The machines are talking to each other. Where is the person in the loop?

Somewhere in New Zealand tonight, someone will send another job application. It is a small act of hope, and it now travels through strange country. AI tools can sit on both sides of it: helping the candidate write, and helping the employer screen or rank what arrives. Whether anything in between checks that the application’s claims are true depends on which tools each side happens to use. The public information reviewed for this article does not establish how consistently that checking happens, and that unsatisfying sentence is the most accurate one available.

So this article does the only responsible thing left. It lays out what is known about the loop, what is claimed but not established, and what New Zealand in particular cannot yet see: we did not identify current NZ-wide measurements of it.

One caution before we begin, because it shapes how you should read everything that follows. Most of the quantitative evidence available for this article comes from US or international studies. It describes the wider conditions surrounding the hypothesis, not how prevalent any of this is in New Zealand. Where a number is local, we say so. Mostly it is not, and that gap is itself part of the story.

What the evidence shows: volume

The volume is the well-documented part of this story, and the numbers have stopped being surprising only because we keep repeating them.

LinkedIn data reported by The New York Times put the flow at roughly 11,000 job applications submitted every minute in mid-2025, up more than 45% in a single year, amid a wave of AI-assisted applying [1]. Recruiting-platform data indicates applications per hire have roughly tripled since 2021, with roles now receiving more than 300 applications per hire on average [2].

Some services sell high-volume automated applying directly, and say so on their own websites. As of August 2026, one auto-apply platform advertises more than two million users; another sells “unlimited access” to its automation tools on subscription and asks, verbatim, “Want to apply to 1000+ jobs while watching Netflix?”, promising “hundreds of applications in minutes.” The same sites advertise outcomes (“80% more likely to get a job faster,” “61% of users get an interview in first 10 days”) with no methodology appearing alongside these claims on the pages we reviewed. We quote them as what they are: advertised claims, from services whose own marketing leads with volume and automation.

On the other side of the desk, the pile lands on software too. Fortune cites a World Economic Forum figure of nearly 90% of companies using AI to screen candidates [5], and in a Gartner survey of 3,290 candidates, 39% said they used AI during the application process [6].

So both ends of the pipeline are now partially automated, at scale. That much is established. Most of what follows from it is not, which is precisely the problem.

What is claimed, but not established

Three widely repeated claims deserve to be marked clearly as open questions, because each one is load-bearing for how seriously you take the rest.

“AI screeners prefer AI-written applications.” Industry commentary asserts that screening models reward the fluent, keyword-dense text that language models produce by default [3]. If true, it would mean the writing machine and the reading machine are, in effect, grading each other’s homework. One controlled experiment supports part of it: a 2025 preprint, later presented at two AI-ethics venues, found large language models preferred AI-written resumes 67 to 82 per cent of the time in simulated hiring pipelines across 24 occupations, and candidates using the matching model were 23 to 60 per cent more likely to be shortlisted than equally qualified applicants submitting human-written resumes [4]. That is evidence from a lab simulation, not from deployed commercial screeners. Whether production systems show the same preference is the question a regulator, a university, or a vendor willing to open its system could answer.

“Automation made hiring worse.” SHRM’s 2025 benchmarking survey reported average cost-per-hire and time-to-hire both rising over the prior three years [7], though its 2026 benchmark complicates the picture, reporting median nonexecutive time-to-fill falling. And 84% of US hiring managers in a Robert Half survey say AI-generated applications have left their teams with heavier workloads [8]. But correlation is doing a lot of work in that sentence. The same years brought tighter labour markets in some sectors, hiring freezes in others, duplicate and stale listings, and changed recruiting practices. The automation-made-it-worse story is plausible; it is not isolated from those confounders by any evidence we have seen. Plausible is not the same as shown.

“The screening can’t be trusted.” Suggestive, not settled: in one 2025 survey, only 21% of recruiters and hiring managers were very confident their systems were not rejecting qualified candidates [9], and in a Gartner survey of 2,918 candidates, only 26% said they trust AI to fairly evaluate them [6]. The same Greenhouse survey also found 70% of hiring managers trust AI to make faster and better hiring decisions, so the picture on the employer side is enthusiasm and doubt together.

The surveys document concern on both sides. Whether the concern is warranted, meaning how often screeners wrongly reject qualified people, is exactly the kind of question for which we identified no current, public, NZ-wide dataset.

The feedback-loop hypothesis

Put the established pieces next to each other and a mechanism suggests itself. We state it here as a hypothesis, the article’s central one, not as a finding:

1. Low response rates may push candidates toward volume.
2. AI and automation can reduce the marginal effort required to send each application, which could raise volume further.
3. Rising volume may make individual attention impractical, encouraging employers to automate screening.
4. Automated screening could in turn lower response quality and increase ghosting (a link none of the evidence we cite establishes), pushing candidates back to step 1, with more automation.

↺ and the loop starts again, one turn tighter

The evidence cited here is consistent with parts of this loop, but it does not establish the complete chain or prove that one step causes the next. What it does document, in US data at least, is the weather at the candidate end.

In one survey of 1,000 US job seekers, 94% said they had applied and never heard back, and by their own estimate roughly two-thirds of their applications drew no response at all; the same survey put the average completed search at 6.6 months and around 62 applications [10]. Burnout mentions in Glassdoor reviews rose 65% year-on-year in the first quarter of 2026 [11].

If the loop is real, the behaviour it produces deserves sympathy before it gets judgment. One possibility worth testing is that some people mass-apply not because they believe it works, but because low response rates make volume feel like the only lever left, and because, when automation reduces the marginal effort to almost nothing, sending one still feels like doing something. Applying becomes the way of not giving up. If a study of jobseeker behaviour bore that out, the flood would read not as laziness but as a rational response to a broken feedback signal, one that then feeds the very flood that broke it.

The incentive structure (analysis, not accusation)

It is worth describing the commercial structure of this market plainly, without attributing intent to anyone in it.

For the candidate-side services examined for this article (a dated sample recorded in our evidence file), revenue was generally tied to access or application volume rather than to a verified hiring outcome. These business models can earn revenue from access or usage without payment being contingent on a verified hire or a shorter job search. That does not describe the whole market: outcome-linked models such as placement fees exist and do not fit this pattern. And it is a description of incentives, not of motives: we have no evidence that any vendor wants hiring to work badly, and plenty of these tools are built by people trying to help. But incentive structures matter independently of intent, and the ones we examined do not reward de-escalation.

The question it leaves a candidate with is the practical one: should I just learn to beat the screener? Keyword-stuff, mirror the ad, let a bot carry the volume. Our analysis, again labelled as such, is not that gaming never works for an individual; we have no evidence either way. It is that widespread gaming may erode the usefulness of the screening signal itself, and that what stays valuable to the human at the end of the pipeline is not specificity, which gaming can fake, but checkability: claims that hold up against the person’s actual documents and under challenge in an interview. Unsupported claims, however specific, are harder to sustain there.

The law: moving elsewhere, unmeasured here

This is not legal advice.

Two developments abroad bear directly on automated hiring, and both circle the same defence: “a human makes the final call.”

In the EU, the Court of Justice’s SCHUFA judgment (C-634/21, 7 December 2023) held that, in the credit-scoring arrangement before the court, an automated score was itself a decision under GDPR Article 22 where the recipient drew strongly on it, on facts where an insufficient score led to refusal in almost all cases. The court reasoned that treating the score as a mere preparatory step would let scoring escape Article 22’s protections [12][13]. The related proposition that token human sign-off cannot take a process outside “solely automated” comes from the Advocate General’s opinion in the case and from EU regulators’ guidance that human oversight must be “meaningful, rather than just a token gesture”, not from the judgment itself. How far that reasoning extends to recruitment screening is a live question in European practice: commentators read it broadly, but its application beyond the facts of the case is legal argument, not settled rule, and we take no position on it here.

In Australia, a new transparency obligation (APP 1.7) commences on 10 December 2026: entities must describe in their privacy policies the kinds of automated programs that make, or do things “substantially and directly related to” making, decisions that could significantly affect individuals. The regulator has signalled a broad reading in which a program that recommends or guides a decision can be caught even where a human decides, if its output is a key factor [14][15]. Recruitment screening appears among the examples discussed. The obligation is transparency in a privacy policy, not a right to contest an automated decision; the regulator’s full guidance is expected by September 2026 and had not been released at the time of writing. Because the Australian Privacy Act reaches entities carrying on business in Australia, some New Zealand organisations will be within its scope [16].

New Zealand’s own Privacy Act 2020 contains no provisions specific to automated decision-making, even after the 2025 amendment, which added only a notification rule for indirectly collected information. The Privacy Commissioner has called for stronger protections, writing that “people should know why an automated decision is taken against them” [17], and has issued expectations-level guidance on AI use (impact assessments, transparency, human review), including that simply having a human in the loop may not be enough. But these are guidance, not statutory obligations [18].

We put the contrast no more strongly than this: from December 2026, transparency about automated screening will be a statutory obligation for entities within Australia’s regime, while in New Zealand it remains a regulator’s expectation. What that difference means in practice for a New Zealand candidate screened by any given tool depends on facts about that tool and its operator that are, in keeping with the theme of this article, not publicly established.

What New Zealand should measure

If there is a single conclusion this article is entitled to, it is this: we did not identify a current NZ-wide dataset measuring the loop’s most important properties (how we searched is recorded in our evidence file). The questions below are ones a regulator, university, or industry body could actually answer:

• What fraction of NZ job applications pass through automated screening before any human sees them?
• How often do automated screeners reject candidates who meet the stated requirements? (The claim that they prefer AI-written text is testable the same way.)
• What are actual response and ghosting rates in the NZ market, by sector?
• Do employers using automated screening disclose it to applicants, and would disclosure change applicant behaviour?
• What does the arms race cost both sides, in time and money, against its benefits?

Answering any one of these questions would provide evidence that is currently missing from the public debate.

Where Rolebird stands in this

Rolebird is one attempt at an evidence-preserving alternative on the candidate side, and it is fair that we state our own position factually: Rolebird checks direct identifiers before documents are sent for AI processing, grounds screening matches in evidence from the applicant’s documents, and does not submit applications on the applicant’s behalf. It sells fixed-use packs, and purchases do not renew automatically.

We are a participant in this market, not a neutral observer, which is exactly why this article stops at the evidence, labels its hypotheses, and puts its central claims on the list of things someone should test. If you want to see what evidence-grounded output looks like in practice, a complete worked example is public, no account needed.

Sources

1. The New York Times (DealBook), “A.I. Sludge Has Entered the Job Search”, 21 June 2025, reporting LinkedIn platform data.
2. Ashby, 2026 Talent Trends Report and press release, 7 May 2026. Dataset: 109M applications, 247K jobs, Jan 2021 to Mar 2026.
3. Metaintro, “AI Resume Screeners Now Prefer AI-Written Resumes”, 2026 (industry commentary).
4. Xu, Li and Jiang, “AI Self-preferencing in Algorithmic Hiring”, arXiv 2509.00462 (2025), presented at EAAMO 2025 and AIES 2025 (non-archival).
5. Fortune, “Inside the AI arms race reshaping hiring”, 1 June 2026, citing World Economic Forum figures.
6. Gartner press release, 31 July 2025: 4Q24 candidate survey (n=3,290) and 1Q25 candidate survey (n=2,918).
7. SHRM, “Recruitment Is Broken. Automation and Algorithms Can’t Fix It.”, 2026, citing SHRM 2025 Benchmarking Survey (member-gated).
8. Robert Half, US hiring-manager survey (n=2,000+), released 10 March 2026.
9. Greenhouse, 2025 AI in Hiring Report, released 19 November 2025 (n=4,136).
10. United Way of the National Capital Area, job-seeker survey (n=1,000, fielded 26 Feb–7 Mar 2026), published 5 May 2026.
11. Glassdoor Economic Research, “The workplace climate is burning us out”, 13 May 2026.
12. Court of Justice of the EU, Case C-634/21 (SCHUFA Holding, Scoring), judgment of 7 December 2023, EUR-Lex CELEX 62021CJ0634.
13. IAPP, “Key takeaways from the CJEU’s automated decision-making rulings”, December 2023.
14. OAIC, Automated Decision-Making Transparency Obligation (APP 1) Issues Paper, 18 May 2026.
15. Allens, “Automated decision-making transparency: what APP entities need to know”, 10 June 2026.
16. Buddle Findlay, “What Australia’s new automated decision-making rules mean for New Zealand organisations”, 29 April 2026.
17. Office of the Privacy Commissioner (NZ), “Privacy Act 2020 turns 5 - changes are needed”, 1 December 2025.
18. Bell Gully, “AI and Privacy in New Zealand: a practical guide”, 20 May 2026.
19. Privacy Act 2020 (NZ), legislation.govt.nz (current as at access date).