Inside Hiring's AI Doom Loop — and Why Detecting AI Resumes Is the Wrong Fight

In July 2026, Greenhouse CEO Daniel Chait gave the hiring industry a name for something it had been living through for over a year without a clean label. Job seekers, squeezed by a brutal market, were using AI to apply to dozens or hundreds of roles at once — some paying roughly $20 a month for mass-apply tools. Recruiters, buried under the resulting volume, were using their own AI tools to filter it back down. Neither side was winning. “We’ve called that the AI doom loop,” Chait told reporters. “Everyone’s using their own AI to solve their own problem, but it’s making the whole system worse” (Fortune; HR Dive). As Chait put it to Fortune, “This is the first time when really both sides have been unhappy.”

Three months later, the survey data is catching up to the anecdote, and it confirms the loop is real — but it also points to a quieter finding underneath it: the industry’s default response, trying to detect which applications are AI-written, mostly doesn’t work. The more durable fixes being tried right now, on both the vendor side and ours, aren’t detection at all. They’re changes to who gets a say in the decision, and when.

The numbers behind the loop

Two separate Robert Half surveys, run six months and a continent apart, landed on the same figure. In March 2026, a US release reported that 67% of HR leaders say reviewing AI-generated applications has slowed their hiring process, with one in five (20%) reporting delays of more than two weeks; 65% said the surge has made it harder to verify candidates’ actual skills (Robert Half). In September, a separate UK survey found the identical 67% figure for hiring managers who say a surge in AI-generated applications has lengthened hiring, with 48% calling it a notable challenge for their organization. Their two leading complaints were an increased volume of applications from unqualified candidates (28%) and difficulty telling AI-written CVs apart from genuine ones (27%) (Staffing Industry Analysts). Two different countries, two different survey waves, the same number — that kind of convergence, from a firm with no obvious reason to inflate it in either direction, is a reasonable signal that “applications are slower to process because of AI volume” is now an ordinary, widely shared experience for hiring teams rather than an outlier complaint.

Candidates, for their part, aren’t hiding that this is a deliberate response rather than an accident. Greenhouse’s own July 2026 reporting found that two-thirds of US job seekers say they’re struggling to land a role, which is the pressure driving the mass-apply tooling Chait described (Fortune). The logic is straightforward and, from an individual candidate’s point of view, rational: if you believe your application is being screened by a model rather than read by a person, spending less effort per application and applying to more of them is the correct response to that belief — even though it’s exactly the behavior that makes the screening problem worse for everyone downstream.

The detection fight is the wrong fight

The instinctive employer-side countermeasure is to try to catch which applications are AI-generated. The evidence that this works at the software layer is thin. Independent testing by two resume-tooling vendors, StylingCV and Jobscan, each checked major applicant tracking platforms (including Workday, Greenhouse, and Taleo) and found none of them flag AI-generated resume text as part of their scoring (StylingCV: stylingcv.com; Jobscan: jobscan.co). That’s consistent with how these systems were built: an ATS parses and ranks text against keyword and fit criteria, which is a different job from classifying who or what wrote it, and a well-edited AI draft and a well-written human draft converge on the same surface features an ATS is designed to reward.

Human reviewers are more confident than the software is capable — Resume Genius’s 2026 Hiring Insights Report, surveying 1,000 US hiring managers, found 80% believe they can spot an AI-written resume at a glance, citing unnatural phrasing, repetitive language, and vague or inflated descriptions as the tells, and 19% say they use AI themselves to screen out applications before any human reviews them (Resume Genius; also reported via Black Enterprise). But self-reported confidence isn’t the same thing as a measured accuracy rate, and “I can usually tell” is a weaker foundation for an automated rejection than it is for a human’s gut check on one resume at a time. Put the two findings together and the picture is a text-authorship arms race where the detecting side has neither reliable software nor a validated human success rate — just the belief that it’s working, applied at a volume where nobody is checking the false-positive and false-negative rate against the people actually filtered out.

Tellingly, Greenhouse itself — the company that coined “doom loop” and runs an ATS used across thousands of employers — doesn’t appear to be betting on better detection either. Its own public fix is a volume-limiting feature called “My Dream Job,” letting a candidate flag one role per month as their actual top priority; Chait has said close to half a million “dream job” applications have been submitted since launch, with those candidates hired at roughly five times the rate of everyone else (HR Dive). That figure is self-reported by Greenhouse about its own product, so it’s worth treating as a claim rather than an independently audited result — but the design choice underneath it is the interesting part regardless of the exact multiplier: Greenhouse isn’t trying to classify which applications are AI-written. It’s trying to change the incentive that produces the flood in the first place, by making one application per candidate per month cost something (attention, not money) that mass-applying doesn’t.

What we do instead, and what we don’t claim

We don’t build or sell AI-authorship detection, and we want to be explicit that nothing below is a claim that NiceHire can tell which applications were AI-written, catch AI-assisted cheating, or verify candidate identity — none of those capabilities exist in our product today. What we can describe precisely is the structural shape of our own screening pipeline, because it takes a different bet than either “detect the fakes” or “limit the volume.”

NiceHire’s hiring pipeline is a configurable sequence of distinct, independently logged stages — resume vetting, an AI screening interview, live human interviews, and so on — rather than one opaque pass/fail decision made from resume text alone. Every candidate who reaches the AI screening stage is scored against the same fixed criteria: technical ability, communication, and cultural fit, each out of 100, against a threshold the employer sets, and every candidate for a given role is asked from the same defined question set in the same order, with the AI interviewer free to ask follow-ups or rephrase but not to swap in different questions for different people. And — this is the piece most directly relevant to the doom loop — when a candidate’s score falls below that threshold, the default behavior is not to auto-reject them. It routes to pending review for a human decision; auto-rejection has to be deliberately turned on, stage by stage, by the employer.

None of that is a detection mechanism, and we’re not presenting it as one. What it is, instead, is a different answer to the same diagnosis Chait and the survey data point to: if text-layer detection of AI authorship is unreliable on both the software and the human side, the stage where volume gets sorted matters less than what happens to the candidates who land near the threshold. A live screening interview is a real-time interaction a candidate has to actually show up and do, which is a meaningfully different ask than editing a resume, independent of whether anyone can tell if the resume was AI-assisted. And keeping a human decision point in the loop by default for the borderline cases — rather than trusting a single automated score to close the file — matters more, not less, as application volume rises and the cost of getting one threshold wrong at scale grows with it.

The honest summary

The AI doom loop is a real, measured phenomenon: two independent Robert Half surveys, a continent apart, landed on the same 67% figure for hiring teams slowed down by AI-generated application volume, and Greenhouse’s own candidate data shows the pressure that’s driving it. But the instinct to fix it by getting better at detecting AI-written text runs into a hard problem — the tooling built to do it doesn’t, by vendors’ own testing, and the humans who believe they can do it have no independently measured accuracy rate behind that belief. The more credible responses on the table right now, Greenhouse’s volume-limiting feature and our own default-to-human-review structure, skip the detection question entirely and change something else instead: how much a single application costs to send, or how much weight a single automated score is allowed to carry on its own. Neither one claims to have solved the doom loop. Both are at least arguing with the right part of the problem.

Sources consulted directly for this piece: Fortune’s and HR Dive’s July 2026 coverage of Greenhouse CEO Daniel Chait’s “AI doom loop” remarks and the “My Dream Job” feature; Robert Half’s March 2026 US press release and Staffing Industry Analysts’ September 2026 coverage of Robert Half’s UK survey; StylingCV’s and Jobscan’s independent testing of major applicant tracking platforms for AI-resume detection; and Resume Genius’s 2026 Hiring Insights Report (also covered by Black Enterprise).