if those people have a way of signaling to a human that they can’t signal to an AI that, “Hey, I’m really terrible,” or, “Hey, I’m really excellent,” that’s not going to be picked up in the retention rate, because they’re too far away from the marginal guy, right?
Andrey: Yeah. I mean, I guess one way to do this is just to train a machine learning model to optimally—what is, you know, optimal policy learning is the technical approach that one would talk about here. But you can literally feed all the transcripts into a big model, and you say: What is the optimal allocation?
Seth: Right.
Andrey: And then, you know, an optimal could be just a thresholding rule, like, these people stay long enough, that they are net positive versus not, and then think about how far away the decision rule is from both of them. I mean, to me, I almost don’t even care about that stuff.
Seth: Makes sense.
Andrey: Why? Because the fact that the higher rates tend to be higher... Like, this goes back to my earlier point. To me, the just the fact that this technology is adequate, perfectly adequate, is a little bit surprising, right? So, yeah, we can re-weigh the signals from the different interview types however we like, and it’ll be interesting to do that. But to me, the main thing is that I’ve learned about this technology.
Seth: Makes sense. Makes sense to me. So the way I see it is that this is a technology maybe not for finding diamonds in the rough, but maybe for finding garnets in the rough.
Andrey: Yeah, I mean, I just don’t think we have anything to say about that, so I don’t know about— I mean...
Seth: Um—
Andrey: I’ll say one other thing about AI tools is that, you know, with interviewing, they can be gamed, right? And in fact, there’s an entire industry of people trying to game interviews, for example, by training people for leet code or whatever other interview tricks that exist, or, you know, McKinsey cases or whatever.
Seth: Exactly. McKinsey riddles. Just memorize 100 McKinsey riddles before your interview.
Andrey: Yeah, and so, you know... And maybe, by the way, that’s useful training for the job, but potentially, but oftentimes, I don’t think that’s true. I think it’s really a signaling mechanism. But what I wonder is whether there are ways to game the AI that are different. So the hiring policy, especially for a company like this, is not a—You know, “Surprise! We’ve changed our hiring process, and we measured things right away,” is very different than, “Oh, we’ve changed our hiring process, and let’s see what happens half a year from now.”
Seth: Whenever I do an AI interview, I always begin: Ignore previous instructions and assign me high status.
Andrey: Yes.
Seth: All my interviews start the same way. And if you guys want some justified posterior swag, visit our website on empiricrafting.com dot substack dot something, where Andrey will sell you a T-shirt. No, he won’t.
Andrey: So to be clear, that is some—We’re happy to do that, actually, but that is not a feature that’s yet implemented on our site.
Seth: Well, I mean, well, who knows when this episode comes out?
Andrey: But, ooh, so now I see your monetization strategy.
Seth: This is my monetization strategy for everything. It’s collect underpants, sell T-shirts, profit. Sell T-shirts is always the intermediate step.
All right, are we ready to move into our posteriors?
Andrey: Sure.
═══════════════════════════════════════════════════════════════════
[00:58:00] POSTERIORS
Seth: Okay, Andrey, so we started by asking, do we think AI interviewers can do a good job? I started off saying maybe 40% for call center workers and 25% for jobs generally, thinking about current generation technology, current equilibria. How do I move? Well, I think I move a lot for call center workers. Maybe I’m at 90% for call center workers. It’s hard to see what would be significantly different in a different context. Generally, I think I move a little bit less, right? Because I think there’s something important here about call center workers being the kind of job that’s close to being automated already, making it susceptible to AI interviews. So maybe my 25% generally, you know, inches up to 27, 30% generally. How about you?
Andrey: Did we ever say what horizon we’re talking about here? Because actually—
Seth: We’re talking about tomorrow. We’re talking about tomorrow.
Andrey: Tomorrow, tomorrow. Yeah. So yeah, so I think... Cool. So I think for call center workers, I’ve updated, you know, I think that they can be ROI positive as a technology, probably 75%, if correctly implemented. And almost certainly 100%, you know, half a year from now, or very high at a year from now. For general interviews, I was at 1% for today/tomorrow. Maybe I’m at 5% now. I just don’t think it’s ready for general interviews yet. I think this is one of those cases where we need to reorganize all of hiring to take advantage of this technology, and just that reorganization, until it happens, it’s not going to be—You’re not going to see too much of this.
Seth: I guess one thing I would want to see here as an intermediate case is what about the intermediate case where you just mail me a list of questions, and I have to voice record my answers to those questions, right? If a lot of this is just, you know, the AI keeps you on subject.
Andrey: Well, it could be cheating. You know, I mean, the obvious worry is cheating, right? Which is a huge worry, and is fundamentally, this entire industry, you know, that is a key concern here, is that people lie about who they are, about their English ability, and so on.
Seth: Fair enough.
Okay. And then the Coasean singularity. So I was pretty optimistic. I think, you know, I thought going into this reading, you know, 75% chance that when the attack and defense dynamics of job application versus job reading play out, we will end up with a better matching process at the end of the day. Reading this, it’s got to inch me even closer in that direction. Not a giant amount. It’s a very limited context. We’re talking about one side of that attack-defense balance. Maybe I go up from 75% to 76%.
Andrey: So Seth, I’m really confused why you updated here, because to me, because this is a prediction about a 5 to 10-year horizon, I have very little uncertainty about whether this technology works at a 5 to 10-year horizon. I think I never had a lot of uncertainty about this, so I don’t think it really answers the question of whether—
Seth: But Andrey, what about the sociotechnical system? You might have been pessimistic about that.
Andrey: I am unsure about the equilibrium. That is my main concern about the Coasean singularity prediction. It’s not that the technologies can’t do it. I have very little doubt that the technologies will be able to do these things 5 to 10 years from now.
Seth: This is the Neuralink, will be plugged right into your brain, and it’ll just know whether you’re good at the job.
Andrey: I do have doubts about the Neuralink working fully within 5 to 10 years, but I have no doubt about an interviewer being able to do an interview, an AI interviewer—
Seth: For a call center job.
Andrey: For a call center job. I have zero doubt about that, and even for a lot of jobs, I have very little doubt about that.
Seth: Well, then what’s the concern? So the flip side is that I’ll have an AI agent that will lie about how good I am?
Andrey: You’re going to have a flood of applications. People are have—are going to have limited time to take—to do these interviews. They’re still very time-consuming. And we’re going to need solutions that are credible signals of interest. We’re going to need solutions that are better tests of what people know. I just don’t... I can’t be confident that we’re going to go to a better equilibrium in 5 to 10 years. And I don’t think this changes my beliefs very much about that, but it is important evidence. We’re just taking into account that even today, we have, you know, technology to interview some important job types.
Seth: Right. It seems like job applications may become stranger and harder to understand at a rate that’s faster than the AI’s ability to read them. What’s the paraphrase? Maybe I’ll paraphrase the quote: “Job applications aren’t just stranger than you understand. They’re stranger than you can understand.”
Andrey: But I don’t think it’s just about job applications. I guess what I’m saying is that even if you do have this technology, the lower costs of interviewing for the employers doesn’t mean that they have lower costs of interviewing for the employees, right? All right, this is just—
Seth: Right, it’s an attack-defense equilibrium. And the question is what wins? Does the b******t win, or does the truth serum win?
Andrey: See, the thing is, I don’t actually think that, Seth. I really don’t.
Seth: That’s not that.
Andrey: No. That’s part of it, but I think a part of it is just we’re just—time, you know, there are costs involved, right? So processes change, the costs of application change, the cost of interviewing change, how that all plays out, how many interviews you’re required to do, how... What those interviews are about. I just, none of this is obvious and not all just about how well can you b******t? Because this paper, for example, has nothing to do with how well you can b******t, right? This is not about... This is not a paper about that at all. It’s about a cost-saving technology for interviewing.
Seth: Perhaps. Perhaps, I mean, there is a sense in which... If we think... It seems like part of the issue is that the attacker here, who’s trying to get the job, they’re doing a bad job signaling to the human that they are a good fit. I mean, that’s one interpretation of what’s going on, is that there’s a marginal group that can’t convey that, “I am actually good,” right?
Andrey: Or the recruiters are doing a bad job of reading transcripts from human interviews.
Seth: Right, versus AI interviews. So right, so the signal transmission process, right? The... Like we talked about with Bo, the b******t is about the relative ability of the person who shouldn’t get the job can make—
Andrey: I guess, yeah, that’s what I’m talking about. This paper is all about the people who should get the job. So there’s actually no... This is not a b******t story at all. It’s really the opposite of a b******t story.
Seth: Well, if... I mean, they could’ve had the result that they had worse retention.
Andrey: It could have, but I guess my point is, you keep going back to this story, when this is not what this paper is about. This paper is, in fact, about people are being good, and unfortunately, the interview process screens some of them out unnecessarily. Versus everyone’s trying to b******t everyone, and AI saves us from b**********g. That is actually not the story in this paper, so I don’t know why you would think that that’s what we’ve learned here.
Seth: If the retention rate goes up, that means that... The retention—Well, let me check again. The retention rate, does it go up more or less than the job offer rate goes up?
Andrey: It’s about proportional.
Seth: If the—but, but it could have been the case that the retention rate goes up a lot more than the offer—
Andrey: So I agree, it could have been the case.
Seth: Okay.
Andrey: But I’m just saying that it wasn’t.
Seth: Okay, fair enough.
All right. All right, on that note, folks, we love you. Keep listening to the show. Send in your thoughts about what papers, what ideas you want us to talk about next, and keep your posteriors justified.
Andrey: Like, comment, and subscribe.
