Job‑Hunt
0 hired
Public research

The AI-Tell Audit

We ran the production checks that scrub every application Job-Hunt generates against 49 real resumes people posted to Reddit. Most of them read as AI-assisted — including, most likely, some that weren’t.

69%
of these resumes trip at least one AI-tell check.
34 of 49 resumes · 33 Reddit users · run on 7 August 2026

Check your CV

We run the same checks on your CV that we ran on the 49 resumes in this audit. Nothing is uploaded.

37%
carry a strong LLM fingerprint — em dash, invisible unicode, or 3+ banned phrases
45%
use vocabulary recruiters now read as machine-written
1 in 10
contain invisible characters no keyboard produces
26/33
users have at least one flagged resume
01

What we found

People post their resumes to Reddit every day asking why they aren’t getting interviews. We took 49 of those resumes and ran them through the exact detection logic that scrubs every application generated on Job-Hunt. No bespoke scoring, no AI detector black box — just the mechanical, vocabulary, and style checks our own output has to pass.

Tells are everywhere

Two thirds of the resumes trip at least one detector, and more than a third carry a strong LLM fingerprint: an em dash, invisible unicode, or three or more phrases from the banned vocabulary list. 26 of the 33 users have at least one flagged resume.

Our sanitizer would rewrite half the corpus

51% of these resumes, had they arrived as LLM output, would be modified by the same 40+ checks that clean every generated application.

Mechanical
20%
Real but rare — zero-width chars in 1 in 10
Vocabulary
45%
Widespread — worst resume packs six banned phrases
Style
61%
Softest signal — the participle-tail cadence detectors flag
02

How we measured

The audit reuses production detection logic. Nothing bespoke, everything reproducible.

  1. 1

    Corpus. 49 resumes from 33 Reddit users, keeping only documents with at least 200 characters of real prose.

  2. 2

    Detectors. The exact method we use in production: mechanical unicode tells, the 35-phrase banned vocabulary list, and the style rules from our generation prompt.

  3. 3

    Scoring. Headline numbers use conservative scoring: the spaced en dash ( – ), a common word-processor artifact, is not counted as a tell. The strict variant is in the comparison table below.

  4. 4

    Privacy. Only aggregate counts left the analysis. No resume content was read, retained, or stored.

03

Characters humans don’t type

Zero-width unicode cannot be typed on a keyboard. It arrives one way: by copy-pasting out of a chatbot. 1 in 10 of these resumes contains literally invisible characters — the closest thing this audit has to a smoking gun, because no keyboard produces them.

Resumes containing each mechanical tell (of 49)
Invisible unicode (zero-width)5
Em dash —4
Smart quotes “ ” ‘ ’2
Exotic spaces (nbsp, thin…)1
Ellipsis character …0
Spaced en dash – (excluded)16

Why the gray bar doesn’t count. The spaced en dash appears in 33% of resumes, but Word and Google Docs autocorrect produce it on their own, so it is excluded from headline scoring. Our product still normalizes it — an ATS parser doesn’t care where a character came from.

04

The vocabulary problem

22 of 49 resumes (45%) use at least one phrase from the banned vocabulary list, and 11 (22%) use three or more. These are the words recruiters and AI detectors now read as machine writing — whether a human typed them or not. “Cross-functional collaboration” alone appears nine times. The single worst resume packs six different banned phrases from the 35-phrase list.

Most frequent banned phrases · total occurrences across all resumes
cross-functional collaboration9
spearheaded7
proven track record6
passionate about5
seamless5
leveraged / leverage4
utilized4
fostering4
results-driven3
robust3
meticulous2
05

Style patterns

Style is the softest signal — it means “reads like AI”, not “is AI”. Two patterns stood out, in opposite directions.

06

How concentrated are the tells?

The tells aren’t spread evenly. Only 15 of 49 resumes are fully clean, and 13 sit at six or more tells — where the AI read is hard to argue with.

Resumes by total tell count · spaced en dash excluded · n = 49
Fully clean15
1–5 tells21
6 or more tells13
51%

of the corpus would be rewritten by our sanitizer if it arrived as LLM output. 26 of 33 users have at least one flagged resume.

07

Strict vs. conservative scoring

Counting the spaced en dash moves the headline number — but not the strong-fingerprint rate, which never depended on it. Headline metrics on this page use the conservative column; strict is shown for transparency.

MetricStrict (en dash counts)Conservative
Any tell at all41 / 49 · 84%34 / 49 · 69%
Any mechanical unicode tell24 / 49 · 49%10 / 49 · 20%
Strong LLM fingerprint18 / 49 · 37%18 / 49 · 37%
Any banned AI phrase22 / 49 · 45%22 / 49 · 45%
Participle-tail sentences30 / 49 · 61%30 / 49 · 61%
Completely clean resumes815
Users with a flagged resume30 / 3326 / 33
08

The caveats we publish with

The honest footnotes that keep the numbers defensible.

Small sample 49 resumes, 33 users. That is why every percentage above ships with its raw count.
Word-processor overlap Smart quotes and en dashes can come from Word or Google Docs autocorrect, not a chatbot. That’s why conservative scoring exists — zero-width unicode and banned phrases are the defensible core.
“Reads like AI” ≠ “is AI” Humans genuinely write participle-tail bullets; detectors flag them anyway. This is a perception problem, not proof of authorship.
Import path Some resumes arrived via PDF import, which preserves the original file’s characters — the tells belong to the author’s document, not our pipeline.
09

Why this matters when you apply

We don’t know for sure how any given company screens resumes — practices differ wildly between a five-person startup and an enterprise ATS pipeline. Our decision is pure caution: English is rich enough to tell your story without a single one of the tells a filter could read as a red flag.

We checked the resumes people bring to Reddit. 69% would trip at least one AI-detection heuristic. Nearly half use vocabulary recruiters now flag as machine-written — and 1 in 10 literally contains invisible ChatGPT characters. That’s why every application we generate is scrubbed through 40+ checks before you send it to any ATS.
Free to start · no card required

Check your own CV. Run these checks on your resume and see which tells it contains. It runs entirely in your browser — the file is never uploaded, and no account is needed.

Reproducible. Every number comes from the audit itself: production detection logic imported from sanitizeLlmText.ts, run on 7 August 2026. Aggregates only — no resume text retained.

69% of the resumes we checked trip an AI-tell check. Yours doesn’t have to.