Signal Lost: The AI Hiring Arms Race, the Applicant Flood, and How Leaders Can Restore Trust and Validity in Selection
Download the article here:
Listen to a review of this article:
Abstract: Generative artificial intelligence has changed both sides of the hiring process at once. Candidates now use AI to write, tailor, and submit applications at scale, while employers increasingly use AI to screen the resulting volume. Workday's October 2026 Global Workforce Report found that 84% of job seekers used AI in their search and that the median number of applicants per filled job rose from 58 to 69 in a year, with no improvement in time to fill. This article examines the resulting arms race: a flood of polished but less informative applications, rising candidate fraud, falling trust on both sides, and growing legal exposure for automated screening. Drawing on personnel selection research, labor economics, algorithmic fairness studies, and regulatory developments, it explains why traditional application signals are losing value. It then reviews five evidence-based responses: redesigning the front door around skills, shifting weight to high-validity assessment, verifying identity proportionately, governing AI screening tools, and adopting transparent AI policies for candidates. It concludes with a framework for treating hiring as an evidence-based measurement system.
Ask any recruiter what the past two years have felt like, and you are likely to hear a version of the same story. Job postings that once drew a manageable stream of applicants now draw a torrent. Cover letters that once revealed something about a person now read as if they were written by the same articulate stranger. Video interviews that once offered a window into how someone thinks now come with a nagging question: is this really the candidate, and are these really their answers?
Candidates tell a mirror-image story. They submit dozens or hundreds of applications and hear nothing back. They suspect, often correctly, that an algorithm rejected them before a person ever looked. Many have encountered job postings that turned out to be fake. Some have concluded that the only rational response is to use AI themselves, to apply faster, to tailor more, and to get past the filters.
This is the AI hiring arms race. In October 2026, Workday's Global Workforce Report put numbers on it. Drawing on its AI@Work Pulse survey of 5,944 workers, Workday found that 84% of job seekers had used AI during their search. Using de-identified data from its customers, it found that the median number of applicants for each filled job rose to 69, up from 58 a year earlier, while the time to fill a role held steady at about 60 days (Workday, 2026). More applications have not produced faster hiring. In financial services, applicants per filled job rose 27%, and in technology and media they rose 40%. In those sectors, 53% and 57% of applicants respectively said that AI had increased how many roles they applied to (Workday, 2026).
The volume problem is only the most visible symptom. Underneath it lies a deeper challenge that personnel researchers would recognize immediately: a loss of signal. Hiring works only when the information employers collect about candidates actually predicts how they will perform. When AI makes it cheap for anyone to produce a polished, tailored application, the polish stops telling employers much about the person (Galdin & Silbert, 2025). At the same time, AI-enabled fraud, from scripted answers to deepfake video, threatens the integrity of interviews themselves (Gartner, 2025; Greenhouse, 2025).
Employers have responded, in part, by deploying their own AI to sort through the flood. That creates a new set of risks. Research has repeatedly found that AI screening tools can reproduce or amplify bias (Wilson & Caliskan, 2024; Armstrong et al., 2024), and the legal environment is tightening, from collective litigation in the United States to new state laws and the European Union's AI Act (Mobley v. Workday, Inc., 2025; Ogletree Deakins, 2026). Trust is eroding on both sides. Only 26% of candidates trust that AI will evaluate them fairly (Gartner, 2025).
This article argues that the way out of the arms race is not more automation on either side. It is a return to the fundamentals of evidence-based selection, adapted for a world where text is cheap and identity can be faked. It first defines the key terms and maps the scale and drivers of the problem. It then examines the consequences for organizations and individuals, reviews five evidence-based responses with named organizational examples, and closes with a framework for building long-term hiring capability.
The AI Hiring Arms Race Landscape
Defining Signal, Noise, and the Arms Race
A few terms help clarify what is happening.
In labor economics, a signal is an observable action or credential that conveys information about a candidate's hidden qualities, such as ability or motivation, because it is costly or difficult to produce for people who lack those qualities. A thoughtful, tailored cover letter has traditionally functioned as a signal because writing one took time and skill. When generative AI makes such letters nearly costless to produce, they lose much of their informational value. Economists describe this as the collapse of a signaling equilibrium.
Noise refers to information that varies across candidates but does not predict performance. Generic applications, keyword-stuffed resumes, and AI-polished text that looks the same across very different people all add noise.
Validity, in personnel selection, refers to how well a selection method predicts later job performance. It is usually expressed as a correlation between scores on the method and measures of performance. Decades of research have established that some methods, such as structured interviews, are considerably more valid than others (Schmidt & Hunter, 1998; Sackett et al., 2022).
Candidate fraud covers a range of deceptive practices, from misrepresenting credentials to having another person take an interview, to using AI-generated scripts, voice alteration, or deepfake video. Prompt injection in this context refers to hidden text that candidates embed in a resume, such as white text on a white background, to manipulate an AI screening tool.
Finally, an automated employment decision tool is software that uses machine learning, statistical modeling, or AI to substantially assist or replace human judgment in hiring decisions. This is the term used in New York City's Local Law 144, and related concepts appear in Illinois, California, Colorado, and EU law.
The arms race describes the feedback loop that connects these elements. Candidates use AI to increase volume and polish. Employers use AI to filter the increased volume. Candidates adapt to the filters, sometimes deceptively. Employers add more filters and more suspicion. Each side's rational response to the other makes the overall system less informative and less trusted.
Prevalence, Drivers, and Distribution
The scale of AI use among candidates is now very high. Beyond Workday's finding that 84% of job seekers used AI in their search (Workday, 2026), Greenhouse's 2025 survey of job seekers and hiring professionals in the United States, the United Kingdom, Ireland, and Germany found that 74% of U.S. job seekers personally used AI and that 49% had submitted more applications than a year earlier (Greenhouse, 2025). Gartner's survey data showed that 39% of candidates used AI during the application process in late 2024 (Gartner, 2025). LinkedIn was reported in mid-2025 to be receiving an average of 11,000 applications per minute, up 45% over the prior year (Jackson, 2025).
Several drivers stand out.
Near-zero marginal cost of applying. When a tool can generate a tailored resume and cover letter in seconds, the cost of one more application approaches zero. The rational strategy for a candidate is to apply widely.
A slower labor market. Workday's report describes a market in which employees are staying put and internal moves are declining (Workday, 2026). When opportunities feel scarce, candidates apply to more of them.
Employer automation. Because many employers already use automated screening, candidates believe they need to optimize for algorithms. Greenhouse found that 41% of U.S. job seekers admitted using prompt injections to get past AI filters (Greenhouse, 2025).
Fraud as an organized activity. Some candidate fraud is not individual opportunism but coordinated activity. The U.S. Department of Justice announced in June 2025 coordinated actions against North Korean remote IT worker schemes, including searches of 29 known or suspected laptop farms across 16 states, and reported that more than 100 U.S. companies had been affected (U.S. Department of Justice, 2025). Gartner has predicted that by 2028, one in four candidate profiles worldwide will be fake (Gartner, 2025).
It is worth emphasizing that neither side is behaving irrationally. From an individual candidate's point of view, using AI to apply to more jobs is a sensible response to low response rates and automated filters. From an individual employer's point of view, using AI to screen a growing pile of applications is a sensible response to limited recruiter capacity. The problem is that these individually sensible choices interact. Each additional AI-generated application makes every application slightly less informative, and each additional automated filter gives candidates more reason to optimize for the filter rather than for the job. This is why the arms race framing is useful: the solution has to change the structure of the interaction, not just ask one side to stop.
The burden is not evenly distributed. Workday's sector data show that the surge is sharpest in technology and media and in financial services (Workday, 2026), sectors with many remote-eligible, knowledge-work roles. Remote roles are particularly exposed to identity fraud because the entire process may happen through a screen. Recruiters bear much of the operational burden. Greenhouse found that 34% of U.S. recruiters spend up to half of their week filtering spam and junk applications (Greenhouse, 2025).
Deception is now widespread enough that most recruiters have encountered it. Greenhouse reported that 91% of U.S. recruiters had spotted candidate deception, and that 65% had caught applicants using AI deceptively, including AI-generated scripts (32%), prompt injections (22%), and deepfakes (18%). Meanwhile, 36% of U.S. job seekers said they had used AI to alter their appearance, voice, or background in video interviews (Greenhouse, 2025). Not all of these alterations are deceptive in intent, since background blurring is common, but the figure illustrates how blurred the line between assistance and misrepresentation has become.
Gartner's candidate survey found that 6% of candidates admitted to interview fraud, either posing as someone else or having someone pose as them (Gartner, 2025). That share may sound small, but in a high-volume process, it represents a meaningful number of fraudulent candidates reaching the interview stage.
Organizational and Individual Consequences
Organizational Performance Impacts
The most important organizational consequence of the arms race is a decline in the quality of hiring decisions, driven by the loss of signal. Galdin and Silbert (2025) provided some of the most direct evidence. Studying a large online freelance marketplace, they used a language model to measure how customized each application was to the job. Before the arrival of large language models, employers were willing to pay more for workers who sent customized applications, because customization signaled effort and fit. After large language models became widely available, that premium disappeared. In the authors' model, the loss of this signal made the market less meritocratic: workers in the top ability quintile were hired 19% less often, and workers in the bottom quintile were hired 14% more often (Galdin & Silbert, 2025). These are counterfactual estimates from one marketplace, but they illustrate the mechanism clearly. When polish no longer distinguishes candidates, employers lose the ability to identify the best ones from written materials alone.
The picture is more nuanced than "AI writing is bad." Wiles, Munyikwa, and Horton (2023) ran a field experiment with nearly half a million jobseekers on an online labor market and found that those given algorithmic help with their resume writing were hired 8% more often, with no evidence that employers were less satisfied with the resulting hires. Their interpretation was that clearer writing helped employers see candidates' real qualifications. Taken together, the two studies suggest that AI assistance can help individual candidates present themselves better, but that when everyone uses it, written materials lose their power to differentiate.
A second consequence is operational cost. When applicant volume rises without any improvement in time to fill (Workday, 2026), the cost per hire rises. Recruiter time is diverted from engaging promising candidates to filtering noise (Greenhouse, 2025).
A third consequence is security risk. Fraudulent hires are not merely poor performers. In some cases they are a direct threat. The KnowBe4 case, discussed below, shows how a fraudulent remote hire can attempt to load malware on a company device within days of starting (Bracken, 2024). The Department of Justice reported that its North Korean IT worker investigations involved more than 80 U.S. persons' identities being compromised, and at least $3 million in damages in one case (U.S. Department of Justice, 2025).
A fourth consequence is legal exposure. Automated screening tools that produce disparate impacts by age, race, sex, or disability can expose both employers and vendors to litigation, as the Workday and iTutorGroup cases illustrate (Mobley v. Workday, Inc., 2025; Duane Morris LLP, 2023). Research continues to find bias in AI resume screening. Wilson and Caliskan (2024) tested embedding-based resume screening across nine occupations and found that it favored White-associated names in 85.1% of cases and female-associated names in only 11.1% of cases. A Bloomberg News analysis found that GPT-3.5 ranked resumes with names distinctive to Black Americans as the top candidate less often than other groups for some roles (Bloomberg News, 2024). Armstrong and colleagues (2024) found that GPT-3.5 showed stereotype-based biases when scoring resumes and, when generating resumes, gave women less experience.
Individual Wellbeing and Stakeholder Impacts
For candidates, the arms race is exhausting and demoralizing. Greenhouse found that 46% of U.S. job seekers said their trust in hiring had decreased, with 42% blaming AI directly. Only 8% said AI makes hiring fairer, and 35% said AI had simply shifted bias from humans to algorithms. Some 69% had encountered fake job postings (Greenhouse, 2025). HR Dive, reporting on Gartner's research, noted that only half of candidates believed the jobs they applied for were legitimate (Crist, 2025).
The trust gap is striking. Greenhouse found that 70% of hiring managers trust AI to make faster and better hiring decisions, while only 8% of job seekers say it makes hiring fairer (Greenhouse, 2025).
Research on applicant reactions helps explain why this matters for employers. Oostrom and colleagues (2024) found across two studies that applicants who were told an algorithm had evaluated them felt more "creepiness" and rated the process as less fair and less valid than applicants told a recruiter had evaluated them. Through lower perceived fairness, algorithmic evaluation indirectly reduced organizational attractiveness and intentions to accept a job offer. In other words, the way a process feels to candidates affects whether the best ones will want to join.
The burden also falls unequally. If AI screening tools carry demographic biases (Wilson & Caliskan, 2024; Bloomberg News, 2024), then the candidates most likely to be screened out unfairly are those already facing barriers. Candidates who cannot afford premium AI tools, or who do not know how to game filters, may be disadvantaged relative to those who do. Paradoxically, Galdin and Silbert's (2025) model suggests that the loss of signal can also harm the strongest candidates, whose genuine effort and skill become harder to distinguish.
Privacy is another concern. When AI hiring systems collect extensive personal data, including interview transcripts, a security failure can expose millions of people. The McHire incident, in which a default password exposed up to 64 million application records, illustrates the scale of what is at stake (INCIBE-CERT, 2025).
For recruiters and hiring managers, the arms race brings its own strain. Filtering junk applications, second-guessing every interview, and worrying about deepfakes are draining tasks. Greenhouse found that 74% of U.S. hiring professionals were more concerned about fake credentials and deepfakes than a year earlier (Greenhouse, 2025).
Candidates with disabilities face particular risks. Automated assessments, timed games, and video analysis may disadvantage people whose disabilities affect speech, movement, or processing speed, even when those differences are unrelated to job performance. California's regulations specifically note that automated assessments that elicit information about disability may amount to an unlawful medical inquiry (California Civil Rights Council, 2025). Accessible design and a clear accommodation process are therefore part of fairness, not an afterthought.
Finally, the broader labor market suffers. When hiring becomes slower, noisier, and less trusted, good matches between people and jobs become harder to make. That has costs for productivity, for workers' careers, and for the economy.
Evidence-Based Organizational Responses
The arms race cannot be won by escalating it. Employers that respond to AI-generated applications with ever more aggressive AI filtering may reduce their workload, but they also risk filtering out good candidates, embedding bias, and further eroding trust. The evidence points instead to five responses that restore signal, protect integrity, and rebuild trust.
Response | Evidence Base | Key Practices | Illustrative Examples |
Redesign the front door around skills | Skills-based hiring research; labor market signaling research | Clear, skills-based requirements; removal of unnecessary degree filters; realistic job previews | IBM; Walmart; State of Maryland |
Shift weight to high-validity, AI-resilient assessment | Personnel selection meta-analyses; structured interview research | Structured interviews, work samples, job knowledge tests, at least one live or in-person stage | Google; Goldman Sachs |
Verify identity and authenticity proportionately | Fraud surveys; law enforcement findings | Risk-based identity checks, device and onboarding controls, recruiter training | KnowBe4; U.S. Department of Justice actions |
Govern and audit AI screening tools | Algorithmic bias research; litigation and regulation | Bias audits, vendor due diligence, human review, data security, recordkeeping | Amazon; Workday; iTutorGroup; HireVue; McDonald's |
Adopt transparent, candidate-respecting AI policies | Applicant reactions research; candidate trust surveys | Clear disclosure of employer AI use, clear rules for candidate AI use, timely communication | Anthropic; Goldman Sachs; Chipotle; L'Oréal |
Redesign the Front Door Around Skills
The first response addresses the source of the flood: vague job requirements and application processes that reward volume. When job postings list generic requirements and a degree filter, they invite generic applications and push employers toward proxies that AI can easily mimic. Clear, skills-based requirements give both candidates and screeners something more meaningful to work with.
Skills-based hiring has been widely promoted but less widely practiced. Research by the Burning Glass Institute and Harvard Business School found that despite many announcements, skills-based hiring accounted for fewer than 1 in 700 hires in the year studied. Only 37% of firms that removed degree requirements drove nearly all of the actual change in hiring; about 45% changed in name only, and roughly one-fifth were backsliders. Among the leaders, however, the results were promising: non-degree hires had retention rates 10 percentage points higher than degree-holding colleagues, and workers without degrees who moved into roles that had previously required them saw an average salary increase of 25% (Sigelman et al., 2024).
The relevance to the arms race is direct. A front door built around demonstrable skills makes polished prose less important and evidence of capability more important. It also expands the talent pool in ways that reduce reliance on easily gamed credentials.
Effective approaches:
Skills-based postings: Describe the specific skills and tasks the role requires, and remove degree requirements where they are not genuinely necessary.
Realistic job previews: Show candidates what the work actually involves so that those who are not a fit self-select out.
Short, specific screening questions: Ask a small number of job-specific questions that require concrete examples, rather than open-ended essays that AI can generate easily.
Follow-through in practice: Change how screeners and hiring managers make decisions, not just what the posting says, to avoid change in name only.
Track outcomes: Measure the retention and performance of hires from different pathways to confirm the approach is working.
IBM stated in 2021 that it had removed bachelor's degree requirements for more than half of its U.S. job openings. An analysis by the Foundation for Research on Equal Opportunity found that only 31% of IBM's software developer and engineer postings requested a degree, compared with 84% at Hewlett-Packard (Cooper, 2022).
Walmart announced in 2023 that it was rewriting job descriptions for some corporate roles to accept either a degree or equivalent skills, noting that a degree is simply unnecessary for many roles, including at its headquarters (Tornone, 2023). In the public sector, the State of Maryland announced in 2022 that more than half of the positions in its state workforce of more than 38,000 could substitute experience, training, or community college for a four-year degree (KATV, 2022).
Common pitfall: Removing degree requirements from postings without changing how candidates are actually evaluated. The Burning Glass research shows that many firms changed their postings but not their hiring decisions (Sigelman et al., 2024). Skills-based hiring requires skills-based assessment, which leads to the second response.
Shift Weight to High-Validity, AI-Resilient Assessment
If written materials have lost much of their signal, employers need to rely more on methods that remain predictive and are harder to fake with AI. Personnel selection research offers clear guidance on which methods work. Sackett, Zhang, Berry, and Lievens (2022) revisited decades of meta-analytic evidence and concluded that earlier estimates had overstated the validity of several methods because of how they corrected for range restriction. In their revised estimates, structured interviews emerged as the top-ranked predictor of job performance, followed by job knowledge tests, empirically keyed biographical data, work sample tests, and cognitive ability tests.
Structured interviews, in which all candidates are asked the same job-related questions and rated against predefined criteria, have been studied extensively. Levashina and colleagues (2014) reviewed roughly two decades of research on how interview structure shapes outcomes, including bias, impression management, question types, rating scales, and applicant reactions, and offered practical propositions for designing structured interviews. Structured interviews and work samples also have a practical advantage in the AI era: when conducted live, they require candidates to demonstrate their thinking in real time, which makes AI-generated answers less useful.
Effective approaches:
Structured interviews: Use standardized, job-related questions with behaviorally anchored rating scales, and train interviewers to use them consistently.
Work samples: Ask candidates to perform a realistic task from the job, ideally in a live or proctored setting, and allow them to explain their reasoning.
Job knowledge tests: For roles that require specific expertise, use well-designed tests of that knowledge.
At least one live stage: Include at least one synchronous interaction, in person where practical, especially for remote and high-risk roles.
Rethink take-home assignments: Either accept that candidates may use AI and evaluate how well they use it, or move the task to a live setting.
Google's chief executive, Sundar Pichai, said in mid-2025 that the company would introduce at least one round of in-person interviews, in response to concerns about AI-assisted cheating in virtual interviews (CXO Digital Pulse, 2025). The move reflects a broader shift. Greenhouse found that 39% of U.S. hiring professionals were doing more in-person interviews to verify authenticity (Greenhouse, 2025), and Gartner found that 62% of candidates said they would be more likely to apply if in-person interviews were required (Gartner, 2025).
Goldman Sachs informed campus candidates in its Europe, Middle East, and Africa recruiting process that it prohibits the use of external sources, including ChatGPT and search engines, during interviews. A spokesperson explained that the firm wanted to hear from applicants in their own voice (Burleigh, 2025b).
Common pitfall: Adding more interview rounds without adding more structure. Unstructured conversations, however many there are, remain less valid than structured ones (Sackett et al., 2022). The goal is better measurement, not simply more contact.
Verify Identity and Authenticity Proportionately
The third response addresses fraud directly. As remote hiring and AI-generated media make it easier to misrepresent identity, employers need verification practices that are proportionate to the risk of each role. A junior retail position and a remote software engineering role with access to production systems present very different risks.
The evidence on the scale of fraud comes from several sources. Gartner found that 6% of candidates admitted interview fraud (Gartner, 2025). Greenhouse found that 18% of U.S. recruiters who caught deceptive AI use had encountered deepfakes (Greenhouse, 2025). The Department of Justice's actions against North Korean IT worker schemes revealed laptop farms, stolen identities, and more than 100 affected companies (U.S. Department of Justice, 2025).
Effective approaches:
Risk-based tiers: Match the intensity of verification to the role's access to systems, data, money, and people.
Identity checks before access: Verify identity documents and match them to the person who appears at interviews and onboarding, before granting system access.
Device and onboarding controls: Ship equipment only to verified addresses, monitor new devices for unusual activity, and limit early access.
Recruiter training: Teach recruiters and interviewers to recognize red flags, such as inconsistent backgrounds, reluctance to appear on camera, or mismatched details.
Respectful process: Explain why verification is needed so that legitimate candidates do not feel accused.
KnowBe4, a cybersecurity awareness training company, disclosed in 2024 that it had hired a remote software engineer who turned out to be a North Korean operative. The person had passed background checks, verified references, and four video interviews, using a stolen U.S. identity and an AI-enhanced stock photo. Shortly after receiving a company laptop, the new hire attempted to load malware. KnowBe4's security operations center detected and contained the activity, and the company reported that no data was lost. Afterward, the company said it was strengthening identity verification before start dates and training its hiring staff on warning signs (Bracken, 2024). The case is notable precisely because KnowBe4 is a security-focused firm with a standard hiring process.
The U.S. Department of Justice's coordinated actions in 2025 involved searches of laptop farms across 16 states, the seizure of financial accounts and websites, and charges in multiple districts (U.S. Department of Justice, 2025). For employers, the lesson is that remote hiring fraud is not hypothetical and is not limited to small firms.
Common pitfall: Applying maximum suspicion to every candidate. Heavy-handed verification for low-risk roles adds friction, signals distrust, and can deter good candidates. Proportionality protects both security and the candidate experience.
Govern and Audit AI Screening Tools
The fourth response concerns employers' own use of AI. Automated screening can be a reasonable response to volume, but only if it is valid, fair, secure, and lawful. The evidence on bias reviewed above (Wilson & Caliskan, 2024; Armstrong et al., 2024; Bloomberg News, 2024) means that employers cannot assume a tool is fair because a vendor says so. Raghavan and colleagues (2020) examined what vendors of algorithmic pre-employment assessments publicly disclosed about how they detect and mitigate bias and validate their tools, and found that the technical choices involved and their relationship to anti-discrimination law deserve close scrutiny from employers.
The legal environment is moving in the same direction, though unevenly. New York City's Local Law 144 requires bias audits and notices for automated employment decision tools, but a December 2025 audit by the New York State Comptroller found weak enforcement. When the city agency reviewed 32 employer and vendor websites, it identified one likely instance of non-compliance. State auditors reviewing the same 32 found 17 (Office of the New York State Comptroller, 2025). Illinois amended its Human Rights Act, effective January 1, 2026, to prohibit the use of AI that has the effect of discriminating in employment decisions and to require notice to employees (Epstein Becker Green, 2024). California's Civil Rights Council regulations, effective October 1, 2025, clarify that automated-decision systems can violate the state's Fair Employment and Housing Act and require employers to retain related records for at least four years (California Civil Rights Council, 2025). Colorado's original AI law was blocked from enforcement in April 2026 and replaced in May 2026 by a narrower, notice-based framework, with a new effective date of January 1, 2027 (McDermott Will & Schulte, 2026). In the European Union, a 2026 agreement delayed the AI Act's high-risk obligations for employment systems from August 2026 to December 2027, but employment AI remains classified as high-risk (Ogletree Deakins, 2026).
Effective approaches:
Validation evidence: Require evidence that a tool predicts job performance for the roles where it is used, not just that it is efficient.
Bias audits: Test tools for adverse impact across protected groups before deployment and at regular intervals, including independent audits where required.
Human review: Ensure that people, not algorithms alone, make final decisions, and that candidates can request human review.
Vendor due diligence: Contract for transparency, audit access, data security standards, and shared responsibility.
Data security and retention: Apply strong security controls to hiring data and retain records as the law requires.
Defensive design against manipulation: Test tools against prompt injection and other manipulation, and avoid tools that reward keyword stuffing.
Amazon offers the best-known cautionary tale. Beginning in 2014, the company built an experimental tool to score resumes, trained on a decade of resumes submitted mostly by men. The tool learned to penalize resumes that included the word "women's" and to downgrade graduates of two all-women's colleges. Amazon edited the tool to remove those specific effects but could not guarantee it would not find other ways to discriminate, and the team was disbanded by early 2017 (Dastin, 2018).
The Workday litigation shows how the legal risk has spread from employers to vendors. In Mobley v. Workday, a job applicant alleged that Workday's screening software contributed to his rejection from more than 100 jobs on the basis of race, age, and disability. In May 2025, a federal court in California preliminarily certified a nationwide collective under the Age Discrimination in Employment Act, and in July 2025 it held that the collective includes applicants scored or ranked by certain AI features (Mobley v. Workday, Inc., 2025). The case remains pending. In online education, the U.S. Equal Employment Opportunity Commission settled a lawsuit with iTutorGroup in 2023, which had used software that automatically rejected female applicants aged 55 or older and male applicants aged 60 or older. The settlement included $365,000 for affected applicants and a five-year consent decree requiring new policies, training, and monitoring (Duane Morris LLP, 2023).
Data security is part of governance too. In 2025, security researchers found that an administrative interface for McDonald's McHire platform, operated by the vendor Paradox.ai, was protected by a default password and lacked multifactor authentication. Combined with a separate flaw, the weakness exposed up to 64 million application records. Paradox disabled the account on the day it was notified, and McDonald's said it would tighten security requirements for its vendors (INCIBE-CERT, 2025).
Some vendors have changed course in response to evidence and criticism. HireVue, whose video interviewing tools are used by many large employers, stopped using facial analysis in its assessments, announcing the change in early 2021 after its internal research found that visual analysis no longer significantly added value (Maurer, 2021).
Common pitfall: Assuming that compliance with the narrowest applicable law is enough. Because laws differ by jurisdiction and are changing quickly, and because litigation can proceed under long-standing anti-discrimination statutes, employers are better served by a consistent governance standard that meets the strictest rules they face.
Adopt Transparent, Candidate-Respecting AI Policies
The fifth response addresses trust. Candidates want to know how AI is being used to evaluate them, and they want clear rules for how they themselves may use it. Greenhouse found that 87% of U.S. job seekers said it was important for employers to be open about their use of AI (Greenhouse, 2025). Gartner found that only 26% of candidates trusted AI to evaluate them fairly, and that 25% trusted an employer less when AI was used to evaluate them (Gartner, 2025). The applicant reactions research shows that these perceptions affect whether candidates are attracted to an organization and willing to accept an offer (Oostrom et al., 2024).
Clear policies also help employers. Without them, recruiters and hiring managers improvise, and candidates face inconsistent expectations. A policy that says when AI use is welcome, such as for polishing a resume, and when it is not, such as during a live assessment, reduces confusion and makes violations easier to identify fairly.
Effective approaches:
Disclose employer AI use: Tell candidates which stages involve AI, what it does, and how they can request human review or accommodation.
Set clear rules for candidate AI use: Specify where AI is acceptable and where it is not, and explain why.
Communicate promptly: Acknowledge applications, provide status updates, and close the loop, which reduces the feeling of applying into a void.
Use AI to improve the candidate experience: Apply automation to scheduling and communication, where it saves time without making judgments about people.
Monitor perceptions: Survey candidates, including those who are not hired, about the fairness and clarity of the process.
Anthropic, the AI company, asked applicants in its job postings not to use AI assistants during the application process, explaining that it wanted to understand candidates' personal interest without mediation through an AI system (Burleigh, 2025a). The policy is notable because it came from a company that builds AI, which underscores that the issue is about the purpose of each stage, not about AI as such. Amazon shared guidelines with its internal recruiters in 2025 indicating that candidates who used AI during interviews could be disqualified from the recruitment process (Burleigh, 2025b).
Some employers have used AI to improve the candidate experience rather than to judge candidates. Chipotle introduced an AI hiring assistant, named Ava Cado and built on the Paradox platform, that chats with candidates, collects information, schedules interviews, and sends offers in several languages. The company said the tool could reduce hiring time for restaurant roles by up to 75% (Kelso, 2024). L'Oréal, which receives about one million applicants a year for about 15,000 positions, has used a chatbot and AI-based analysis of open-ended answers in its recruiting. For one internship program, which drew about 12,000 applicants for about 80 places, recruiters said the approach saved 200 hours and produced the most diverse cohort to date (CNN Business, 2019). Both examples are company-reported, but they illustrate a design choice: automation aimed at speed, responsiveness, and access rather than at replacing human judgment about fit.
Accommodation belongs in any transparent policy. Candidates should know how to request an alternative to an automated assessment or an AI-led interview, and recruiters should know how to respond quickly and without stigma. This is both a legal safeguard and a signal of respect (California Civil Rights Council, 2025).
Common pitfall: Banning AI use by candidates while relying heavily on undisclosed AI screening. Candidates notice the asymmetry, and it erodes trust. Policies are more credible when employers hold themselves to the same standard of transparency they ask of candidates.
Building Long-Term Hiring Integrity Capability
The responses above address immediate problems. Building a durable advantage in an AI-saturated labor market requires a broader shift in how organizations think about hiring. Three pillars support that shift.
Treat Hiring as a Measurement System
At its core, hiring is a measurement problem: the organization is trying to predict how people will perform in roles they have not yet held. The AI arms race has exposed how much of that measurement relied on cheap proxies. A long-term capability means treating each stage of the process as a measurement instrument and asking what it actually measures.
That requires data. Organizations should link selection scores to later outcomes, such as performance ratings, retention, and promotion, and use that evidence to refine their methods. They should monitor adverse impact at each stage. They should also retire methods that add cost without adding predictive value, as HireVue did with facial analysis (Maurer, 2021). The meta-analytic evidence offers a strong starting point (Sackett et al., 2022), but each organization's roles and context are different, and local validation provides the most useful guidance.
Structured interview design offers a concrete example of what this looks like in practice. It begins with a job analysis that identifies the critical tasks and competencies of the role. Questions are then written to target those competencies, using past-behavior or situational formats, and rating scales are anchored with examples of weak, adequate, and strong answers. Interviewers are trained, panels calibrate their ratings, and the organization periodically checks whether interview scores relate to later performance and whether they show adverse impact (Levashina et al., 2014). None of this is new. What is new is that the alternatives, such as reading cover letters for signs of effort, have become far less informative, which raises the value of doing it well.
Build Verification and Security Into the Talent Lifecycle
Identity and authenticity cannot be checked once and forgotten. The KnowBe4 case shows that a fraudulent hire can pass every standard hiring check and reveal itself only after onboarding (Bracken, 2024). A long-term capability treats verification as a continuum that spans recruiting, onboarding, and the first months of employment.
This means close collaboration between talent acquisition, information security, and legal teams. It means risk-based policies that are consistent across business units. It also means treating hiring data as sensitive data, with security standards for vendors that match those applied to customer and financial data, as the McHire incident underscores (INCIBE-CERT, 2025).
Make Trust a Talent Asset
In a market where only a small share of candidates believe AI makes hiring fairer (Greenhouse, 2025), employers that earn candidates' trust have a competitive advantage. Trust is built through transparency about how decisions are made, consistency in how candidates are treated, respect for candidates' time, and fairness that can be demonstrated rather than merely asserted.
Trust also has an internal dimension. Recruiters and hiring managers need confidence in the tools and processes they use. When they understand what each stage measures and why, they are better able to make good decisions and explain them to candidates. Training, clear governance, and honest communication about the limits of AI tools all contribute.
Conclusion
The AI hiring arms race is a classic case of individually rational behavior producing a collectively worse outcome. Candidates use AI to apply more widely and present themselves more effectively. Employers use AI to manage the resulting volume. The result is more applications without faster hiring (Workday, 2026), application materials that reveal less about candidates (Galdin & Silbert, 2025), rising fraud (Gartner, 2025; U.S. Department of Justice, 2025), growing legal risk (Mobley v. Workday, Inc., 2025), and falling trust on both sides (Greenhouse, 2025).
The way forward is not to abandon AI or to escalate the arms race, but to return to the fundamentals of evidence-based selection. Structured interviews, work samples, and skills-based requirements remain highly predictive and are more resistant to AI manipulation than written materials (Sackett et al., 2022). Proportionate verification can counter fraud without treating every candidate as a suspect. Rigorous governance of AI tools can reduce bias and legal exposure. Transparent policies can rebuild trust.
Practical takeaways for leaders:
Fix the front door: Write skills-based job postings, remove unnecessary degree requirements, and use realistic job previews to reduce noise.
Measure what matters: Shift weight from written materials to structured interviews, work samples, and job knowledge tests.
Include a live stage: Build at least one synchronous or in-person interaction into the process, especially for remote and high-risk roles.
Verify proportionately: Match identity and authenticity checks to the risk of each role, and coordinate with security teams.
Govern your AI: Demand validation and bias audit evidence from vendors, keep humans in final decisions, and secure hiring data.
Be transparent: Disclose how AI is used in your process and set clear, fair rules for candidates' own AI use.
Close the loop: Link selection data to later outcomes, and keep improving the process based on what the evidence shows.
When text is cheap, evidence becomes valuable. The organizations that will hire best in the coming years are those that stop trying to out-automate their candidates and start measuring what actually predicts success.
Research Infographic
References
Armstrong, L., Liu, A., MacNeil, S., & Metaxa, D. (2024). The silicon ceiling: Auditing GPT's race and gender biases in hiring [Preprint]. arXiv.
Bloomberg News. (2024). OpenAI's GPT is a recruiter's dream tool. Tests show there's racial bias. Bloomberg.
Bracken, M. (2024, July 24). Cyber firm KnowBe4 hired a fake IT worker from North Korea. CyberScoop.
Burleigh, E. (2025a, February 4). AI company Anthropic's ironic warning to job candidates: 'Please do not use AI.' Fortune.
Burleigh, E. (2025b, June 11). Goldman Sachs wants students to stop using ChatGPT in job interviews with the bank. Fortune.
California Civil Rights Council. (2025, June 30). Civil Rights Council secures approval for regulations to protect against employment discrimination related to artificial intelligence [Press release].
CNN Business. (2019). [Report on L'Oréal's use of artificial intelligence in recruitment]. CNN.
Cooper, P. (2022, February 26). Employers are finally dropping college degree requirements. Foundation for Research on Equal Opportunity.
Crist, C. (2025, August 8). By 2028, 1 in 4 candidate profiles will be fake, Gartner predicts. HR Dive.
CXO Digital Pulse. (2025, August 26). Google reinstates in-person job interviews amid rising AI cheating concerns.
Dastin, J. (2018, October 10). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters.
Duane Morris LLP. (2023, August 11). EEOC settles its first discrimination lawsuit involving artificial intelligence hiring software. Duane Morris Class Action Defense Blog.
Epstein Becker Green. (2024, August 14). Illinois prohibits discriminatory artificial intelligence in employment decisions. Workforce Bulletin.
Galdin, A., & Silbert, J. (2025). Making talk cheap: Generative AI and labor market signaling (arXiv:2511.08785) [Preprint]. arXiv.
Gartner. (2025, July 31). Gartner survey shows just 26% of job applicants trust AI will fairly evaluate them [Press release].
Greenhouse. (2025, November 19). An AI trust crisis: 70% of hiring managers trust AI to make faster and better hiring decisions, only 8% of job seekers call it fair [Press release].
INCIBE-CERT. (2025, August 19). Security flaw in McDonald's AI recruitment system exposes data of millions of applicants. Spanish National Cybersecurity Institute.
Jackson, F. (2025, June 25). Job seekers, some using AI, flood LinkedIn with 11,000 applications a minute. eWeek.
KATV. (2022, March 16). Hogan cuts college degree requirement for thousands of Maryland jobs.
Kelso, A. (2024, October 22). Chipotle rolls out an AI platform to reduce hiring time. Nation's Restaurant News.
Levashina, J., Hartwell, C. J., Morgeson, F. P., & Campion, M. A. (2014). The structured employment interview: Narrative and quantitative review of the research literature. Personnel Psychology, 67(1), 241–293.
Maurer, R. (2021, February 3). HireVue discontinues facial analysis screening. SHRM.
McDermott Will & Schulte. (2026, May 27). Colorado AI law in flux: Comprehensive replacement bill signed after federal court blocks predecessor's enforcement.
Mobley v. Workday, Inc., No. 3:23-cv-00770 (N.D. Cal. 2025).
Office of the New York State Comptroller. (2025). Enforcement of Local Law 144: Automated employment decision tools [Audit report].
Ogletree Deakins. (2026, June 16). EU nears approval of agreement to delay rules for AI use in employment decisions.
Oostrom, J. K., Holtrop, D., Koutsoumpis, A., van Breda, W., Ghassemi, S., & de Vries, R. E. (2024). Applicant reactions to algorithm- versus recruiter-based evaluations of an asynchronous video interview and a personality inventory. Journal of Occupational and Organizational Psychology, 97(1), 160–189.
Raghavan, M., Barocas, S., Kleinberg, J., & Levy, K. (2020). Mitigating bias in algorithmic hiring: Evaluating claims and practices. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency Association for Computing Machinery.
Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2022). Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range. Journal of Applied Psychology, 107(11), 2040–2068.
Schmidt, F. L., & Hunter, J. E. (1998). The validity and utility of selection methods in personnel psychology: Practical and theoretical implications of 85 years of research findings. Psychological Bulletin, 124(2), 262–274.
Sigelman, M., Fuller, J., & Martin, A. (2024). Skills-based hiring: The long road from pronouncements to practice. Burning Glass Institute and Harvard Business School Project on Managing the Future of Work.
Tornone, K. (2023, September 29). Walmart scraps degree requirements for some corporate jobs. HR Dive.
U.S. Department of Justice. (2025, June 30). Justice Department announces coordinated, nationwide actions to combat North Korean remote information technology workers' illicit revenue generation schemes [Press release].
Wiles, E., Munyikwa, Z. T., & Horton, J. J. (2023). Algorithmic writing assistance on jobseekers' resumes increases hires (NBER Working Paper No. 30886). National Bureau of Economic Research.
Wilson, K., & Caliskan, A. (2024). Gender, race, and intersectional bias in resume screening via language model retrieval. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7(1), 1578–1590.
Workday. (2026, October 5). Workday Global Workforce Report: AI is rewriting jobs more than it's cutting them [Press release].
Jonathan H. Westover, PhD, Chief Research Officer (Nexus Institute for Work and AI); Co-Founder & Chief Workforce and Learning Officer (Future State University); Founder & CEO (Human Capital Innovations); Professor of Organizational Leadership & Change (UVU). Read Jonathan Westover's executive profile here.
Suggested Citation: Westover, J. H. (2026). Signal Lost: The AI Hiring Arms Race, the Applicant Flood, and How Leaders Can Restore Trust and Validity in Selection. Human Capital Leadership Review, 40(2). doi.org/10.70175/hclreview.2020.40.2.5
