The 14% Overlap: Why AI Screeners Shortlist Different People Each Run

If you got rejected last Tuesday and shortlisted this Tuesday with the same resume, you are not imagining it. In February 2026, recruiting commentator Greg Savage went on episode 300 of The Resilient Recruiter and dropped a stat that has been bouncing around recruiting outlets ever since: run the same AI screener twice over the same resumes, and the two shortlists overlap by just 14%. That is not a filter. That is a weighted coin flip with a confident tone of voice.

What the 14% actually measures

The 14% is the agreement rate between leading AI screeners asked to shortlist the same resumes for the same role. It comes from Eunomia HR's "LLM Reality Check" field experiment in May 2025, which ran 300 head-to-head resume screens using ChatGPT-4o, Gemini 2.0 Flash, and Grok 3 against 109 anonymized HR Business Partner resumes for a global Meta role.

Translated into something you can picture:

• Two AI recruiters looking at the same CVs will reach different top-ten lists four times out of five.
• Three AI recruiters converging on a candidate is the equivalent of three human recruiters agreeing on one or two "top" picks out of ten.
• Human recruiters are not perfect either (inter-rater reliability sits around κ ≈ 0.49), but LLM overlap at 14% is roughly half the human band, with twice the volatility.

The mechanism is boring and important. Large language models sample from probability distributions. Temperature settings, prompt wording, and the order resumes are fed in all nudge the output. Your "fit score" is not calculated. It is drawn.

Why AI hiring tools are inconsistent by default

AI hiring tools are inconsistent because almost every employer runs one and almost none of them audit it. The World Economic Forum reported in March 2025 that roughly 88% of companies use AI for initial candidate screening, and the number has climbed since. Fewer than 10% of candidates are clearly told an AI evaluated them, per Enhancv's survey data.

Three things stack on top of the 14% overlap problem:

1. The skills-mismatch tax. Workable found that 63% of candidates with the right skills for a role get screened out by ATS and AI tools.
2. Keyword fragility. A 2026 ResumeAdapter analysis of 10,000 anonymized resume scans found that 82% of rejected resumes had fewer than 50% of the posting's required keywords present, even when the candidate had the actual experience in their work history.
3. AI-washing. Savage's follow-up piece in March 2026 pointed out that a lot of "AI parsing" is regular expressions from 1987 in a new jacket. His quick test: if the tool cannot match "Led development teams" to "Team Leader," it is not AI, it is grep.

A rejection email covers at least three failure modes: stochastic LLM variance, keyword regex, and vendors who claim intelligence they do not have. You cannot tell which one rejected you. The vendor often cannot either.

The pool-size math nobody tells applicants

The practical cost of 14% overlap depends on how big your candidate pool is, and it is bigger than most applicants assume. In Refolk's index of professional profiles, the pools for common US titles are large enough that tiny screener variance produces huge downstream differences in who gets interviewed.

Candidate-pool and employer figures from Refolk's index. Noise cost derived from the Eunomia 14% top-ten overlap finding.

Two things fall out of this table that most career advice ignores:

• In a 353,689-strong SWE pool, variance is the enemy. You beat a coin flip with more flips, not with a shinier coin. Submit more tailored applications.
• In a 6,965-strong PMM pool, the same twenty names likely surface across every screener in town. More applications barely help. Referrals and warm intros do the heavy lifting.

Tailored resume vs generic: the only move that still pays

A tailored resume does not fix the randomness. It raises the floor of the distribution so that when the dice do roll your way, the model has something defensible to anchor on.

Here is the math people are missing. Expected shortlists are roughly:

applications × tailor-quality × screener-luck

Only the first two are inside your control. Rewriting your resume for the eleventh time against the same posting does not move either number. Submitting a tailored version to a second employer moves the first. Rewriting against a different posting moves the second.

This is the exact work Refolk takes off you: paste the posting, get your own resume back rewritten for it, with the cover letter drafted and a fit score that tells you whether the posting is worth the application in the first place. The point is not a magic resume. The point is collapsing the per-application cost of "tailored" to something close to the cost of "generic," so you can run more flips.

Why chasing a "perfect resume" is the wrong response

Perfection does not exist in a stochastic system; distribution-shaping does. The right question is not "is this resume good?" but "across ten screener runs, how often does this resume clear?"

A few concrete reframes:

• Stop rewriting the same resume for the same posting. Diminishing returns kick in immediately. If version three did not clear, version seven will not either.
• Kill the white-text keyword hack. Greenhouse's 2025 survey found 41% of US job seekers said they had hidden instructions or keywords in a CV, but when Sapia researchers scanned about 197,000 real CVs they found it in roughly 1%. Most candidates either cannot execute it or get caught. Modern parsers flag invisible text.
• Treat the fit score as a go / no-go, not a grade. If a tailored version of your resume still scores poorly against a posting, the posting is wrong for you. Spend the next hour on a different one.
• Keep one canonical version a human has read. Referrals, recruiters, and hiring managers still see PDFs. The resume that wins a 14% lottery still has to survive a thirty-second human skim.

The bias that is not about bias

The scariest part of the 14% number is that it is bias-neutral, and most readers hear "AI bias" and assume a biased direction. The Eunomia result shows the opposite failure mode: the models disagree with each other as much as they disagree with humans. The randomness itself is a kind of discrimination because it punishes anyone whose candidacy sits in the middle of the pack.

Directional bias is still real on top of that. The most-cited 2024 study from the University of Washington (Wilson and Caliskan, AIES 2024) found that language-model-based resume rankers preferred white-associated names 85.1% of the time and Black male names 0% in head-to-head tests. A 2026 Pin audit of 33,000 job searches found that employer-prestige filters appear in 70.7% of recruiter searches, meaning even if you clear the AI, a sourcer filter has already run against your logos.

So the applicant is being filtered by:

1. A stochastic LLM with 14% self-consistency.
2. A directional bias that favors certain names and logos.
3. A regex pretending to be AI.
4. A sourcer prestige filter upstream of all of it.

No single resume beats all four. A tailored pipeline across many postings has a chance.

The legal ground is finally moving

Legal exposure for employers using AI screening is rising fast, which means the "AI rejected you" era has concrete standing for the first time. Three dated facts worth knowing as a candidate:

• Workday ADEA case. In May 2025, Judge Rita Lin in the Northern District of California conditionally certified an ADEA collective action against Workday on behalf of applicants aged 40 and over rejected by its AI screening since 2020.
• EEOC enforcement. 2026 enforcement data found that 74% of organizations investigated for AI hiring practices failed to maintain proper audit documentation, and 62% could not demonstrate meaningful human oversight.
• UK Data (Use and Access) Act 2025. Since February 2026, solely automated decisions are allowed on more grounds in the UK, but candidates must be told, must be able to challenge the decision, and must be able to get human intervention.

None of this gets you an interview tomorrow. It does mean that asking "was my application reviewed by a human?" is now a reasonable question, not a conspiracy theory. In some jurisdictions, you are entitled to the answer.

A one-week playbook built around the overlap

The one-week playbook that actually responds to a 14% overlap is wide, tailored, and auditable. Not perfect, not clever, not stuffed with invisible keywords.

• Monday. Pick 20 postings that match your actual experience. Discard any where a tailored version scores poorly.
• Tuesday to Thursday. Submit a resume tailored per posting. Draft a short cover letter that names the specific team or product line. Refolk handles both from the posting URL plus your history, which is where the per-application time cost otherwise explodes.
• Friday. For the 5 highest-fit postings, find a human. One LinkedIn message to someone on the team beats three more applications.
• Saturday. Audit. Which applications got auto-rejected inside an hour (regex), which got rejected after a day (human-in-the-loop or slow AI), which went silent (ghost queue)? Pattern-match for next week.
• Sunday. Rest. Variance resets on Monday.

Every input you can control (tailoring, volume, referral surface) is cheap and compounding. Every input you cannot (screener seed, model temperature, sourcer mood) is noise you simply have to out-sample.

FAQ

Does a longer, keyword-dense resume beat the 14% overlap?

No. Keyword density past a point triggers two separate failure modes: modern parsers flag suspiciously stuffed resumes, and LLM-based screeners often penalize incoherent bullets even when the keywords are present. The ResumeAdapter analysis found that most rejections correlate with missing relevant keywords, not with having too few total words. Target the keywords actually in the posting, written as real sentences, and stop there.

Should I apply to the same job twice if I got rejected?

Sometimes, but only with a materially different resume and a two to three week gap, and only if the posting is still open. Given 14% overlap, a second submission has a real chance of landing differently. But most ATS systems flag duplicate applicants, and recruiters notice. A better move is to apply to a sibling role at the same company or a similar role at a competitor.

How do I know if a company is using AI to screen me?

Mostly you cannot, which is itself the problem Enhancv's data captures: fewer than 10% of candidates are clearly told. Signals that AI is in the loop: rejection within minutes of submitting, generic rejection language with no human name, a company that publicly lists an AI hiring vendor, or a job description that reads like it was written by one. In the UK and parts of the EU, you can now ask directly and are entitled to an answer.

Is it worth paying a human resume writer in a 14% overlap world?

For one canonical version, yes. For every tailored variant, no, because the economics do not work. A human writer at typical freelance rates cannot keep up with the twenty tailored versions a serious job search now requires per week. The pattern that works is one human-edited master, then per-posting tailoring handled by a tool that rewrites against the specific job and scores the fit before you spend the application slot.