The AI-Written Resume Tell Reference, and the Human Rewrite for Each

You drafted your resume with AI, and now you need to strip the parts that make a recruiter think "a machine wrote this" before you send it, without losing the keyword match that gets you past the parser. This is a lookup document for exactly that edit: a table of the words, phrases, and structural patterns that read as machine-generated to a human reviewer, a tested human rewrite for each, and a note on which so-called tells are false positives worth keeping. Jump to the row you need, fix the line, and reconcile it against your keyword score.

The load-bearing fact, and the one most listicles get wrong, is this: recruiters do not run AI detectors, and no major applicant tracking system detects authorship. What people react to is genericness. So the job is not to evade a machine. It is to make every line sound like you, specifically, while holding the keyword match.

What actually flags a resume as AI-written to a human

A human reviewer flags genericness, not authorship. In an Insight Global survey, 53 percent of hiring managers said they can tell when a candidate used AI, but what they are recognizing is a generic resume: the same tell-words and the same unquantified claims arriving in wave after wave of near-identical applications.

This distinction changes everything about how you edit. If the problem were specific words, you would delete them and be done. But the problem is a texture: interchangeable bullets, capability statements instead of outcomes, and uniform structure. One recruiter at Zapier reports that nearly 25 percent of the resumes she sees are clearly AI-written. She cannot recite a banned-word list. She pattern-matches on interchangeability - a bullet that would fit on a stranger's resume without a single change.

The mechanism matters because it tells you where to spend effort. Genericness is what remains when specifics are absent. Adding one real number - a team size, a latency figure, a named tool version - breaks the pattern faster than deleting ten flagged words. The disqualifier survives even when AI use is disclosed: only 54 percent of managers said they would even care whether a resume was AI-written, and 99 percent said they use AI in hiring themselves. What gets rejected is the low-effort signal, and genericness reads as "did not care."

The word-level tell reference

Most flagged words are ordinary human vocabulary, and the tell is clustering, not the individual word. Below is the reference table: the word, the plain human rewrite, and the verdict on whether it is a genuine machine signature or a false positive you can keep when it is isolated and true.

The em dash earns its place at the top. In a test where a writer generated 500 ChatGPT resumes, the em dash appeared in 92 percent of them. Meanwhile the words the internet loves to ban from chatbot essays - "delve" and "tapestry" - appeared zero times. Those are essay-era myths, not resume tells. Do not waste an edit on them.

The plain-verb swaps are the safest fixes because they lose nothing. "Utilize" becomes "use." "Leverage" becomes "use," or better, name the thing you used it for. These words appear in ChatGPT output at rates five to ten times higher than in human writing, so removing them shifts the density signal without touching your meaning.

Why the most-flagged words are usually false positives

The words listicles tell you to delete are, in most cases, long-standing human resume vocabulary. This is the single most important correction to the standard advice, and it comes straight from Refolk's index of professional headlines.

In Refolk's index, "results-driven" sits in 68,248 US professional headlines and "spearheaded" in 661. Both are human vocabulary. Deleting "spearheaded" on the theory that it is a machine word is a false positive: the word appears in hundreds of real headlines written by people who never touched a chatbot. The verdict rule is simple. Is the word clustered with three or more other tell-words, and is it unquantified? Then it is filler and you replace it. Is it isolated and describing something you actually did? Then keep it.

This advice is also market-specific, which no listicle mentions.

The same word is roughly 11 times more prevalent in US headlines than in UK ones. A tell list written for the US market overstates the risk for a UK reader, and the reverse holds too. Before you delete a word because a blog flagged it, sanity-check whether it is actually common in your market. If tens of thousands of real professionals in your region use it, it is not a machine signature.

The structural tells that survive a word swap

Structure gives you away even after every flagged word is gone. The documented signature is uniformity: bullets that all start with the same verb, all run the same length, and all follow the same grammatical skeleton - verb, adjective, noun, prepositional phrase - repeated line after line.

This is the tell that a punctuation hunt misses entirely. You can remove every em dash and still have a resume where every bullet reads "Spearheaded cross-functional initiatives to drive measurable growth." The fix is rhythm, not vocabulary. Vary the length. Start some bullets with the object instead of the verb. Let one bullet be short and blunt.

The deeper structural tell is capability-statement phrasing over outcomes. AI drafts describe what a role generally involves rather than what you specifically did, so a bullet reads like a job description instead of an accomplishment. "Responsible for managing stakeholder relationships across the organization" is a capability statement. "Convinced three regional VPs to standardize on one reporting stack, cutting month-end close from nine days to four" is an outcome. The second one cannot be pasted onto a stranger's resume, which is the whole point.

The matrix carries the counterintuitive rule: a specific bullet that still contains "spearheaded" beats a word-clean bullet that could belong to anyone. Reviewers forgive a mild tell-word attached to a concrete accomplishment. They do not forgive a line that fits a hundred candidates.

The read-aloud and stranger-swap tests

Two named checks catch the pattern faster than any word list. Read three bullets out loud. If any one of them could be pasted onto a stranger's resume without changing a word, it reads as AI. That is the stranger-swap test, and it is more reliable than scanning for vocabulary because it targets the actual signal - interchangeability.

The read-aloud test is its companion. Machine-drafted prose has a rhythm that sounds fine on the screen and wrong in the mouth. When you read a bullet aloud and it sounds like something no human would say in an interview, it needs work. The repair is almost always the same: cut the adjectives, keep the numbers. "Cut onboarding time 30 percent" beats "results-driven professional dedicated to operational excellence" every time, and it survives the interview because it is true.

1. Name the object: what exactly did you touch? (a named tool, feature, team, or process)
2. Name the change: what measurably moved? (a number, a before/after, a time saved)
3. Name the constraint: what made it hard? (a deadline, a legacy system, a budget cut)

Before: Leveraged cross-functional collaboration to streamline key processes and drive results.
After:  Rebuilt the returns workflow across support and warehouse ops during a hiring freeze, cutting refund turnaround from 6 days to 2.

The "before" line fits any operations role at any company. The "after" line fits exactly one person. That is the difference a reviewer registers in half a second, and no humanizer tool produces it, because the tool does not know your returns workflow. You do. This is the part of the edit that only you can supply, which is also why running your own history through Refolk - which drafts from your actual work rather than from a template - starts you closer to the "after" line than a blank chatbot prompt does.

The reconciliation: keeping keyword match while cutting tells

This is the core tension, and the one the listicles skip because they are selling a humanizer, not a scanner. Removing tells can lower your keyword match, and forcing keywords back up recreates the tells. Jobscan recommends a match rate of at least 75 percent, with 80 percent as ideal. The trap is pushing higher.

Pushing a match score above 80 to 85 percent almost always makes the resume worse for human readers. The keywords that take a resume from 72 to 88 percent are typically the ones already present in context in a well-written bullet, just not in the exact phrasing the job description uses. Adding them as standalone lists or forcing them into awkward constructions creates the generic pattern recruiters identify immediately. A resume at 74 percent that tells a coherent story of specific accomplishments gets called. A resume at 88 percent that reads like a keyword list gets rejected at that stage.

The 75 to 80 percent band is the only place both gates open. Below it the parser filters you out. Above it you generate the texture a human flags. Scan first, learn which keywords are load-bearing, keep those in real sentences, and stop chasing the score once you clear 75 percent. Keyword stuffing has its own penalty: SHRM research found 43 percent of recruiters say it makes candidates look dishonest.

The step-by-step humanizing procedure

Work in this order. The order matters because scanning before editing tells you which keywords you cannot afford to lose, and testing after editing confirms you did not lose them.

There is a live disagreement worth naming: some sources say fix the bullets first, since that is where recruiters look, while others start with the summary. Either order works. What does not work is skipping the specificity step, because that is the step that actually removes the signal.

Where this goes wrong: the false positives

Most damage comes from over-correcting. Here are the failure modes, each with the check that catches it before you send.

• Deleting a word that was never a tell. Cutting "spearheaded" or "results-driven" because a listicle said so, when they appear in hundreds or tens of thousands of real headlines. Check: is the word clustered with three or more others and unquantified? If it is isolated and true, keep it.
• Chasing the ATS score into a keyword list. An 88 percent match that reads as a template. Check: re-read after every keyword you add. If a bullet no longer sounds spoken, revert it.
• Trusting an AI detector to "clean" the resume. A real human bullet flagged as AI - detectors false-flag 61 percent of non-native-speaker essays. Check: never edit to satisfy a detector score. Edit for specificity instead.
• Hunting the em dash while ignoring rhythm. Removing every dash but leaving uniform verb-adjective-noun bullets. Check: do two consecutive bullets share the same skeleton? Vary structure, not just punctuation.
• Adding a fabricated metric to sound human. AI will invent numbers like "increased revenue 40 percent" that you never claimed. Recruiters spot fake numbers fast, and they collapse in the interview. Check: every number must survive being asked about.
• Fixing the resume but leaving a mismatched profile. A polished resume next to a casual, generic online profile. Check: if the resume voice and your online presence do not align, questions arise. Align them.
• Sending the same humanized resume everywhere. "Humanized" but identical across roles. Check: is it tailored to one job? The same document sent everywhere is what "AI-written" actually smells like.

The through-line across all seven is the same insight: detection is a confidence illusion. Managers believe they can spot AI, but they are anchoring on genericness and calling it "AI." So the fix is never evasion. Evasion produces a word-clean, still-generic resume, which fails on the exact axis reviewers actually judge.

The tailoring failure deserves its own emphasis because it is invisible to you. A resume can pass every word and structure check and still read as machine-made if it is the same file you sent to twelve other postings. Tailoring to one job is not a nicety; it is the difference between a document that argues for you and one that argues for no one. If you want to see what tailored, specific accomplishment bullets look like at scale, search for them directly.

Verify before you send

Run this checklist against the final draft. Every item is a checkable statement, not a topic, and each maps to a failure mode above.

If the draft clears every item, stop editing. The evidence is that algorithmic writing help does not hurt you when the output is specific: a field experiment covering 480,948 job seekers found those who got algorithmic writing help were 8 percent more likely to be hired, with no drop in employer satisfaction. AI is a fine drafting tool. The failure is mailing the raw output. Your job is the last mile - the numbers, the named objects, and the voice that only you can supply.

Keeping the reference current

Two things in this document will drift, and both have a mechanism you can re-check yourself rather than a value to memorize. The tell-word list moves as models change their output distribution, so re-run the density check on any new draft rather than trusting a fixed list - the signal is always clustering, not the specific word of the season. And the ATS match threshold is a published recommendation that can shift, so confirm the current minimum from the scanner you use before you optimize to it, and hold the discipline of stopping at the low end of the recommended band regardless of where it sits. The one thing that will not drift is the principle underneath all of it: a specific, tailored, defensible bullet reads as human because it is human, and no reviewer and no detector can take that away from you.