Two hundred and fifty-four applications per posting. A synthetic candidate who clears a screening call. A recruiter buried under a 412 per cent increase in volume. These read as three separate crises. They are one, and it has a single cause.
Recruiting teams are currently running two emergency workstreams that nobody has connected. One is about volume: too many applications, too little signal, recruiters drowning. The other is about fraud: candidates who are not who they say they are, interviews conducted by a face that does not exist, and in the worst cases an employment relationship that turns out to be a foreign revenue operation.
They are the same problem, viewed from two ends. The marginal cost of producing a job application collapsed to approximately nothing. The marginal cost of evaluating one did not move.
The volume side, with numbers attached
Greenhouse's chief executive Daniel Chait described the mechanics to Fortune on 27 July 2026, and the figures are worth quoting precisely because they come from platform telemetry rather than a survey. Across roughly 175,000 live jobs on the platform: an average of 254 applications per posting, and applications per recruiter up 412 per cent. In the UK, more than 1.2 million applications for fewer than 17,000 graduate roles last year. Candidates are paying around $20 for tools that apply on their behalf, at scale, while they sleep.
These are Greenhouse's numbers about Greenhouse's customers, and should be attributed that way rather than treated as a market-wide measurement. But the direction is not seriously contested by anyone hiring at volume this year.
Chait's own term for it is a doom loop, and the loop is the interesting part. Candidates automate because response rates are terrible. Response rates are terrible because volume is overwhelming. Employers respond by adding screening, which raises the cost of applying — but the candidate has automated that cost away, so the screening lands almost entirely on the genuine applicants who are still doing it by hand. Each turn of the loop degrades the signal further.
Every intervention that makes screening harder is paid for by the candidates who were not gaming the system. That is why the loop tightens rather than settling.
The fraud side is the same economics, further along
On 6 May 2026, the US Department of Justice announced the sentencing of two US nationals, Matthew Isaac Knoot of Nashville and Erick Ntekereze Prince, each to 18 months for running "laptop farms" that enabled North Korean IT workers to hold remote roles at US companies. Roughly 70 US companies were affected across the two cases, generating over $1.2m for the DPRK. Prince's scheme alone accounted for $943,069 in salary payments and more than $1m in remediation costs; Knoot's for over $250,000 paid and more than $500,000 in remediation.
Note the remediation line. In both cases it exceeded or approached the salary paid. The cost of a fraudulent hire is not the wage — it is the forensic exercise afterwards.
On prevalence, the honest position is that good independent data does not exist. The best available measurement is vendor-funded: research from GetReal Security released 11 December 2025, surveying 668 IT, security, fraud and risk leaders at organisations of 1,000-plus employees, fielded September 2025. It found 41 per cent of enterprises saying they had hired and onboarded a fraudulent candidate, 88 per cent encountering deepfake or impersonation attempts at least occasionally, and only 40 per cent believing their defences adequate. Methodology is disclosed, the commercial interest is obvious, and both facts should travel with the number.
One correction worth making, because this publication keeps encountering it. The widely-repeated line that one in four candidate profiles will be fake by 2028 comes from a Gartner press release dated 31 July 2025, attributed to senior research director Jamie Kohn and drawn from a survey of 3,000 job candidates. It is a forward-looking prediction. It is now routinely reprinted as though it describes hiring today. It does not, and anyone building a business case on it should know which it is.
Why they are one problem
Fraud at scale depends on precisely the condition that produces the volume crisis: applying is free. A synthetic candidate is only economically viable if you can generate hundreds of them for the cost of an API call and put them through a process that never requires simultaneous, verified, human presence. Remove either condition and the arithmetic stops working — not because you detected the fraud, but because you made it uneconomic.
This is why the framing matters. Treat it as a detection problem and you buy a detection vendor, add a screening step, and pay for it with the genuine applicants' patience. Treat it as an economics problem and the intervention set changes entirely.
Visual 1 — Interventions, and who actually pays for them
Intervention | Effect on cost of applying | Who bears it |
|---|---|---|
More knockout questions | Near zero — automated tools answer them | Genuine applicants only. Net negative. |
Longer application forms | Near zero for automated applicants | Genuine applicants. Net negative. |
CV-parsing fraud detection | None — acts after the application exists | The employer, in licence fees and false positives |
Synchronous live interaction early in the funnel | High — requires real-time presence per application | All applicants equally; automation loses its leverage |
Verified identity before interview | High for synthetic identities, low for real ones | Asymmetric in the right direction |
Fewer, better-specified postings | Unchanged per application, but shrinks the target surface | The employer, in hiring-manager discipline |
How to read it: Only interventions that reintroduce a real cost on the applicant side — presence, or verified identity — change the underlying economics. Everything above the line taxes the honest applicant and leaves the automated one untouched.
What the market actually shipped
Greenhouse completed its acquisition of Ezra AI Labs on 27 May 2026, bringing conversational voice AI into the hiring process, with founder Ophir Samson joining as head of voice AI. It launched US-only.
It is worth being precise about the pitch, because it is easy to misread. Ezra is not marketed as fraud detection or identity verification. The mechanism is the one in the table above: a mandatory synchronous voice interaction early in the process imposes a real-time cost per application, and that cost is what breaks mass-applying. The fraud benefit, to the extent there is one, is a side effect of the economics rather than the product claim.
The other detail worth noting is governance. Ezra falls under Greenhouse's programme of monthly independent bias audits published via Warden AI on a public assurance dashboard. Given where the regulatory floor is heading — Colorado's surviving 30-day explanation and human-review duties from January 2027, New York City's Local Law 144 already in force — publishing audits monthly rather than annually is a stronger position than most of the market holds. It also creates an artefact that exists in writing, which is a double-edged thing in a jurisdiction with active AI-hiring litigation. Mobley v. Workday is still live in the Northern District of California, where in March 2026 the court largely refused dismissal and allowed applicant claims under the ADEA to proceed.
What to do
Measure your own cost per evaluated application, not per hire. Cost per hire hides the problem entirely — it divides by the successes. The number that has moved is what you spend looking at the ones you reject.
Move one synchronous step earlier in the funnel. It does not have to be an AI interview. A short live interaction of any kind, placed before the expensive human review, changes the economics for every automated applicant at once.
Decide where identity verification sits, and say so publicly. Candidates tolerate verification far better when it is disclosed up front than when it is sprung at offer stage. If you operate in the EU, note that any conversational AI in that flow now carries a disclosure obligation under Article 50 of the AI Act, live since 2 August 2026.
Cut postings before you cut candidates. Volume is partly self-inflicted: vague, duplicated and speculative postings attract exactly the untargeted applications teams then complain about.
Assume your bias audit is discoverable. Whatever assurance programme you run, write it as though it will be read in a deposition, because the trajectory of the litigation suggests it eventually will be.
The instinct in most talent functions right now is to buy something that reads applications faster. It is the one intervention guaranteed not to work, because reading faster is a response to symptoms in a system whose defining feature is that producing the input is free. The only durable answers put a price back on applying — and the uncomfortable truth is that every one of them makes the process slightly less convenient for the candidates everyone says they are trying to attract.
Sources and method. An HRHubsMedia original. Application-volume figures from Greenhouse CEO Daniel Chait via Fortune, 27 July 2026 — these are Greenhouse platform figures, not market-wide measurements. DPRK IT-worker sentencings per the US Department of Justice, 6 May 2026; note that a separate, similar DOJ prosecution exists with different figures and the two should not be conflated. Prevalence data from GetReal Security, released 11 December 2025 (n=668, fielded September 2025) — vendor-funded research with disclosed methodology. The "1 in 4 fake profiles by 2028" figure is a Gartner prediction dated 31 July 2025, reported via HR Dive, and is not a measurement of current conditions. Ezra AI Labs acquisition per Greenhouse, 27 May 2026. Mobley v. Workday, No. 3:23-cv-00770-RFL (N.D. Cal.), March 2026 ruling per HR Dive; the case remains ongoing and nothing here characterises its merits. We found no independent, non-vendor prevalence measurement of candidate fraud dated within 2026. Journalism, not legal advice. Corrections will be made openly on this article.
