Robot Recruiters: How AI is Helping Decide Who Gets Hired – Terms of Service with Clare Duffy

All right. I’ve got my resume. I am on jobscan.co, and I have just uploaded it. And I’m going to see how good of a fit it thinks I am for my boss’s job, so the CNN tech editor job. The other day, I tried uploading my resume into an online resume scanner, which promised to evaluate it and give me a score. And theoretically, that score would tell me how I would be perceived by recruiters or hiring managers. It looks like I have lots of issues to fix with my resume. There’s searchability issues, there’s hard skills issues, there’s lots of soft skills issues, and there is at least one formatting issue. Many people are using tools like this nowadays, and it makes sense for job seekers to get a bot’s take on their resume before hitting apply, because these days bots are often screening resumes before people do. It has picked up on the fact that there are a bunch of skills listed in the job description that aren’t necessarily listed on my resume. So it says that I didn’t list media industry on my resume, although that’s in the job description. That seems weird, like a thing I would never actually write on my resume, but it likes that I included journalism, management, editing. These are all things that were in the job description that were on my resume.

These days, AI is widespread in the recruiting world. It’s more than possible that an everyday job seeker will be put through some kind of AI evaluation, no matter what industry they’re applying to. So if you’re applying for a job, what should you know about how this technology is evolving? Where can AI help in the hiring process and where can it go wrong? I got a taste of this for myself when I found out that my resume was rated just a 60% fit for a CNN tech editor position, which I thought was surprisingly low. And don’t worry, Lisa, I’m not coming for your job. But to get some more concrete answers about how AI is changing hiring, I talked to Hilke Schellmann, an investigative reporter and assistant professor at New York University. Hilke has tips for how to stay ahead of the curve and what to know about AI in the workplace once you’re hired. I’m Clare Duffy, and this is Terms of Service.

Well, Hilke, thanks so much for doing this with us.

Yeah. Thank you for having me.

So, when did you first hear about AI being used in hiring?

‘It’s been a while. It was kind of an interesting, fun chance encounter. I found out about AI and hiring by taking a left right. I asked the driver, how are you doing? And he said, you know, it’s been a weird day. And I was like, oh yeah, why are you having a weird day? And he’s like, you know, I was just interviewed by a robot. I was like, what? He’s like, yeah, I applied for a job as a baggage handler at a local airport. And that day, he got a call from a quote unquote robot asking him three questions. So today, we know is probably a pre-recorded voice, but, you know, to him, this was in late 2017. This felt like a robot was calling. And…

Yeah.

he was definitely kind of weirded out. He had a weird day. And I was like, what? Robots doing job interviews. You know, I never heard of it. And then, you know, I forgot about it until a couple months later, I was at a conference and somebody who had just left the Equal Employment Opportunity Commission, spoke there in the afternoon about how she can’t sleep at night because companies use very basic algorithm, this was in early 2018, to go through the people’s their employees calendars to check how many hours you are at work. I started talking to her and I was like, have you heard of these like robot job interviews? And she’s like, oh yeah, there’s a company called HireVue that does one way video interviews with AI. You should call them. And you know, that’s how it all started. And little did I know that there was a whole world out there. You know, at the time, in early 2018, there was already Unilever, Hilton, like large companies, were using AI based video interviews for hiring. And it was, you know, just took off from there. They just happened to have this chance encounter so early on, right, that I got this tip from the Lyft driver that this is happening. And, you know, and I think that speaks to sort of a an occupational divide that like, you know, we see a lot of these AI hiring tools used on what the industry calls high turnover, high volume jobs. So meaning like often hourly workers in businesses that have to hire lots and lots of people. Right. You see this in retail if you want to have a job with any of the large retailers in the US, they usually have some sort of a AI screen, right? In fast food. And then over the years we’ve seen it creep up. But, you know, I hadn’t been looking for a job for years, so I would have never encountered it.

We’re going to get into sort of the nitty gritty of how it all works, but how like from 2018 to now, how have you seen AI’s role evolve and expand in the workplace?

I can say, you know, it’s kind of interesting. And, you know, this started like sort of in the early aughts when, like LinkedIn, Indeed, Monster, all of those job platforms, they started with saying, like, we’re going to democratize hiring. Like, you can find all the job openings here. You take a couple of clicks, you can write a couple letter, send it out your resume. And so starting about 20 or so years ago, like recruiter started complaining about a very high volume of resumes and that actually now it’s sort of in a frenzy, right? With the introduction of ChatGPT and other large language models. What we’ve seen is like, oh, instead of like, you know, an average of maybe 200 or so applications per job, we now see about 50% more. And it’s probably because job seekers use ChatGPT and other tools to like, sort of help them generate cover letters and help them with their resumes and so they can apply to more jobs. We have recruiters just drowning in resumes and job applications. So there’s a very, very large applicant pool. So AI vendors have started to like you know know about this. So they pushed into the space and say like no, we can make this easy for you, right. Like we going to build AI tools that will sort through all of the candidates. They do it in a bias free way. It’s going to be much more efficient. You know, you need way less labor to do this kind of work. And we’re going to find the most qualified candidates. So when I found out it is true, it saves company lots of money. It’s way more efficient. What I found actually counter evidence is that actually these tools are not bias free and they don’t necessarily pick the most qualified candidates. But that’s sort of the promise of these tools.

What these AI tools are doing for companies is whittling down and ranking a pool of applicants based on certain criteria that’s built into their algorithms. So what are these algorithms looking for in job candidates?

There’s a couple ways that what we call resume parsers. There’s like basic technology that looks at scans, all the resumes and sort of figure out, like, are you qualified for the job or not? And there’s sort of very basic algorithms that look at the job description and check out the candidate’s resume and sort of understand like, okay, how much overlap is there? They’re often very inefficient because it turns out job seekers are pretty smart, right? They understand, okay in the job description it asks for all of these skills, I’m going to make sure I’m going to address all of these skills in my resume. Right. So like you want to screen down the applicant pool, but if you have a basic screen like that, you actually don’t lose a lot of job seekers in this process. Right?

Right.

There is another way that companies use resume screeners, where they actually feed the resume screeners for the resumes of people currently in the role. So I don’t know, you want to hire new accountants. You have 50 accountants or 100 accountants that already work for you. And presumably those are sort of the successful people. Right? So you sort of declare them successful, take their resumes, feed them into the system and tell, you know, an AI tool, hey, figure out what all of these people have in common, and then you compare incoming resumes against that model.

In theory, it’s a compelling idea for large companies to use AI to comb through hundreds of resumes. But Hilke has seen the algorithm get some things very wrong because, as we’ve talked about on other episodes, AI tools can sometimes recreate rather than prevent the biases we often see among humans.

Some of the companies do their due diligence, they bring in outside counsel who looks exactly like, okay, what are the keywords that the resume parser uses to infer that somebody is going to be successful? Right. So what we’ve seen is the tool does a statistical analysis, and in some examples found out that the word Thomas was predictive of success.

Oh wow.

And other times it was locations like Syria, Canada, and you know, it’s those kind of keywords could get you actually into legal trouble. Right? And then we’ve also seen other keywords like Africa and African American that could point to racial discrimination. In one resume parser, the tool gave more points to people who had the word baseball on the resume, and if you had the word softball on your resume, you got fewer points.

Wow!

And you’re discrimination, right? Because in the US, more women play softball and more men play baseball. And you know, this wasn’t a baseball coaching position, right. This just was like some random job that have nothing to do with sports. And probably what happens, the hypothesis is in this stack of resumes that the parser analyzed, maybe there were a bunch of people who had baseball on their resume, and the tool did a statistical analysis and found out exactly. Oh yeah, it’s totally significant, except that AI tools don’t have a consciousness, right? They wouldn’t understand that like, wait a second, baseball has nothing to do with the job, right? Just because you know baseball, you’re not a better accountants. So I think that the problem that we see again and again, again and again replicating in these tools at the resume parser level. You know, it’s not that humans don’t discriminate against other humans in hiring, right? But the threat of AI could be that it’s used at a scale that is unprecedented.

And when we talk about the AI hiring tools, how effective is that technology at actually finding employees that are the right fit for the jobs that companies want to fill?

Oh my God, I wish I could tell you that. I wish I could answer that question.

Yeah.

‘What we do know is that company executives themselves are starting to be dubious that these tools really work. So what we know is, like Joe Fuller, who’s a business professor at Harvard, did a large survey of over 2000 C-suite executives in Germany, UK, and the US and what was interesting was when the company used AI tools in hiring, almost 90% of C-suite leaders said that they know that the tools reject qualified candidates. So they know that they don’t work necessarily, as promised, right, that a lot of people get rejected who are qualified. You know, it takes hundreds and hundreds of applications these days to find a job. And it’s really hard on job seekers, right? You know, after that survey, you often feel like we have so many recruiters and folks in HR complain that they don’t find the right qualified candidate and job seekers who complain that they are qualified but they can’t get through. It’s like maybe it is the AI tool sometimes that, you know, like are responsible for this mismatch.

Some companies have made an effort to do something about this. In 2021, a group of big brands including Nike, Walmart, Meta, the NFL and others formed the Data and Trust Alliance to teach corporate HR teams how to evaluate these tools and avoid bias. Still, there may be more work to be done to make sure these tools are working for everyone. So the companies are using AI as this first step, both to sort of like get people to apply for jobs and then sift through their resumes and cover letters. Will you talk through the next steps of the hiring process and how AI plays in, in things like an AI powered interview?

‘Yeah, totally. So we see the resume process at the beginning. If you, as a job seeker, have gone through the first screen, what we often see is that companies are asking job seekers to do a one way video interview or audio interview. So what often happens is like, you get a link sent to you, and then you open the link on your desktop or on your phone, and you go into sort of like a tool and there’s no one on the other side just basically having a job interview with yourself. You get pre-recorded questions. You know, there might be a video of a human saying, hey, welcome to company X. I want to ask you a few questions. You know, why are you interested to work here? And then you get a couple of minutes to prepare your answer, and then you basically record yourself answering. Job seekers have told me it’s like inhumane, disturbing. It’s just also really weird. I’ve done many of them now where you just sort of like want to try to sound really chipper and excited, but there’s no one on the other side.

You know, talking to a human. Yeah.

You know, it is like a kind of weird experience. I do have to say. And they overwhelming amount of job seekers that I’ve talked to they don’t like this kind of interviewing. They don’t have the chance to ask questions, right? And they don’t know what happens to their interviews as well. Right. Like as human watching them as an AI watching them, it’s not necessarily always clear what happens. And, you know, sometimes you get rejected at that stage. You know, you wonder like what happened to this interview with anyone analyzing it or watch it. Unclear.

So do we know anything about how it typically works? Like is it an AI system that’s looking at those answers and trying to glean some information, or do humans watch them?

We know very little. So what I can tell you is, like the largest vendors in this space often have two products, right? They have a basic straight up video interview where you’re record yourself and like a human on the other side, watches those videos. And then most of the companies also offer an AI solution where an AI tool then quote unquote watches the videos, right? I can usually what happens is like they take the audio. It gets transcribed into text. And then the AI tool infers, based on the text, how likely you are to be successful in the job. Right? It’s like infers like maybe things like how good of a teamwork are you? What are you strength in capabilities? Based on the words that you used. A few years back companies also use the emotion facial expression analysis, the tone of candidates voices. They still sometimes use that. But I do have to say the largest vendors stop using it because I and others have reported on that there is no science underneath this, right? Like we don’t know what facial expression you need to be in a job interview to be successful in the job. There’s no signs of that. You know, the question is like, technically, yes, you could scan my face for facial expressions, but is it meaningful?

Right? Just because you can do it, doesn’t that mean you should be doing it.

Exactly. And that is asking the question, right? Like we have the technical capabilities, but is it meaningful?

And what about the employees, the job seekers that you’ve spoken with? Does interviewing with an AI or knowing that your resume is going to have to go through one of these AI screeners, are folks put off by that? Like, does that make them less likely to potentially work with a company that they’re doing an AI interview with?

Yes. You know, I’ve talked to so many job seekers. There were a couple who did like actually the AI video interviews that they felt like, you know what, I can do this at like 2 a.m. on a Sunday when I have a minute. That’s great. And they also felt like, you know, I can talk about myself. There’s no one in the room. I’m not that nervous. So it appeals to some people. But the overwhelming majority felt like this is dehumanizing. They don’t get to ask questions, and some people actually took themselves out of the process. And so I’m not going to keep interviewing with this company because I just feel not treated well. And if they treat me already this badly as a potential colleague, then I don’t want to work there. And the thing is, as a job seeker back in the day, you could be very choosy. It’s harder to ignore those video interviews because they’re just more so, more ambiguous, right? Like, if you want to have a job today in banking or investment banking as a young graduate, you have to go through these AI screens. I don’t think there’s a way around it.

Hilke wanted to test one of these AI interviewing tools, so she did an experiment where instead of answering any of the questions, she just read a psychology related Wikipedia page entirely in German. The AI tool rated her a six out of nine in English competency and a 73% match for the job, putting her in the top half of applicants. This raised some big questions about the effectiveness of this technology.

So I do something that I call like maybe prodding the algorithm like needling a little bit, right? Like, you know, I’m not a coder. I can’t do like tests with like thousands of entries or something like that. But, you know, I can try some of these tools myself. So these were two different tools. And, you know, one criteria was like, how well do you speak English as a candidate? So I did that interview myself, speaking English, I got at 8.5 out of nine, English proficient. And then I thought, what happens with people who have an accent or have a speech disability? Because what the tool does, right, it doesn’t take the audio recording. It does a transcription of the speech into text. So this speech to text transcription, doesn’t treat everyone fairly. Right. So I felt like, you know, well might as well speak German. It’s my first language. And I think I still speak it really well. And so I thought like, let me read that Wikipedia entry on psychometrics, which is sort of the science underneath this kind of testing. It’s called psychometry. And so I just read that article in German every time I was prompted to give an answer, and I thought I would get an error message. But lo and behold, I got a score and I got a six out of nine English proficient! And I was like, wow, I didn’t say a word of English.

Yeah, haha.

I mean, look at that. Even speaking German, the tool found ou I speak really well, English or something, but you know. So I think those are kind of like little tests that I think actually anyone can do, right? All of us now get pitches for AI tools all the time. So can we test those algorithms to sort of understand, like how do they work and do that work? Right. Like I did another analysis where I just said I love teamwork 50,000 times as answers. It’s the only four words I used, and I still get a super high rating on another algorithm. Like, you know, I was like 72% qualified for the role. And I was like, I didn’t say anything except like four words.

What is your takeaway there that the tool just, like doesn’t work, or that it might be looking for something different from what you thought it was?

Yeah, I would say that it probably doesn’t work, right. If you speak German and you get a rating that you’re proficient in English, that probably doesn’t really work, right? The AI vendors really need to think through like, how do our tools work, what happens in these edge cases, and do they point to something larger? Right. Because their lack of an edge case doesn’t work. So does the tool work at all?

So the jury is still out on how effective AI powered recruitment tools actually are, and whether they can really pick the best candidates out of the big lineup. But like it or not, these tools seem to be here to stay. And once you’ve landed a job, you might find AI following you into the workplace, too. That’s after the break.

So once you’re hired, how are companies using AI as sort of tools to monitor employees ongoing performance? And do you have a sense of what metrics those tools are relying on?

Yeah, technology to sort of monitor people has been around for a while. I think it got a huge push during the pandemic? Right. When a lot of people started working from home, a lot of managers and like higher ups at companies, get really worried, like is the person working at home. And so then, you know, much more of this kind of software was being installed on people’s computer. And it goes from like keystroke tracking, where like everything you hit on your keystroke gets recorded to screenshots of you sitting in front of your computer. We also now have sentiment analysis, and it’s called employee listening, which I think is a euphemism for listening to everything employee’s write. Meaning there’s tools that check slack channels, emails, every kind of text trace or video trace or anything that employees leave behind. It also checks meetings and everything like that to check for like, are there any signals that people are unhappy with their company? Are there signals that maybe you’re moving large swath of data, meaning you might be leaking information, or you’re trying to leave the company and take with you some trade secrets, right? They’re looking for compliance. Some of the tools say they can find bullying, discrimination and all kinds of things. And then those signals get sent up to legal, IT or HR, right. Whatever the AI tools find. So we see in some instances, we see a whole lot of monitoring of employees that they may or may not be aware of. Right. You know, many companies do predictive analytics on what what it’s called on their employees. And some of those, for example, if you’re are flight risk, are you likely to leave the company in the next 12 month? And printing a lot might be an indicator, right, that you are like printing out information. You get ready updating your LinkedIn so it takes internal and external signals as well into consideration. And it may or may not be true. I think that’s the problematic thing here that we see. We see a lot of inferences. But again, you know, sort of similar to hiring. Like we see a lot of technology that can track certain things, but are they meaningful is sort of the next question.

And what about AI being used in employee performance reviews? Will you talk about how that works and is that an effective use of this technology?

Yeah. I mean, employee performance is really difficult, right? Because, you know, outside of like a sales job, it’s actually really hard to say who’s successful at their job. Right? And what we now often see is like, you know, you benchmark certain employees against each other, meaning they’re like two vice presidents of product or whatever in the company, one in Europe, maybe, and one in the U.S. so the too will record all of their data and sort of just like push them against each other, meaning that like, you know, you get like, well, this other VP sends 200 emails a week, why don’t you send 200 emails a week? You should close more tickets. Everything gets brought down to sort of technical things that can be counted, but it doesn’t really take into account, you know, maybe you have helped a lot of your employees this week, right? And you’ve maybe done this outside of your computer. You didn’t send 200 emails, but you mentored them and helped them to be successful. None of that gets recorded, right? Only these very clear tasks that a computer can record get recorded. So we see this a lot. It’s actually really, really hard to understand what makes you successful. And it’s especially hard to use, like sort of like technological metrics for that.

What types of businesses are using these tools? You talked about some of the really big players, but is this being used by smaller companies as well?

Yeah, I think I think especially in like startups and tech companies we see this. You know, it doesn’t really matter the size because the technology is like super accessible. Like a lot of times it comes built in. So like you run a conference call on video chat, right? Then there’s maybe analytics already built in. So there’s almost no threshold anymore to use these tools, anyone can use them.

Have you heard of any companies who have actually changed or walked back their use of AI in hiring or monitoring employees because of these issues, either with bias or just with the overall effectiveness of the technology?

Yes. Yes, I have heard that. I have talked to some, you know, HR managers who are like, oh yeah, we use the same questions and we really didn’t think it was working, so we stopped using it. And I was like, can you talk, please, about this publicly so that the next company also understands that you did the work and you analyzed and you found that this isn’t effective and they’ll like, oh no,

Yeah…

We can’t talk about this publicly. So I think that’s sort of the problem in this, that the way the system is set up, right, That like the AI vendors obviously wouldn’t want to talk about if their technology gets dropped. Right. But the companies that dropped the technology also don’t want to talk about it because they are afraid that they have liability. Right. And then the job seekers have no idea how these algorithms work. Right. Um, so they’re none the wiser. So I think that sort of leads to this unfortunate loop that we don’t learn because no one publicly acknowledges, oh, this tool didn’t work. So I think that is sort of the unfortunate way that this field is set up. And I think that’s why we haven’t seen a lot of litigation, because like in court, you have to prove that you have been harmed. But, you know, job seekers get rejected all the time. That’s part of the game, right? Like, what do I know when I get rejected after a one way video interview that I don’t even know that AI is probably used to me, let alone do I have any evidence that I’ve been harmed by AI, right? And you have to show harm to bring forward a legal case. So I think that’s kind of shielding the vendors and companies in this case.

And what about are we seeing lawmakers take any interest in this application, this use of AI by companies?

Yes we do. So I have testified in front of California lawmakers about the things that are found in other people have, as well in other lawmakers. And we’ve seen not a whole lot of appetite on the federal level to do a whole lot. But we see like different states, right? Like Colorado has started having law. There’s a law in Illinois, in Virginia. And I think that makes it then hard for vendors and companies, because we see there’s sort of this patchwork of regulation in the US. You know, I wish we could mandate at least transparency and accountability that companies would have to tell you, here’s how your information was used. Here’s how we came up with the score. Companies often don’t want to do that because they feel like, you know, this is like their proprietary algorithm. This is like a trade secret. So I wish we had that, but we really do not have that yet.

Yeah. Knowing all of this. Do job seekers need to be changing how they apply for and interview for jobs to keep up with this use of AI?

Yeah, I mean, it’s a really difficult question because often it’s really hard to say what every individual model does. I mean, I think in general, like I’ve learned sort of best practices that will generally apply, especially for resumes, right? You know, your resume has to be machine readable. It used to be that we would say like find a way to stand out. No, please don’t stand out anymore. Don’t have any graphics like none of that, a computer can ingest or read. Like it has to be short sentences to the point. Quantifiable results. Like, you know, you didn’t save the company money, you saved it $1.2 million. So anything that’s quantifiable. We also see a lot of companies trying to hire from with skills. So I think it could be helpful to possibly have like your own skill section on your resume. There are also tools out there where you can upload the job description and your resume will tell you how much overlap there is. And you probably want to aim for like 80, 90% of overlap. Don’t aim for 100% overlap, because some resume screeners will reject you, because it will infer that you just copied the job description. So you don’t want to do that. For the video interviews, that is incredibly hard. I think what we know is it’s like short answers are not helpful. And I know that from the AI vendors themselves that they said they had very low scores for people who give very short answers. So like expanding, talking about examples where this apply, right? Not like just like, tell me one time you were a team player and you’re like, you know, I helped a colleague once, like, no expand. Like what exactly happened? How did you do it? Like, I think that can be really helpful.

What about job seekers using AI themselves? Like, is that a way to make sure that the AI screening systems, like your resume or your cover letter, if you use AI to write them? Or is that a bad idea?

‘It’s really hard to tell. I mean, you know, recruiters are in generally upset about the higher volume of candidates. So they’re coming through, right? They’re also upset. So like oh like so much is ChatGPT and other generated. It sounds the same. But I mean, you know it’s really hard to tell you because of resumes often do sound very similar. Right. It’s like semi-structured language where we sort of talk about the skills all the time. So we do know that job seekers use it a lot. And if it gets you, you know, especially if, like English isn’t your first language, it can be really helpful to have ChatGPT or other tools, check your resume, right? So I think you do want to obviously fact check because, you know, all the tools are prone to hallucination. So it may make up that you wrote a book when you didn’t, or something. So you want to absolutely fact check that because you don’t want to lie on your application.

Are there takeaways for you for company HR, or hiring managers about whether and how to use these kinds of AI tools for hiring?

Oh, totally. Like, I wish HR folks would be much more skeptical about these tools. Like, ask very basic skeptical questions and how the tools are trained to pilot studies before unleashing quote unquote, the algorithm. Right? Like, you know, have like a sandbox where you test them and you check the results. So I think, like much more scrutiny and skepticism is really helpful doing these pilot studies and in general sort of thinking through like, okay, I have a problem to solve is there actually like technology, who can solve that problem?

Awesome. Thank you so much. Really appreciate it.

Oh, I really appreciate it. Thank you for having me.

So if you’re job hunting right now and feeling overwhelmed by the AI powered world of hiring, here are some things to keep in mind: As this technology evolves, it can be hard to know what tactics will be effective and make you stand out in a crowd of job applicants. That being said, think about having a skill section on your resume with keywords that overlap with those in the job descriptions for the roles you’re applying for. You don’t want it to be too matchy matchy, but make sure you’re using the most important keywords. Next, you can try using a resume scanner, like I did, to get a sense for how your resume will look through the lens of AI recruiting tools. And if you find yourself in an AI interview, like a phone or a zoom interview without another human at the other end, Hilke recommends trying to use longer descriptive answers with concrete examples. If your answers are short, the AI will have less to work with in evaluating you. Finally, if you are employed, be aware that companies may be using AI to monitor your conversations and activity on company devices. So keep private conversations on a personal phone or computer. Thanks so much for listening to this week’s episode of Terms of Service. I’m Clare Duffy. Catch you next week.

Terms of service is a CNN Audio and Goat Rodeo production. This show is produced and hosted by me, Clare Duffy. At Goat Rodeo, the lead producer is Rebecca Seidel, and the executive producers are Megan Nadolski and Ian Enright. Production support on this episode from Jay Venables and Hazel Hoffman. At CNN, Matt Martinez is our senior producer and Dan Dzula is our technical director. Haley Thomas is senior producer of development. Steve Lickteig is the executive producer of CNN Audio. With support from Kyra Dahring, Emily Williams, Tayler Phillips, David Rind, Dan Bloom, Robert Mathers, Jamus Andrest, Nicole Pesaru, Alex Manasseri, Leni Steinhardt, Jon Dianora, and Lisa Namerow. Special thanks to Katie Hinman, David Goldman, and Wendy Brundige. Thank you for listening.