Which words give away an AI-written cover letter?

What did the new study find about AI writing tells?

A study published at the start of October 2026 catalogued roughly 13,000 words and phrases that show up at least twice as often in AI-written prose as in human writing. The method was clean: take 10,000 articles written before ChatGPT existed, summarise each one, then ask current frontier models to rewrite the article from that summary. What's left is the model's own voice. Claude Opus 5.5's loudest single tell turned out to be the word "dependable", appearing 23 times more often than in the human control group. Its strongest habit is explaining significance — "this matters" cropped up 116 times more often than people write it, and "why X matters" 92 times more often.

Other models leave different prints. OpenAI's Astra leans on what the researchers call corrective framing: telling you something is "not simply" one thing, or nudging you to act "rather than relying on" something else. Those constructions appeared over 100 times more often in Astra's prose than in human writing, next to soft hedges like "can provide" and "may provide". The em dash, meanwhile, has quietly been retired — Opus 5.5 uses it 99% less than Opus 5 did, Astra uses it 88% less than human writers, and Gemini 3.1 Pro has almost eliminated it. So the old party trick of counting em dashes to catch a bot is already obsolete.

Here's why a word-frequency study belongs in a job-search article. That vocabulary is sitting in thousands of cover letters right now, because the letters were drafted in a chat window and sent with the phrasing untouched. A recruiter needs no detector to react to it; they've read "I am writing to express my strong interest" four hundred times this quarter and they know what usually follows. The research hands you something better than a gut feeling: a concrete, checkable list of the words that mark your letter as machine-made, and a principle you can apply in ten minutes. One researcher's warning is worth holding on to — labs keep killing the famous tells, and new ones keep surfacing with every release.

Here's why a word-frequency study belongs in a job-search article. That vocabulary is sitting in thousands of cover letters right now, because the letters were drafted in a chat window and sent with the phrasing untouched. A recruiter needs no detector to react to it; they've read "I am writing to express my strong interest" four hundred times this quarter and they know what usually follows. The research hands you something better than a gut feeling: a concrete, checkable list of the words that mark your letter as machine-made, and a principle you can apply in ten minutes. One researcher's warning is worth holding on to — labs keep killing the famous tells, and new ones keep surfacing with every release.

Which words give away an AI-written cover letter?

The giveaways cluster in four groups: significance signposts ("this matters", "why this matters", "underscores the importance of"), character adjectives with no evidence behind them ("dependable", "passionate", "results-oriented", "detail-oriented", "proven track record"), corrective framing ("not simply a role", "rather than relying on spreadsheets"), and hedged verbs ("can provide", "may offer", "aims to ensure"). Add the stock vocabulary any recruiter can recite in their sleep: leverage as a verb, seamless, robust, delve, navigate, elevate, spearhead, dynamic. None of these words is wrong on its own. The problem is density. Four of them in one paragraph reads as a template, and a template reads as someone who didn't bother.

Sentence patterns betray you faster than single words. Watch for the rolling three-clause rhythm, where every sentence lands the same way: a short setup, a comma, a middle clause, a comma, then a tidy summarising phrase. Human paragraphs lurch. Yours should have a six-word sentence next to a twenty-eight-word one. The other fingerprint is the inflation construction — "this role is more than a job, it's a chance to grow" — which models still produce constantly even after dropping the older "it's not X, it's Y" formula. If a sentence could be pasted into any application for any employer, it's doing nothing except filling space a recruiter wanted facts in.

Register mismatch is the tell people miss entirely. A graduate three months out of a bootcamp writing about "driving transformative organisational change across the enterprise" is flagged instantly, not because the grammar is off but because nobody at that stage talks that way. The same goes the other way: a head of operations whose letter reads like a careers-service handout. Your letter should sound like the person who walks into the interview, using the words you'd use on a call. One practical test — read the letter out loud to someone who knows your job. If they stop you and say "you'd never say that", delete the sentence and tell them the story instead. Write that down.

Register mismatch is the tell people miss entirely. A graduate three months out of a bootcamp writing about "driving transformative organisational change across the enterprise" is flagged instantly, not because the grammar is off but because nobody at that stage talks that way. The same goes the other way: a head of operations whose letter reads like a careers-service handout. Your letter should sound like the person who walks into the interview, using the words you'd use on a call. One practical test — read the letter out loud to someone who knows your job. If they stop you and say "you'd never say that", delete the sentence and tell them the story instead.

Can recruiters really tell, and does it cost you the job?

They catch generic writing far more reliably than they catch AI. Industry surveys of hiring managers put outright rejection for suspected AI content at roughly one in five, with much higher numbers — around half in some samples — when the question specifies unedited or unpersonalised output. Treat the spread with care: most of these figures come from self-reported surveys run by careers companies, not peer-reviewed research, and the wording of the question moves the answer enormously. The consistent finding across all of them is the useful one. Hiring managers aren't anti-AI; plenty use it to screen applications themselves. They're anti-generic, and they reject letters that could have been sent to anyone.

Blind tests puncture the confidence. Recruiters overwhelmingly say they can spot a machine-written letter, yet when they're handed a mixed pile with no labels, accuracy drops close to a coin flip. Detection software fares no better. Tools that advertise near-perfect accuracy miss badly on real applications, and false positives fall hardest on non-native English speakers and anyone with a formal, structured writing style — which describes most people writing a cover letter. Here's the contrarian part: don't write for the detector. Chasing a low AI score produces mangled sentences and deliberately clumsy grammar, and it optimises for a tool nobody on the hiring side trusts. Write for the human who has forty letters to read before lunch.

One myth worth killing: the applicant tracking system isn't checking whether a model wrote your letter. An ATS parses text, matches keywords, titles and dates, and routes your file — detecting authorship isn't in its job description. Where AI suspicion bites, it bites a human reader, usually in the first four lines. That's also where the fix pays off fastest. If you want a second opinion on whether your wording reads as machine-made before you send it, a check for AI-sounding phrasing on your CV and summary will surface the same habits that show up in your letter, because most people paste both from the same chat window.

How do you strip the AI fingerprint in one editing pass?

Run a find-and-replace pass with a fixed search list, then a read-aloud pass. Search your draft for: dependable, this matters, why this matters, passionate, proven track record, results-oriented, detail-oriented, dynamic, leverage, robust, seamless, delve, navigate, elevate, spearhead, "I am writing to express", "not simply", "rather than relying on", "is more than", "deeply", "truly", "ensure", "in today's". Delete each hit rather than swapping in a synonym — a thesaurus pass just moves the tell one word sideways. Where the deleted phrase was carrying a claim about you, replace it with the evidence that claim was standing in for: a number, a tool name, a deadline, a named outcome your interviewer could ask about.

Then fix the rhythm, because vocabulary alone won't save a letter that breathes like a machine. Break every third sentence in half. Start one sentence with "I" and the next with a time or a place, so your openings stop repeating. Cut adverb-adjective pairs — "highly motivated", "extremely passionate", "consistently reliable" — down to nothing; they add syllables, never information. Replace abstract nouns with the concrete thing they describe: "stakeholder engagement" becomes "the weekly call with the warehouse team". Read the whole letter aloud at normal speed. Any sentence where you run out of breath or stumble gets rewritten, and any sentence that bores you will bore a recruiter twice as fast.

Here's the pass in miniature. Before: "I am a dependable, results-oriented professional with a proven track record of leveraging data to drive seamless process improvements — and that's why this matters for your team." After: "In my last role I rebuilt the returns report in SQL, which cut the Monday reconciliation from four hours to twenty minutes. Your job ad mentions the same backlog." Same length, nothing invented, every claim checkable. If your CV summary is carrying the same template vocabulary — and it usually is — the free CV analysis scores clarity and structure alongside ATS compatibility, so you can see whether the wording is doing any work before you copy it into another letter.

How should you use AI on a cover letter without sounding like one?

Use AI for everything except the sentences. Models are genuinely good at three jobs here: pulling the real requirements out of a messy job ad, interrogating you until you produce specific evidence, and checking structure — does paragraph two actually answer the ad's main ask? They're weakest at the part everyone delegates, which is phrasing, because phrasing is where the training distribution shows through. So invert the normal workflow. Let the assistant build the skeleton and stress-test your examples, then write each sentence yourself in your own words. Twenty minutes of your own typing beats an hour of prompt engineering, and it's the only version that survives contact with an interviewer who asks you to expand.

Two prompts carry most of the value. First: "Here's the job ad and my CV. List the five requirements I have the weakest evidence for, and ask me one question about each." Answer those out loud, in a voice note if it helps, then type the answers in roughly as you said them. Second: "Read my draft and flag any sentence that could appear in an application for a different company." That single question deletes more filler than any rewrite request. Avoid asking for "a more professional tone" — professional tone is exactly the register that produces dependable, robust and seamless three sentences apart. Ask for shorter, blunter, and more specific instead.

Where HRLens fits is after the thinking, not instead of it. An analysis of your CV scores clarity and structure, ATS compatibility and impact out of 100 for free, and the paid Full Analysis at $3.99 adds the fix list, your blind spots and the reasons you might be getting rejected — which is usually where people discover their evidence is thin rather than their adjectives. For letters, the AI cover letter writer builds from the role, company and industry you give it, and All-Access subscribers get a CV-aware version on the analysis results page. Use whichever draft it produces as a frame, then put your own sentences in it.

Frequently asked questions