Greenhouse CEO Daniel Chait gave the current state of hiring a name this year that stuck: the "AI doom loop." Candidates use AI to apply to more jobs, faster, than any human could write cover letters for. Employers respond by using AI to filter those applications faster than any recruiter could read them. Each side's tooling exists specifically to beat the other side's tooling. Nobody designed this on purpose, and — based on survey data from the company that sells the filtering half of the loop — neither side thinks it's working.
The numbers behind the name
The volume growth is not a vibe, it's measured. Ashby's May 2026 analysis of more than 100 million applications and 200,000 job postings found applications per hire have roughly tripled since 2021, with many roles now averaging 300 or more applications before a single hire. A separate Ashby breakdown of EMEA data (released September 2026) found the growth is uneven by role and region — EMEA technical roles average 254 applications per hire versus 610 in the Americas — but the direction is the same everywhere it measured. Gem's 2026 recruiting benchmarks report, drawn from 165 million applications and 1.2 million hires, puts it in blunter terms: recruiters are handling 93% more applications per hire than they were in 2021. None of these are competitors citing each other — they're three ATS and recruiting-ops vendors independently measuring their own customers' pipelines and landing on the same slope.
The platform-level number people actually recognize is LinkedIn's: in mid-2025 the company told the New York Times it was processing roughly 11,000 job applications per minute, up 45% year over year, with generative AI tools cited as a direct driver (as relayed by eWeek). That figure is now over a year old and nobody has published a 2026 refresh in our research, so treat it as a baseline rather than today's rate — but even as a floor, it's the context for everything downstream of it.
Greenhouse, whose CEO coined the "doom loop" phrase, backed it with its own November 2025 AI in Hiring survey of more than 4,100 job seekers, recruiters, and hiring managers across the US, UK, Ireland, and Germany. The split is the real story: 70% of hiring managers said they trust AI to help them make faster, better hiring decisions. Only 8% of job seekers said they think AI-driven screening makes hiring fairer. And on the recruiting side specifically — the people actually running the filtering software — only 21% said they were "very confident" their own systems weren't screening out qualified candidates. That's not candidates distrusting a system recruiters stand behind. That's most recruiters not fully trusting their own tools either, while still using them, because the volume leaves no other option.
Why "better filtering" doesn't break the loop
The instinctive fix — make the employer-side AI smarter, stricter, more selective — is the one every ATS vendor is currently shipping, and it's worth being precise about why it doesn't close the loop so much as reset it one level up.
An applicant using an AI tool to tailor a resume isn't guessing at what a human reader wants. Increasingly, they're optimizing directly against the pattern that automated resume-matching systems reward: keyword density against the job description, phrasing that mirrors the posting, structured bullet formats that parse cleanly. When the employer's screening model gets more sophisticated, the applicant-side tools — many costing a flat monthly fee to blast an application to every opening that matches a profile — adapt to the new pattern just as fast, because they're built to. Both sides are training against each other's model, not against the job. Chait's framing, as reported by Fortune in July 2026, is specifically that each side is using AI to try to help itself, and the two efforts cancel out: application volume keeps climbing while the share of applications that actually convert to interviews or hires keeps falling. Tightening one side's filter doesn't restore signal. It just raises the bar the other side's model learns to clear next.
There's a legal dimension layered on top of the volume problem, and it's not hypothetical. Mobley v. Workday, the closest thing the US has to a landmark AI-hiring case, is still active: a federal judge in California allowed disparate-impact discrimination claims against Workday's screening tool to proceed past a motion to dismiss, on the theory that an AI vendor can be sued as an employer's "agent" when its scoring shapes who gets hired. A second case against Sirius XM, built on its iCIMS-based screening and tracked by the OECD AI Incidents Monitor, alleges similar effects through proxies like ZIP code and school. Neither case has reached a final judgment as of this writing, and the legal theory (vendor-as-agent liability) is being tested, not settled — but the practical lesson for anyone buying or building screening AI is already visible: funneling every applicant through one black-box matching score, with no visibility into what it actually measures, is a liability surface as well as a trust problem. Tightening that same single filter doesn't reduce the exposure. It concentrates more hiring decisions in the one model a plaintiff's lawyer would need to depose.
What would actually interrupt it
If the loop is driven by both sides training against the same single gate, the structural counter isn't a smarter gate — it's more than one gate, measuring different things, so that gaming one doesn't clear all of them. A resume-keyword-optimization tool is well-matched against a resume-keyword-matching filter. It's a much weaker bet against a live, unscripted follow-up question, because generating a plausible resume bullet and sustaining a coherent real-time answer are different skills to automate. A pipeline that separates "does the application match the role on paper" from "can this person actually discuss the work" from "can they do a bounded version of the task" forces an applicant-side tool to win three different games instead of gaming one shared pattern.
That's the architectural bet behind NiceHire's own pipeline, and it's worth being exact about what we can actually stand behind rather than rounding it up. Every candidate who goes through NiceHire's AI screening interview is scored against the same fixed criteria — technical fit, communication, and cultural fit, each out of 100 — against a threshold the employer sets, not a model-specific score that shifts session to session with no visible rubric. Separately, that AI screening step is available to candidates on their own schedule, without waiting on a recruiter to open a calendar slot — a real availability property, not a claim about how much faster it makes the overall funnel.
What we're not claiming is that this ends the arms race or produces a fairer outcome in any audited sense. It's still an AI system scoring people, it carries the same category of scrutiny Mobley is currently testing against Workday, and — as we've written before — no US jurisdiction outside New York City currently mandates the kind of independent bias audit that would let any vendor, including us, say "verified unbiased" instead of "consistent criteria." Consistent, disclosed scoring is a mechanism we can point to in the code. It is not a substitute for the audit nobody in this market has actually run.
The practical read
The doom loop isn't a temporary spike that AI-detection tools or stricter filters will quietly resolve — the vendors building both halves of it are reporting record volume and record distrust in the same breath, from the same surveys. The fix that's actually being tested right now, in court as much as in product design, isn't "better at catching AI-written resumes." It's whether hiring decisions run through one opaque gate or several disclosed ones. That's a question every team evaluating screening tools in 2026 can ask a vendor directly, regardless of which one they pick.
Sources: Greenhouse's "An AI Trust Crisis" report (November 19, 2025); Fortune's July 27, 2026 coverage of Greenhouse CEO Daniel Chait's "AI doom loop" framing; Ashby's May 2026 global application-volume release and its September 2026 EMEA-specific release; Gem's 2026 Recruiting Benchmarks Report; the New York Times' mid-2025 LinkedIn application-rate reporting as relayed by eWeek; and Bloomberg Law's coverage of the Mobley v. Workday motion-to-dismiss rulings alongside the OECD AI Incidents Monitor's record of the related Sirius XM/iCIMS suit.
Ready to transform your hiring?
See how NiceHire's AI-powered hiring platform works for your team.
Get Started