A Debate about the AI Hiring Arms Race

[00:00:00] Welcome to The Debate. Imagine a submarine navigating through deep, pitch black water. Okay. To understand its surroundings, it relies entirely on sonar. It sends out a single acoustic ping and then the crew just waits. They listen carefully for the echo. Right, because that return signal tells them exactly what is out there in the dark. Exactly. Okay. It is a clean, reliable, highly informational signal. And because the environment is relatively quiet, the submarine can trust the echo.

[00:00:27] Yeah, it functions perfectly because there is a one-to-one relationship between the object and the water and the sound bouncing back. Precisely. But, uh, imagine if suddenly millions of autonomous drones were dropped into that same ocean. And every single one of those drones began broadcasting perfect synthetic sonar echoes all at once. Oh, wow. Just total chaos. Right! The submarine isn't blind, but it is entirely deafened by the noise. The radar screen just becomes this solid wall of glowing green.

[00:00:55] The crew literally can't tell the difference between a trench, a passing whale, or an enemy vessel synthesizing an allied ping. I mean, that is an incredibly vivid way to describe exactly what is happening in the modern labor market right now. The signaling environment hasn't just gotten a little noisy. It's, the ecosystem has experienced a total collapse. And that collapse is the central issue we are exploring today. We are looking at the core problem raised by Dr. Jonathan H. Westover's October 2026 paper in the Human Capital Leadership Review.

[00:01:25] He argues that we are currently trapped in a massive AI hiring arms race. Right. On one side, job candidates are utilizing generative AI tools to mass-produce highly polished, perfectly tailored job applications at a scale we've just never seen before. And on the other side, employers are responding to that massive volume by deploying their own automated AI filters just to survive the deluge. Exactly. And the result is surging applicant volumes, an explosion of sophisticated fraud, and really plummeting trust across the board.

[00:01:54] So the central question we have to answer today is this. How do organizations restore validity to the candidate selection process? Like, do they retreat to traditional, high-friction, human-centric assessments that got through the noise? Or do they have to adapt their systems to evaluate candidates within a new AI-integrated baseline? Well, I take the position that organizations must return to the fundamentals of evidence-based selection. We have to prioritize AI-resilient, structured human assessments.

[00:02:23] You simply cannot fix an automation crisis by piling on more automation. We have to get back to evaluating actual human capability. And I look at this entirely differently. I argue that returning to traditional, high-friction processes is a completely unscalable fantasy. It fundamentally ignores the physical reality of modern hiring and, honestly, the immense benefits of AI. But at what cost? Well, the focus should be on transparent algorithmic governance and redesigning our assessments for the modern reality,

[00:02:53] rather than retreating to a past that, let's be real, didn't work as well as we pretend it did. Let's start by looking at what we've actually lost here, because I think we need to understand the mechanics of the crisis. In labor economics, there is a concept called a signal. Right. A signal is an observable action that conveys a hidden quality about a candidate, like their true ability, their work ethic, or their specific fit for a role. But here's the key. A signal only works if it is costly or difficult to produce.

[00:03:22] Yet it has to take effort. Exactly. A carefully tailored cover letter, for example, used to be a highly effective signal precisely because it took time, research, and effort to write. Generative AI has essentially driven the marginal cost of producing that polish to absolute zero. The numbers definitely reflect that drop in cost. I mean, if you look at the October 2026 Workday Global Workforce Report, 84% of job seekers are now using AI in their search. 84%. Yeah.

[00:03:50] Because candidates can now use automated tools to sort of spray and pray, applying to hundreds of jobs overnight with perfectly tailored AI cover letters, the median number of applicants per filled job jumped from 58 to 69 in just a single year. Right. The volume is skyrocketing. But here is the critical part of that same report. The time to fill the position remained totally stagnant at 60 days. Interesting. Employers are drowning in more applications, but they aren't hiring any faster. They are just wading through more noise.

[00:04:20] And frankly, this environment is actively penalizing top talent. Wait, how does it penalize top talent? If everyone has the same tools, shouldn't the best still rise to the top? You would think so, but the data says otherwise. There was this fascinating 2025 study by Galden and Silbert that looked at a massive online freelance marketplace. They found that when AI makes tailored applications costless, the market actually becomes remarkably less meritocratic. Really? Yes. Less meritocratic. Yes.

[00:04:48] In their study, workers in the top ability quintile were hired 19% less often. Think about why that happens. Previously, a client could spot top-tier talent because their custom proposal contained unique, specific insights that took real expertise to generate. Oh, I see. When LLMs wiped out that premium, when bottom-tier workers could suddenly generate a proposal that sounded just as polished and insightful as the top-tier workers, clients literally couldn't tell who was actually good anymore. So we'd essentially randomize the selection process.

[00:05:19] Exactly. When everyone sounds perfect, true excellence is completely camouflaged. To restore predictive validity to actually find the best people, we have to shift our weight back to methods AI cannot easily fake. And that means live, structured interviews and real-time work samples. See, you are mourning a loss of signal. But I am arguing that this so-called loss of signal is actually a necessary, long-overdue exposure of how deeply flawed and noisy our traditional signals always were. How were they flawed?

[00:05:49] We operated for decades under this massive, unexamined assumption that a beautifully written cover letter correlated with being a great software engineer, or a brilliant financial analyst, or a top-tier supply chain manager. It didn't. Well… It just correlated with having strong writing skills, or perhaps having a socioeconomic background that afforded you the leisure time to perfect your application formatting. AI hasn't destroyed a valid signal. It has finally leveled a very uneven playing field.

[00:06:15] But masking a candidate's actual voice with a machine's voice doesn't level the playing field. It just replaces the human element entirely. But it removes an artificial barrier that had nothing to do with the job. Consider a brilliant data scientist who's neurodivergent and struggles with persuasive writing, or someone who speaks English as a second language. Okay, sure, but… In the old system, their resume gets tossed by a recruiter because of a slightly awkward sentence structure. Even though they can build a predictive model better than anyone else in the pile. Right.

[00:06:44] There was a fantastic 2023 field experiment by Wiles, Munique, and Horton with nearly half a million job seekers. They gave some of these candidates algorithmic writing assistance. The result? Hires increased by 8%. But were those good hires? Yes. There was absolutely no reduction in employer satisfaction with those new hires. Why? Because the AI helped clarify the candidates' true qualifications. It removed the noise of poor writing that previously obscured their actual technical capabilities.

[00:07:13] I hear that. But let's dig into the mechanics of what generative text actually does to human communication. Because I think your data scientist example misses a crucial point about how these models function. Go on. When a candidate uses a large language model to write their application, the employer is no longer evaluating the candidate. They are evaluating the software. And how does that software work? It doesn't possess unique insight. Right. It predicts text. Exactly.

[00:07:38] AI text generation is mathematically designed to predict the most likely next word based on massive data sets. It is designed to be highly plausible and therefore deeply average. It generates a massive homogenization of the applicant pool. What do you mean by homogenization in this context? Think about the cover letters or professional emails you read today. Have you noticed how they all have that exact same overly polite, frictionless, robotic tone? Yeah, I suppose.

[00:08:05] I am highly motivated to leverage my synergies to drive impactful results. Yes. It strips away the idiosyncratic quirks that actually signal a human being's unique genius or specific passion. And this disagreement over what we are measuring naturally leads to how we evaluate it. Okay. If the written application is completely decoupled from the candidate's actual capability, if it's just a homogenized output, it is fundamentally useless as a predictive tool. We have to rely on live human-centric evaluation.

[00:08:35] This isn't just theory. Major organizations are already recognizing the necessity of this shift. You mean returning to the classic in-person interview. Exactly. Google's chief executive explicitly announced the reintroduction of in-person interviews specifically to combat AI cheating and virtual assessments. Right. I remember that. And Goldman Sachs went so far as to prohibit external sources like ChatGPT in their interviews because they stated they need to hear a candidate's, quote, own voice in real time. And they are following the science here. What science?

[00:09:04] Let's look at the 2022 Sackett meta-analysis. For a long time, personnel psychology had a massive blind spot called range restriction. Imagine trying to figure out if SAT scores predict college success, but you only study the kids who actually got into Harvard. You are missing the whole picture because you aren't looking at the people who were rejected. Bingo. Sackett corrected this historical error across decades of hiring data. When you zoom out and look at the actual math, structured live interviews emerge as the absolute top predictor of job performance.

[00:09:31] You just can't fake a live structured conversation with ChatGPT. Look, I understand the appeal of the live interview, but your approach completely ignores the physical reality of the modern labor market. Scalability is an absolute unavoidable constraint. How so? In mid 2025, LinkedIn was receiving an average of 11,000 applications every single minute. That was a 45% increase from the year prior. That is a lot of volume. It's an avalanche. Just picture what a recruiter's applicant tracking system actually looks like on a Monday morning.

[00:10:00] Relying on live structured interviews and in-person assessments creates an absolute impossible bottleneck. Well, I'm not saying... We already know from Greenhouse's data that a third of US recruiters are spending up to half of their week just manually filtering out spam and junk applications. You cannot put 11,000 people a minute through a structured in-person interview to find the right one. The math simply does not work. But I am certainly not suggesting a recruiter sit down and interview 11,000 people. I am suggesting we rebuild the front door of the process. Meaning what exactly?

[00:10:30] We need to stop relying on easily fake text documents and instead filter efficiently using automated but strictly skills-based criteria, like actual tests of capability, and then rely heavily on the high-validity live assessments for the actual decision-making. But the filtering itself requires AI to be manageable. Look at L'Oreal. They receive about a million applicants a year for roughly 15,000 positions. Right. They successfully utilize the chatbot and AI-based analysis of open-ended answers to process these candidates.

[00:10:58] For just one internship program, recruiters reported that this AI-integrated approach saved them 200 hours of manual labor. Okay, but… And crucially, it produced their most diverse cohort to date. If they had relied on high-friction human assessment from the jump, the system would have collapsed under its own weight. The AI didn't just save time. It allowed them to cast a wider net and actually evaluate people who would have otherwise been completely lost in the pile. You bring up L'Oreal saving 200 hours by using a chatbot. Sure, that saves manual labor.

[00:11:27] But when an automated system is completely blind to the physical human on the other end, it opens a massive, unprecedented attack vector for fraud. Fraud? Yes. Let's go back to our submarine analogy. If your ship is entirely reliant on automated systems to process incoming signals without visual confirmation, you are going to get hit by an enemy vessel synthesizing an allied ping. Okay, I follow. Let me walk you through the KnowBefore case from 2024, because it perfectly illustrates the catastrophic security risk of the highly scalable,

[00:11:56] remote, AI-mediated hiring model you are defending. KnowBefore is a cybersecurity training company, right? Yes. Security is literally their entire business model. They hired a remote software engineer. This individual went through the standard automated systems. They passed the background checks. They verified references. Wow. They even sat through four separate video interviews. The automated checks gave the green light, but it wasn't a real candidate. It was a North Korean operative. That is wild. Right?

[00:12:25] They used a stolen U.S. identity and AI-enhanced deepfake technology to alter their appearance in real time. KnowBefore shipped this quote-unquote employee a company laptop, and almost immediately the operative attempted to load malware onto the corporate systems. Oh, man. This isn't just a candidate exaggerating their Excel skills on a resume to get an interview. This is organized, malicious deception. The only way to stop this kind of sophisticated, AI-enabled fraud is proportionate, rigorous identity verification.

[00:12:55] Physical, in-person touch points where possible, and a dramatic reduction in our reliance on automated trust. The actual AI filters are a security nightmare. No, it's a terrifying case, and I absolutely acknowledge the severe threat of organized fraud. We saw the Department of Justice crack down heavily on these exact schemes in 2025. They raided 29 laptop farms across 16 states that were facilitating these North Korean IT worker operations. Exactly. It is a coordinated criminal enterprise. It is.

[00:13:24] But listen closely to your own example. You just said the operative sat through four separate video interviews. Four. Yeah. Human interviewers are just as easily fooled by deepfakes if they are unaided by technology. The know-before operatives passed the human test. A human recruiter cannot detect the micro-fluctuations in a video stream that indicate a real-time face overlay. Only another AI can do that. Relying purely on human intuition to spot a deepfake is a losing battle.

[00:13:51] But the failure there was trusting the digital medium entirely. If they had required an in-person assessment or a highly structured life environment, the fraud would have failed immediately. You cannot demand in-person assessments for a global, remote-first workforce. But more importantly, focusing entirely on candidates tricking the system ignores the other half of the equation. Which is? We also have to look at the immense damage caused by employers' flawed, unmonitored use of AI screening. You're worried about candidates faking a signal.

[00:14:20] I'm worried about the automated radar system itself being fundamentally broken. Hold on. Those are two entirely different categories of risk. A targeted nation-state cybersecurity threat is a malicious attack. An employer's algorithmic bias is a flaw in software design. They are both profound failures of the hiring system's predictive validity. To use your analogy, if candidate fraud is an enemy vessel faking a ping, algorithmic bias is a corrupted sonar system built right into your own ship's radar.

[00:14:48] It lies to you about what's out there. Okay. Fair point. And when employers use automated systems without oversight, the scale of the damage is astronomical. Look at the pending Mobley vs. Workday litigation. An applicant alleged that the automated screening software discriminatorily rejected him from over 100 jobs based on race, age, and disability. Or look at Amazon's infamous 2014 resume screening tool. Do you know how machine learning trains on past data?

[00:15:16] Well, it looks for patterns in historical success. Exactly. It looks at 10 years of successful resumes. In Amazon's case, those historical resumes were heavily dominated by men. The AI system literally taught itself to penalize resumes that included the word women's, like women's chess club captain. Wow. Because it learned the historical bias of the human recruiters and automated it at scale. It was throwing out top talent just because they didn't fit a historical demographic profile. That is terrible predictive validity.

[00:15:45] But the solution to both candidate fraud and employer bias isn't throwing out the AI. The solution is strict algorithmic governance. I mean, governance is a nice buzzword. But how does it actually fix the underlying problem of signal loss? It fixes it through mandatory transparency and testing. We need rigorous enforcement of policies like New York City's Local Law 144, which requires independent audits for automated employment decision tools to ensure they aren't generating disparate impacts. Right.

[00:16:15] We need transparent policies like Anthropics' clear disclosure rules, which explicitly tell candidates how and when AI is being evaluated. Robust governance, you know, testing the AI, auditing the algorithms, defining the parameters. It protects candidates from systemic failures far better than relying on the subjective, historically flawed intuition of individual recruiters trying to manually sift through thousands of applications. Okay. Governance is absolutely essential. I don't disagree with you there.

[00:16:43] Any tool you use needs to be audited. But governance alone doesn't restore the fundamental signal. How doesn't it? Think about the mechanics of what you are proposing. If an AI screening tool is rigorously audited to ensure it has no demographic bias, that's wonderful. But if that tool is still just analyzing cover letters and resumes that were entirely generated by other AI tools, what are we actually doing? We are just governing a feedback loop of synthetic noise.

[00:17:09] We are ensuring that our automated systems are fairly and equitably reading the automated outputs of other systems. That doesn't predict job performance. It just fulfills a compliance checklist. To actually predict if a human being can do the job, we have to measure human behavior. So back to the live assessment. Yes. We must evaluate a candidate's actual ability to do the work, which means bringing them out from behind the algorithm. See, I completely agree that we are stuck in a feedback loop.

[00:17:36] This arms race where candidates optimize for the algorithms and employers escalate their automated filters to handle the output is highly destructive. Highly. And the data bears out exactly how alienating this is to the candidate experience. Gartner found that only 26% of candidates currently trust AI to evaluate them fairly. It really feels like shouting into a void. Absolutely. But I maintain that the path forward isn't pretending we can put the AI genie back in the bottle. Generative AI is here. It is a fundamental part of knowledge work now.

[00:18:05] We have to design systems that assume AI assistance is the new baseline and build transparent, well-governed frameworks that measure how well candidates solve problems with these tools at their disposal. And I maintain that restoring trust and genuine predictive validity requires us to prioritize high-validity, human-centric structured assessments. We need stringent identity verification and evidence-based measurement to escape the automated noise. The traditional human signals certainly had their flaws.

[00:18:32] But abandoning human assessment for algorithmic convenience just to manage inbox volume is a cure far worse than the disease. But it seems we absolutely agree that evidence-based measurement and strict continuous oversight of whatever systems we use are non-negotiable moving forward. The arbitrary metrics of the past just cannot survive this technological shift. Without question. It is an incredibly complex challenge.

[00:18:55] How do you design a fair, highly predictive hiring system in an era where digital identity, text, and even video are infinitely malleable? It's a huge question. It really is. The source material in Dr. Westover's paper outlines a rich, multilayered framework for building long-term hiring integrity. And there is so much more nuance in there to explore. Ultimately, the market is going to decide which approach yields the strongest teams.

[00:19:19] But for now, we leave you, the listener, to look out into that dark water, listening to the echoes, and decide for yourself which signal you can truly trust. We'll see you next time. We'll see you next time. Bye. Bye.