AI-written CV tells: which words give you away?

What did the new AI writing tells study actually find?

A study released on 1 October 2026 found 13,000 phrases that turn up at least twice as often in AI-rewritten text as in human writing, and it named the exact words. The method was clean: researchers took 10,000 articles published before ChatGPT existed, asked frontier models to rewrite each one from a short summary, then compared word frequencies and sentence patterns across the two piles. Opus 5.5's single biggest tell was "dependable", which showed up 23 times more often than in the human samples. Its other habit is explaining significance — the phrase "this matters" appeared 116 times more often, and "why X matters" 92 times more often. So the problem isn't bad grammar. It's a fingerprint in the vocabulary.

The old tells are fading and new ones keep arriving. In those samples, Opus 5.5 used the em dash 99% less than its predecessor, and a rival frontier model cut it by 88% against human writing. The "it's not X, it's Y" construction is largely gone from Opus 5.5 too, but it still reaches for "is more than an X, it's a Y". Another lab's model leaned on hedges like "may provide" and "can provide", plus corrective framing — "not simply X", "rather than relying on X" — more than a hundred times as often as people do. The researchers' conclusion is the useful bit: training strips the famous markers and grows fresh ones, so the total count holds roughly steady.

Here's why a job seeker should care. Those phrases land in exactly two places on a CV: the professional summary and the first three words of every bullet point. Ask a model to "write a professional summary for a project manager" and you'll get character adjectives plus significance framing, because that's the shape of the training data it's pattern-matching. My own contrarian take: the em dash witch hunt is a year out of date. Deleting every dash while keeping "dependable professional committed to delivering robust results" fixes nothing, because the dash was never the tell — the evidence-free adjective was. Fix the vocabulary and the punctuation argument stops mattering.

Which AI-written CV tells show up in your summary and bullets?

On a CV, the tells cluster in four predictable spots: character adjectives with no evidence behind them, significance framing that explains why your work was valuable instead of showing it, three-item lists, and bullet openers that all share one verb shape. "Dependable and results-driven operations lead with robust stakeholder management skills" contains zero facts — no headcount, no budget, no region, no system. A recruiter reading forty CVs an hour has seen that sentence fourteen times already this morning. The second giveaway is symmetry. Human bullets are ragged: one runs eleven words, the next runs twenty-eight because the story needed it. Model-drafted bullets arrive the same length, same rhythm, same strong verb in the first position, every single time.

Triplet lists deserve their own warning. Models love "planning, execution and delivery", "accuracy, efficiency and compliance", "strategy, process and people". Three abstract nouns feel balanced to the ear, which is exactly why the pattern survives the drafting process — it sounds like writing. Count the triplets in your summary; if there's more than one, you've got a tell. The verb openers go the same way. "Spearheaded", "orchestrated", "championed" and "leveraged" are not words most people use when they describe their job out loud to a friend. Swap them for what actually happened: built, fixed, cut, negotiated, migrated, hired, shipped. Plain verbs carry a number better, and numbers are what a hiring manager scans for.

Then there's the residue, which is less a stylistic tell than an unforced error. Unedited drafts ship with square-bracket placeholders like "[insert company mission]", with a line still addressed to the wrong employer, or with metrics that all land suspiciously on a round 20%, 30%, 40%. Models invent plausible-looking numbers when you don't give them real ones, and a hiring manager who asks about that 40% in the interview will watch you stall. The deepest problem, though, is interchangeability. If your CV could be handed to any other candidate with the same job title and still read true, the writing is doing nothing for you — whoever or whatever typed it.

Does it matter if a recruiter thinks your CV is AI-written?

It matters, though not because using AI is forbidden. It matters because an unedited draft signals a candidate who couldn't be bothered to add their own facts. Surveys of hiring managers through 2025 and 2026 put the picture plainly: roughly 58% say they've received AI-generated CVs or cover letters, about half say they'd dismiss an application they identify as machine-written, and around 51% name unnatural phrasing as the main thing that tipped them off. Read that last figure carefully — phrasing, not provenance. Nobody is running forensics on your file. They're reacting to a summary that sounds like every other summary, and the rejection is really a rejection of sameness.

Two honest caveats keep this in proportion. First, those detection claims are self-reported confidence, not measured accuracy; no study has verified that suspicious recruiters are actually right. Second, automated AI detectors are unreliable, with false-positive rates that run high for people writing in a second language — so don't paste your CV into one and panic at the result. The other half of the picture is the double standard: the same employers screening you with automated tools are irritated when you draft with them. Your job isn't to prove purity. It's to make the document specific enough that the question never comes up, which is what a check for AI-sounding phrasing is for.

Keep the priority order straight, because the phrasing panic can crowd out bigger problems. A CV that reads beautifully human but never gets parsed properly loses before anyone judges your word choice. Dates in a header, skills inside a text box, a two-column layout, contact details in an image — those failures happen upstream of taste. So fix structure and parsing first, then evidence, then voice. In practice the three overlap: when you replace "dependable professional" with "managed a 14-person warehouse team across two sites, cut picking errors 31%", you've solved the keyword problem and the human-voice problem in the same keystroke.

How do you humanize an AI CV draft line by line?

Keep the draft and edit it in layers — one pass per problem, because trying to fix everything at once is how tells survive. Pass one: delete every adjective describing your character. "Dependable", "driven", "passionate", "detail-oriented" all go, with nothing put in their place. Pass two: find each sentence that explains why something mattered and replace it with the outcome. Pass three: break every triplet into either one concrete item or a full sentence. Pass four: vary bullet length deliberately — let one bullet run two lines and the next run six words. Pass five, the one that does the real work: add back the detail only you could know.

Here's the shape of it. Before: "Dependable engineer who leveraged robust CI/CD practices to deliver seamless deployments — this matters because reliability builds trust." After: "Moved our Jenkins pipeline to GitHub Actions; deploys went from 40 minutes to 9, and on-call pages dropped from about 12 a month to 3." The second version is shorter, names two tools, carries three numbers, and no model would have produced it because no model knew those facts. That's the whole trick. Specificity is the humanizer — not quirky phrasing, not deliberately clumsy grammar. Tool versions, team sizes, ticket volumes, client regions, the constraint you worked around: those are unforgeable.

Two tests finish the job. Read each bullet aloud; if you'd never say it to a colleague in the pub, rewrite it in the words you would actually use. Then ask whether anyone else with your job title could sign the sentence. If yes, it's filler. Keep your own phrasing even when it feels less polished — slightly plain and unmistakably yours beats smooth and anonymous every time. If starting from a blank page stalls you, paste your old CV into HRLens's chat-based CV builder, then push back in the chat until each bullet carries a number; it exports to PDF and Word, and Hebrew CVs come out written in Hebrew and laid out right-to-left.

How does a clarity-and-structure score catch what you stop noticing?

By the third editing pass you've stopped reading your own words — you're recognising them, which is a different mental act entirely. A scored analysis reads the file cold. HRLens's free CV analysis takes a PDF, a Word file or even a photo of a printed CV, with Hebrew OCR supported, and returns an overall score out of 100 banded as Weak, Decent, Strong or Excellent, plus five category scores: experience, tech stack, impact and ownership, clarity and structure, and ATS compatibility. It also lists your strengths and runs a visual layout analysis. Clarity and structure is the one that catches sameness; ATS compatibility catches the parsing faults your eye forgives.

Pair it with the job you're actually applying for. Paste the job description into the analysis — paste the text rather than a link for LinkedIn and Indeed, which block fetching — and you get a list of the skills the ad asks for that your CV never mentions. That list is the cure for generic phrasing, because it tells you exactly which real experience you buried under an adjective. Every analysis also generates ATS-friendly rewritten versions of your CV across six templates that mirror the ad's keywords, which is a faster route to a tailored application than editing the same master file for the ninth evening in a row.

Be clear about what's free and what isn't. The score, the five categories, the strengths and the layout analysis cost nothing. The Full Analysis at $3.99 adds the fix list, a "why you might be rejected" breakdown, your blind spots, a hiring-probability estimate in interviews per ten applications against the market, and a salary estimate; comparing a new version against an earlier analysis sits in the Premium Package at $8.99 or All-Access at $24.99 a month. My advice: run the free check first, do the five-pass rewrite yourself, then run it again. And remember the actual goal — not passing some AI detector, but being the only person who could have written your CV.

Frequently asked questions