Can recruiters tell if you used ChatGPT on your resume? The 7 AI tells (and the fix)

65 percent of candidates use AI somewhere in the application process. Hiring managers know this. The question that matters is not "can I use AI" (you can, every recruiter quoted below uses it themselves), but "what does AI-slop look like, and how do I avoid producing it." That is what this post is about.

We build an AI resume tool (the free CV score and the Glow Up rewrite), so we are not neutral. The advice below applies whether you use ours, ChatGPT, Claude, or paste from anywhere else. What recruiters spot, and how to fix it.

The ATS does not detect AI (and the detectors are unreliable)

Two myths to clear before the rest of the article makes sense.

Myth 1: the ATS flags AI-written resumes. It does not. Jobscan confirmed directly with Workday, Greenhouse, and Lever, each spokesperson said there is no feature that detects AI-generated writing in any of those systems. iCIMS, SAP SuccessFactors, Oracle Taleo, Workable and BambooHR also do not run AI detection at parse time. Greenhouse Real Talent (launched June 2025) and Workday HiredScore (acquired 2024) are matching systems, not provenance detectors.

Myth 2: AI detectors catch AI text reliably. They don't, especially not on short, listy, bullet-heavy text. GPTZero requires 250 characters minimum and drops to about 62 to 88 percent real-world accuracy depending on content type. Sapling acknowledges higher false-positive risk on "short, polished, or formal" writing, which describes every resume. Winston AI's own model increases false-positive rate by about 5 percent on bullet-heavy resumes. OpenAI shut down its own classifier in July 2023 after correctly classifying only 26 percent of AI-written text. Worst of all: a Stanford study of seven detectors on 91 TOEFL essays found a 61 percent average false-positive rate on non-native English writers, with 97 percent of essays flagged by at least one detector. If your company runs AI detectors at all, it's a coin flip on AI and a coin flip on whether your applicants even speak English natively.

The detector industry is not catching AI. Hiring managers are, with their eyes, in 20 seconds.

The 7 tells recruiters spot

These are the patterns named, repeatedly, in 2024-2026 recruiter interviews and surveys. Each one has a concrete swap. The fix for AI slop is not "use less AI," it's "rewrite the seven tells."

1. The vocabulary cluster: delve, robust, spearheaded, leveraged

A 2025 Florida State University study found words like delve, robust, pivotal, intricate, meticulously, commendable spiked over 50 percent in published writing after ChatGPT launched. Independent analysis of 14 million PubMed abstracts (Kobak et al., 2024) found delves appearing 25 times more often than baseline, showcasing 9 times more, underscores 9 times more. These are now AI tells in any document, but especially on resumes, where space is tight and every word gets weighted.

The retire-these list:

• Verbs: spearheaded, orchestrated, leveraged, harnessed, embarked, navigated, showcased, delved, fostered, augmented
• Adjectives: robust, intricate, comprehensive, meticulous, dynamic, seamless, synergistic, holistic, scalable, transformative, results-driven, detail-oriented, passionate
• Nouns: realm, tapestry, landscape, paradigm, testament, ecosystem, synergy

The fix: every weak verb becomes a concrete one. Built, shipped, closed, wrote, sold, hired, fired, replaced, rewrote, cut, doubled. If a verb could appear in any industry on any CV, it's the wrong verb.

2. The three-adjective tricolon

The single most distinctive AI tic. Three abstract adjectives stacked before a generic noun: dynamic, results-driven professional, passionate, detail-oriented self-starter, innovative, scalable, mission-critical solutions. Humans rarely write this way. RLHF-trained language models do it constantly.

Spot it by counting adjectives in a row. Two is fine, three plus and you're reading AI output. The fix: pick one adjective and make it concrete, or drop the modifiers entirely and lead with the noun.

3. The identical bullet skeleton

AI generates bullets by template: action verb plus what-you-did plus metric plus outcome, every single time. Over a full resume, the rhythm becomes obvious. Every line has the same length, the same structure, the same beat. Gem and Resume Pilots both name this as the easiest tell on the recruiter side because it's visible in two seconds of vertical skim.

The fix: vary structure deliberately. Lead some bullets with the result, some with the verb, some with the scope. Vary length. A three-line bullet next to a one-line fragment reads more human than five identical seven-word bullets.

4. Round-number metrics (the plausibility problem)

Bonnie Dilber at Zapier says she sees 25 percent of applications come through with metrics ending in 0 or 5. Increased revenue by 50 percent. Reduced cost by 30 percent. Grew team by 10. Real numbers don't cluster like that. Real numbers are 47.3 percent, $1.2M, 11.4 points. The rounded versions read as fabricated because they are: an LLM defaulted to a plausible-sounding figure because no actual measurement informed the answer.

The fix has two parts. First, use real numbers when you have them, including the awkward ones. Second, when you only have rough estimates, use ranges ("saved 12 to 18 hours per week") or non-numeric scale signals ("cut a manual weekly process to automated, freeing one analyst-day"). Ranges and before-after states are honest. Round percentages with no source are not.

5. Executive voice on a junior CV

The Interview Guys, Resume Pilots, and Willo all name the same tell: when an entry-level candidate writes about "spearheading go-to-market strategy" or "orchestrating cross-functional alignment," the voice doesn't match the experience. A new grad who sounds like a CMO is using AI. A 12-year veteran who sounds like a new grad is too, just from the other direction.

The fix: write in your own register, then ask AI to tighten, not generate. If you wouldn't use the phrase in a 30-second conversation with a friend in the same field, it doesn't belong on the resume.

6. Leftover placeholder text

Tejal Wagadia, a tech recruiter, says she still sees ChatGPT output pasted in raw, with the parentheses left in: (add company name here), [insert mission statement], {insert metric}. Bonnie Dilber says she sees opening lines like "Company's mission of [insert mission statement] resonates with me" multiple times a week. This isn't even an AI tell, it's carelessness, but it instantly gets the application rejected.

The fix is read-aloud. If you can't pronounce a sentence cleanly, the brackets are still in.

7. The interview mismatch (the failure mode that hurts most)

Even if all six tells above are fixed, there's a seventh that surfaces later: when the resume reads at one seniority and the phone screen reveals another. Curtis Britt at Korn Ferry says about 5 to 10 percent of job offers are rescinded due to resume dishonesty discovered post-interview. That number is rising, not falling, because recruiters in 2026 are running "forensic follow-up questions designed to expose the delta between what a resume claims and what a candidate actually did" (Interview Guys, 2025).

The fix is the SOAR test. For every bullet on your CV, can you narrate the Situation, Obstacle, Action, and Result out loud, in three minutes, with details? If not, delete the bullet. Not because it's false (it might be true), but because the interview will catch you unprepared on something that was your own work.

The numbers, dated and sourced

• 88 percent of hiring managers say they can tell when AI was used (Insight Global 2025 AI in Hiring Survey, 1,005 HR managers).
• 80 percent say they can often tell from the resume itself (Resume Genius 2025, 1,000 hiring managers).
• 33.5 percent can spot an AI resume in under 20 seconds (TopResume 2025, 600 hiring managers).
• 19.6 percent would reject an AI resume outright. 62 percent reject if not personalized (TopResume + Resume Now 2025).
• 67 percent of HR leaders say AI applications are slowing hiring; 20 percent report delays of two weeks or more (Robert Half, November 2025, 2,000+ HRs).
• 65 percent of candidates use AI somewhere in the application process (Career Group Companies 2025 Market Trend Report).

The pattern is consistent: recruiters use AI themselves (99 percent in the Insight Global survey), recognize AI slop on sight, and reject it not for being AI but for being slop.

When AI is the right move (most of the time)

Tessa White, a senior recruiter quoted in CNBC's April 2025 piece, calls it hypocritical for recruiters to criticize candidates for using AI. Harvard's Mignone Center for Career Success officially supports AI as a revision tool: "generative AI can be useful in the editing process, for example, helping you brainstorm revisions to bullet points or incorporate keywords from a job description. However, generative AI should not be the primary author, not least because its output will likely be very generic."

The four legitimate uses, in order of value:

1. Keyword gap analysis against a specific job description. AI is great at this. It is what every resume scanner does internally.
2. Tightening bullets you already wrote. Paste in your weak bullet, ask for three sharper variations, pick the one closest to your own voice. Then edit it again by hand.
3. Translating the same experience for a different audience. Same achievements, different industry vocabulary. AI saves you 20 minutes of thesaurus work.
4. Structural critique. Does this resume have a clear narrative? Are the dates consistent? Is the seniority arc obvious? AI is a fine reader.

What AI cannot do well: generate experience you don't have, invent metrics, judge what to leave off, or speak in your voice. That work stays yours.

The humanize checklist (run this before submitting)

Eight passes to strip the seven tells out of an AI-assisted draft. None of them take more than five minutes.

• Read every line aloud. If a sentence doesn't sound like you, rewrite it. Two-pass minimum.
• Cut every third clause. AI bloats. Real writing is tighter.
• Replace every retire-list word (spearheaded, leveraged, robust, comprehensive, etc.) with a specific verb or delete the sentence.
• Vary sentence length. Short fragment. Then a longer clause that adds context and a number. Short again. Mix the rhythm.
• Check every number. Anything ending in 0 or 5 gets a sanity check: is this the real number or a plausible-sounding round-up? If the latter, replace with a range or a non-numeric scale signal.
• Adjective audit. Search the document for two-adjective stacks. Cut all three-adjective stacks. One concrete adjective is fine, three abstract ones is AI.
• SOAR-test every bullet. If you can't narrate the Situation, Obstacle, Action, and Result for three minutes from memory, the bullet isn't ready.
• Search for brackets: [, {, (insert. The number of resumes submitted with placeholder text still in them is depressingly high.

Related reading

If you're running the AI workflow on your own, our ChatGPT resume prompts post has 15 specific prompts that respect the seven tells above. Our AI resume enhancer guide covers the broader question of what a good AI resume tool actually does, and the bullet point examples give you 80 plus role-specific bullets that already pass the SOAR test.

Key takeaways

• The ATS does not detect AI. Workday, Greenhouse, Lever, iCIMS, Workable all publicly confirm there is no AI-detection layer in their systems.
• AI detectors are unreliable on resume-length text. GPTZero, Sapling, Winston AI all admit higher false-positive rates on short, formal, bullet-heavy text. Stanford found 61 percent false-positive rate on non-native English writers.
• 80 to 88 percent of hiring managers say they can tell from the writing itself. 33.5 percent in under 20 seconds. The detection is human, not algorithmic.
• The 7 tells: AI vocabulary (delve, spearheaded, robust), three-adjective tricolons, identical bullet skeletons, round-number metrics, executive voice on junior CV, leftover placeholder text, interview mismatch.
• Each tell has a concrete swap. The fix is not less AI, it's rewriting the slop.
• Legitimate AI uses: keyword gap analysis, tightening existing bullets, vocabulary translation, structural critique. AI cannot generate the experience itself, only sharpen the description.
• The humanize checklist: read aloud, cut every third clause, replace retired words, vary sentence length, check every number, audit adjectives, SOAR-test every bullet, search for placeholder brackets.

If you want the seven tells caught for you, the free CV score flags vocabulary and structural risks in 90 seconds, and Glow Up rewrites the bullets without producing the patterns above. Same input, no slop, same voice.