# Recruiters Can't Spot AI Writing. They Spot Bullets With No Nouns

_6 in 10 US job seekers changed or considered changing how they write to avoid sounding like AI. Recruiters can't tell. Vague bullets are the real tell._

- Canonical URL: https://www.refolk.ai/candidates/blog/recruiters-cant-spot-ai-writing-fix-resume-nouns
- Published: 2026-11-08
- Reading time: 10 min
- Keywords: how to make resume not sound like AI, AI generated resume detection recruiters, cover letter sounds like ChatGPT, do recruiters reject AI written resumes

If you've been scrubbing em dashes and "spearheaded" out of your resume so it doesn't read as AI, you're fixing the wrong thing. In controlled tests, recruiters barely tell AI prose from human prose. What they do notice is a bullet with no detail that only you could know.

In October 2026, Greenhouse reported that 6 in 10 US job seekers had changed, or considered changing, how they write so they don't sound like AI. The survey covered more than 2,800 job seekers across five countries, 1,200 of them in the US. The numbers that matter for your resume:

- **39%** had already rewritten an application so it wouldn't read as AI-generated
- **21%** had considered doing the same
- **77%** used AI somewhere in their search in the past year
- **4 in 5** were rejected without a single conversation

Greenhouse CPO Sharawn Tipton put the employer side bluntly: employers "sift through AI-polished resumes that all look the same." Read that closely. Her complaint is that the resumes *look the same*. She doesn't say they sound robotic. Sameness comes from content, and swapping punctuation doesn't change your content.

## Can recruiters actually detect AI-generated resumes?

No, not reliably. In controlled tests, recruiters identify AI-written applications at the rate of random guessing or only slightly better, even though most of them believe they can.

| Study | Who judged | Result |
|---|---|---|
| Tilburg University (Journal of Labor Economics) | Recruiters from Philips, PwC, Rabobank, VodafoneZiggo | No better than chance at spotting AI cover letters |
| TopResume, 2025 | Reviewers picking the AI resume out of four | 33.5% correct vs. a 25% guessing baseline, in about 20 seconds |
| Resume Genius, 2026 | 1,000 US hiring managers | 80% say they can spot AI writing |
| StylingCV ATS audit | Major applicant tracking systems | AI and human resumes parsed and scored identically |

The applicant tracking system (ATS), the software that parses and ranks applications before a person reads them, doesn't care who wrote your resume. The human reading it afterward thinks they can tell, but in testing they mostly can't.

So why do 80% of hiring managers feel confident? They're reacting to something real and calling it the wrong thing. A vague bullet triggers the same "this is AI" reaction whether a model wrote it or you did at 1 a.m. The tell is how empty the bullet is.

On em dashes specifically: I found no survey data showing they're a meaningful rejection trigger, and none showing the opposite. When a lever is that hard to find evidence for, it isn't the one to pull.

## What actually makes a resume read as AI-written?

Generic content is what gives a resume away: bullets that describe a category of work instead of a specific piece of work. In Resume Genius's 2026 survey, 76% of hiring managers said AI-written resumes make it harder to understand what a candidate actually did. "Repetitive or overly generic language" was the top giveaway, at 44%.

Generic content is any sentence that would still be true if you swapped in another candidate's name. "Drove cross-functional alignment to deliver strategic initiatives" fits every operations manager in the country. That's why it reads as machine-made.

Give a language model "I managed databases" and nothing else, and it writes the most probable sentence about managing databases. You get fluent, average copy. Compare:

- **Generic:** "Spearheaded database modernization."
- **Specific:** "Postgres 14 migration, 2TB, zero downtime."

Deleting "spearheaded" isn't what makes the second version better. It's better because it names a system, a scale, and a constraint, and no model could have guessed those.

## Why did AI polish stop helping applicants?

Once every applicant's grammar is clean, grammar stops separating candidates. In the Tilburg study, AI-assisted cover letters scored higher on average, but they did not lead to more interview invitations.

The Tilburg authors (Kian Abbas Nejad, Giuseppe Musillo, Till Wicker and Niccolò Zaccaria) found that AI mostly polishes the generic sections of a letter, while recruiters focus on personal motivation and clarity. AI helped weaker writers the most. That made it harder for recruiters to identify the best candidates, because the floor rose to meet the ceiling.

Field data says the same thing. When Freelancer.com rolled out its "AI Bid Writer," 62% of eligible users tried it at least once. The paper "Signaling in the Age of AI" found that after the tool arrived, cover letters became less correlated with hiring. Tailoring to the specific job still predicted callbacks and awards.

Greenhouse CEO Daniel Chait described the endgame: "as the number [of applications] goes to infinity, the value [of each application] is going to go to zero." Polish is the part that went to zero.

> Polish is free now. The only thing that still costs something is a fact only you could know.

### Making clean writing messier backfires

Don't add typos or clunky transitions to look human. Tilburg found that candidates who already write strong letters gain little from AI and may even be disadvantaged if recruiters get suspicious of AI use. Making your writing worse adds risk without telling the reader anything new.

## Why do recruiters only give you 20 seconds?

Recruiters skim because their workload roughly quintupled while their teams shrank. Greenhouse's North America benchmark shows annual applications per recruiter climbing from 146 in 2022 to 746 in 2025.

> **746** Annual applications per recruiter in North America, 2025
>
> Up from 146 in 2022, a 411% rise in Greenhouse's benchmark data.

The 411% figure is specific to North America. Greenhouse's global preview uses a different series, from 1,610 to 4,890 (+203.7%). Both point the same direction. Around the same period:

- Applications per job rose 111%, from roughly 115 in 2022 to 244 in 2025
- Recruiting teams shrank 55% since 2022
- Ophir Samson, Greenhouse's Head of Voice AI, told WIRED about two thousand applicants arriving in twenty-four hours for a single role

Who is doing that reading? Refolk's index of professional profiles shows the recruiter supply behind those numbers:

| Metric | Figure | Source |
|---|---|---|
| US Recruiter / Technical Recruiter / TA Partner titles | 122,703 | Refolk's index |
| UK, same titles | 9,384 | Refolk's index |
| US-to-UK recruiter multiple | ~13.1x | Derived |
| Top US title in a 25-profile sample | Technical Recruiter (9 of 25) | Refolk's index |
| Top UK title in a 25-profile sample | TA Partner (13 of 25) | Refolk's index |
| Applications per recruiter, 2025 (N. America) | 746 | Greenhouse benchmark |

The title samples are small, so treat them as directional. Still, the most common US title being Technical Recruiter tells you something about the reader. Technical recruiters are often not engineers, so they match on named systems they can check against the req. They can match "Postgres 14." They can't verify "modernization," so they skip past it.

That's the math behind the silence. Four in five applicants in Greenhouse's survey were rejected without a single conversation, and at 746 applications a year per recruiter, a lot of good candidates never get a real read. You can rewrite every bullet perfectly and still be one of 2,000 files that showed up the same day.

## Was generic resume language invented by ChatGPT?

No. Generic phrasing was a habit long before language models, and senior people used it too. AI just made it cheaper to produce.

> **69,914** US professionals with "results-driven" in their headline
>
> Refolk's index. Roughly one for every two US recruiters.

In a 25-profile sample of those headlines, the top titles included CFOs, a CTO, co-founders and an SVP Chief Medical Officer. These are people with real, specific accomplishments who still describe themselves with the most interchangeable phrase in hiring.

That changes how you should read the "sounds like AI" worry. AI didn't invent the problem. It scaled a habit people already had, until recruiters started noticing it everywhere. The fix is the same one that worked before ChatGPT: say what you did, to what, at what size, under what limit.

## How do you make a resume not sound like AI?

Give every bullet four things: the system you used, who the work was for, a number, and the constraint you worked under. Once those are in place, the adjectives usually have nothing left to do.

Work through each bullet in this order:

1. **Name the system.** Use the exact product name and version, spelled the way the posting spells it: "Postgres 14," not "relational databases." A recruiter can match a name against the req. They can't match a category.
2. **Name who the work was for.** That might be a client, an internal team, a region, or a product line. If an NDA covers it, describe it instead: "a top-10 US regional bank."
3. **Add a number.** Scale, dollars, time saved, volume, headcount. One per bullet is enough.
4. **State the constraint.** A deadline, a budget, a two-person team, a compliance regime, zero downtime. Constraints make a bullet believable because nobody invents them.
5. **Cut the adjectives last.** After steps 1 to 4, "strategic," "innovative," and "spearheaded" usually look silly next to the facts, and you'll delete them on your own.

Here's what that looks like across roles. Brackets mark where your own tools and figures go:

| Role | Generic | Specific |
|---|---|---|
| Engineer | Spearheaded database modernization | Postgres 14 migration, 2TB, zero downtime |
| Analyst | Leveraged data to drive insights | Rebuilt churn model in [tool]; flagged [N] at-risk accounts for [team] |
| Marketer | Drove multi-channel growth | Cut paid search cost per lead [X]% by moving [Y]% of budget to [channel] |
| Operator | Streamlined finance processes | Moved month-end close from [X] to [Y] days in [system] with a two-person team |

If you draft with ChatGPT or any other model, the same rule applies to your prompt. Give it the facts before you ask for prose. A model can't add nouns it was never given, so a prompt with no facts gets you average copy. This is the gap [Refolk](https://www.refolk.ai/candidates) closes: it writes your resume from your own work history, so the systems, clients, and numbers come from your record, not from the model's guesses. Then you paste a posting and get your resume back rewritten for that job, with the tools the posting names moved up front wherever you've actually used them.

## How do you fix a cover letter that sounds like ChatGPT?

Replace the generic motivation paragraph with one specific reason you want this job at this company, tied to something in the posting, and back it with one proof point. Tilburg found recruiters focus on personal motivation and clarity. Those are the two things AI polish doesn't supply.

A cover letter sounds like ChatGPT when the opening could go to any company: "I am excited to apply for this opportunity to leverage my skills." A three-part structure fixes it:

- **The reason:** one line about something specific in the posting or the company's work that matches a problem you've actually solved.
- **The proof:** one bullet-grade sentence with a system, a number, and a constraint.
- **The ask:** a short, plain close. Leave out the summary of everything above.

The Freelancer data backs this up. Tailoring to the job content kept predicting callbacks even after AI made generic letters cheap. If you can't come up with a specific reason, that tells you something about the job. Refolk drafts the cover letter from the same history as your resume and scores how well you actually fit the posting, so you can spend your best letters on roles where the reason is real.

## FAQ

### Do recruiters reject AI-written resumes?

Not because a model wrote them. Recruiters in the Tilburg study were no better than chance at spotting AI cover letters, and TopResume's reviewers picked the AI resume only 33.5% of the time against a 25% guessing baseline. What gets resumes passed over is generic content. In Resume Genius's survey, 76% of hiring managers said AI-written resumes make it harder to see what the candidate did. The real problem is vague bullets, whoever wrote them.

### Will an ATS flag my resume as AI-generated?

No. A StylingCV audit found every major applicant tracking system parses and scores AI-written resumes the same way as human-written ones. That's one more reason to name your systems exactly as the posting does, since a named tool is something the software can match and an adjective is not.

### Should I remove em dashes and words like "spearheaded"?

You can, but it isn't the fix. I found no survey data showing em dashes are a meaningful rejection trigger. The top giveaway hiring managers cite is repetitive or overly generic language, at 44%. A bullet with "spearheaded" and a real system, number, and constraint still tells a recruiter more than a dash-free bullet that says nothing.

---

_From the Refolk job search blog. More posts: https://www.refolk.ai/candidates/blog/index.md_
