Can Employers Tell If AI Wrote Your Cover Letter?
Ask a recruiter and you will hear yes, obviously, every time. Ask the researchers who test detection under controlled conditions and you will hear no, not reliably, not even close. Both groups are describing something real, and the gap between them turns out to be the only part of this question worth your time.
Quick Answer
- What recruiters report: 88% of 1,005 US hiring managers said they can tell when candidates use AI — though the question bundled applications, cover letters and resumes together
- What research measures: in the three main experiments of a 4,600-participant study, people identified AI-written self-presentations with 50–52% accuracy, against a 50% chance baseline
- Software does not rescue it: OpenAI withdrew its own detector in July 2023 for low accuracy, and independent detectors misclassify human writing by non-native English speakers at high rates
- The resolution: what readers reliably catch is generic, not AI. The two overlap because most AI letters are generated from the posting alone
- The fix is additive: a letter fails when it contains nothing only you could have written. You solve that by putting something in, never by taking something out
What Recruiters Say They Can See
The number that gets quoted at job seekers comes from a hiring-manager survey, and it is a big one.
88% of hiring managers surveyed said they can tell when candidates are using AI to help with applications, cover letters or resumes.
Source: Insight Global, 2025 AI in Hiring Survey Report, fielded by Atomik Research among 1,005 US hiring managers, 17–22 October 2024 (margin of error ±3 percentage points).
Read the question rather than the headline and three things become clear. It bundles applications, cover letters and resumes into one item, so it cannot tell you anything specific about cover letters. It is US-only. And, most importantly, it measures confidence, not accuracy. Nobody in that sample was handed a stack of documents and scored on how many they got right.
That distinction matters more than it sounds, because recruiters have no feedback loop. A recruiter who decides a letter was machine-written never finds out whether they were correct. The letters they correctly flag confirm the skill; the ones they wrongly flag, and the AI-assisted letters they never noticed at all, stay invisible forever. Confidence under those conditions rises regardless of accuracy. It would in anyone.
The same survey found that 54% of hiring managers would care if a job seeker applied with a resume or cover letter written by AI — a smaller and stranger number than the coverage suggests. Asked why they care, those respondents did not line up behind disapproval. The most common single answer, given by 36% of them, was that it shows expertise in AI and technology. The two objections you would expect came in behind it: 24% said it makes them think the candidate is not putting in the effort to get the job, and 17% said it feels impersonal.
Source: as above. A fourth answer, keyword optimisation for applicant tracking systems, drew 23% and is left out above because it is not an objection. The report states no base for the four shares; they appear to be shares of the subgroup who said they care rather than of all respondents.
What Detection Research Actually Measures
Now the other side of the ledger, where people are scored instead of asked.
A large controlled test of human detection of AI self-presentation was published in PNAS in 2023. It ran six experiments with 4,600 participants in total. In the three main ones — professional profiles, hospitality listings and dating profiles — participants identified which self-presentations a language model had written with 50% to 52% accuracy, on a two-way choice where guessing scores 50. Paying them more did not help. Training them with feedback did not help. The authors traced the failure to intuitive cues that people trust and that do not actually distinguish the two: the confidence was real, the signal was not.
Source: Jakesch, Hancock & Naaman, Human heuristics for AI-generated language are flawed, PNAS, 2023 (N = 4,600 across six experiments; the 50–52% figure is from the three main experiments). The limits are worth stating plainly: these were short self-presentation profiles rather than full cover letters, and the text was generated by 2022-era models. It is the best controlled evidence available, and it is not a direct test of this exact question.
Software is the obvious next hope, and a thinner one than most people assume. OpenAI shipped a classifier for detecting AI-written text in January 2023 and withdrew it six months later.
OpenAI's own classifier correctly identified 26% of AI-written text, while incorrectly flagging 9% of human-written text as AI. It was taken offline on 20 July 2023, in the company's words, “due to its low rate of accuracy”.
Source: OpenAI, New AI classifier for indicating AI-written text, update dated 20 July 2023.
Be fair about what that does and does not prove. It is one old tool, and vendors have shipped detectors since that claim better numbers. But it remains striking that the lab with the most direct knowledge of how its own models write could not build a classifier it was willing to keep online — and a 9% false positive rate, applied across an application pile, means many honest applicants accused of something they did not do.
Both Sides Are Right About Different Things
Put the two findings side by side and they look like a contradiction. They are not, because they answer different questions. The research asks: given this text, can you name the author? The recruiter is answering: given this letter, did I learn anything?
Those come apart completely. A recruiter who says “I can always tell” is, almost always, describing a letter that praised the company mission, restated the job title, listed three adjectives and closed with an eagerness to contribute. They are right that something is wrong with it. They are guessing about the cause.
The guess correlates well enough to feel like a skill, and the mechanical reason is not complicated. The common pattern of use is the same one everywhere: paste the job posting, take the first output, send it. A model given only the posting has only the posting to write about. It can describe the role in fluent, accurate, well-organised prose, and it cannot say one word about the applicant, because nothing ever told it anything about the applicant. That is not a flaw in the writing. It is an absence of input.
This is also why the advice to write “more naturally” misses. Style is not the problem, so style is not the fix. A letter can be warm, contraction-heavy and idiosyncratic and still contain nothing a stranger could not have written about you. What it needs is material: a project with a name, a constraint you actually worked under, a number you are prepared to defend, a reason for this employer that came from something you read or someone you spoke to.
A disclosure is owed at this point. This site sells a cover-letter generator, so treat the argument above as one with an interest attached and weigh it accordingly. The honest version of our position is narrow: a generator is a drafting tool, it cannot invent the specifics for you, and a letter produced from a job posting alone will be generic no matter how good the model behind it is.
The Same Posting, Twice
Here is the difference, illustrated. One job posting, two letters: the first is the shape you get from the posting alone, the second the shape you get once the applicant's own background is in the room. The only variable is what the writer had to work from.
“Customer Success Manager, mid-market. You will own a book of roughly 60 accounts, drive renewals and expansion, and work with Product on churn signals we currently spot too late. We are looking for someone comfortable running quarterly business reviews with commercial stakeholders.”
Fluent, accurate, and about nobody
“I am writing to express my strong interest in the Customer Success Manager position. With a proven track record of managing accounts and driving renewals, I am confident I would be a valuable addition to your team. I am experienced in running quarterly business reviews with commercial stakeholders and passionate about identifying churn signals early. I would welcome the opportunity to contribute to your continued success.”
Nothing in Letter A is badly written. Nothing in it is false. And every sentence of it survives being posted to another employer unchanged — swap the job title and it is a different company's letter. That is what the recruiter is reacting to when they say they can tell.
Same posting, one extra input
“You mention spotting churn signals too late. At Brightline I inherited 54 mid-market accounts where renewals were reviewed 30 days out, which was already too late to fix anything, and moved that review to 90 days with a three-question health check the CSMs could run in ten minutes. Renewal conversations stopped being negotiations about price. The part I got wrong was the first version of the health check — it measured logins, which told us about habit rather than value, and we replaced it after a quarter. I ran QBRs with commercial stakeholders throughout, usually with a finance lead in the room, which changes what the deck has to prove.”
That is the whole mechanic, and it is why the background you supply matters more than the model does. Our guide to tailoring a cover letter to the job description walks through choosing which parts of your history to feed in, and whether you need a different letter for every job covers what to keep and what to rewrite when you are applying at volume.
The Non-Native English Problem
If you write English as a second language, there is a specific unfairness here that deserves naming, and it is the strongest practical argument against employers screening on detector scores at all.
A 2023 study in Patterns ran seven widely used GPT detectors over essays that were unambiguously human-written, and the results split along one line.
Across 91 human-written TOEFL essays, the seven detectors produced an average false positive rate of 61.22%. All seven agreed in flagging 19.78% of those essays as AI-authored, and 97.80% were flagged by at least one detector. On 88 essays written by US eighth-graders, the same detectors were near-perfect.
Source: Liang, Yuksekgonul, Mao, Wu & Zou, GPT detectors are biased against non-native English writers, Patterns, 2023. The detectors tested are of that generation, and the sample is student essays rather than job applications — but the bias it identifies is a property of register, not of essays.
What that data is picking up is register, not dishonesty. Careful, formal, deliberately plain English — the register most people write in when working in a second language, and very nearly the register everyone adopts for a job application — is precisely the writing those tools scored as machine-made. Native-speaker essays written loosely came out human. The careful writing came out as a machine.
We are not going to tell you to write differently in order to move that number, and you should be wary of anyone who does. A detector score is not evidence, and no reader should be reshaping their own voice to satisfy a tool that flagged six in ten honest essays by non-native writers. The useful move is the same one as everywhere else in this piece: specific content does something a detector score cannot argue with. A named project and a real constraint are checkable facts about you. A probability score is an opinion from software that mistakes carefulness for automation.
And if your applications are not converting and you have been assuming detection is the reason, that assumption is worth testing. Our breakdown of why you are not getting interviews works through the causes that actually show up in the data, and what to write when you do not meet the requirements handles the case where the real obstacle is the posting rather than the prose.
Being Straight About It
There is a version of this question that has nothing to do with detection: not will they find out, but what is actually fine here. It has a clean answer, and the line falls on evidence rather than on tools.
- Drafting, structuring and editing with a model is ordinary tool use. So is a template, a spellchecker, a friend who reads it, or a careers adviser who rewrites your opening. Nobody has ever been expected to produce a job application unaided.
- Letting a model assert things about you is not. If a sentence claims experience you do not have, a scale you never worked at, or a result whose mechanism you cannot describe, that is a false statement on an application, and which tool wrote it makes no difference to how serious that is.
- If the employer states a policy, follow it. Some postings and application forms now say something about AI use. Some ask you to declare it; some ask you not to use it for a specific exercise. Those are conditions of applying, not suggestions.
- If you are asked directly, answer honestly. “I drafted it with an AI tool and then rewrote it around the Brightline work” is a completely acceptable answer, and it is available to you only if the second half is true.
- You are not obliged to volunteer it unasked, any more than you would volunteer which template you started from. The line is denial, not silence.
The real audit is not a detector anyway. It is the interview. Every specific claim in your letter is a question someone can ask you about for five minutes, and a letter you cannot discuss fluently is a liability whoever drafted it. That is also the cleanest reason to build it out of your own material: the letter and the interview then say the same thing, because they came from the same place. We have written separately on whether using AI to prepare counts as cheating, which runs the same evidence-based line through interview preparation.
Frequently Asked Questions
Can employers tell if your cover letter was written by AI?
Not reliably, as a technical matter. In a 2023 PNAS study (N = 4,600 across six experiments), participants in the three main experiments identified AI-written self-presentations with 50 to 52 percent accuracy, which is chance. What readers do detect reliably is a letter containing nothing specific to the applicant, which is a content problem rather than an authorship one.
Do recruiters use AI detectors on cover letters?
Some do, but the tools are weak. OpenAI withdrew its own text classifier on 20 July 2023, citing low accuracy: it caught 26 percent of AI text while wrongly flagging 9 percent of human writing. A score from such a tool is not evidence, and treating it as evidence produces false accusations.
Is it acceptable to use AI to write a cover letter?
Using a model to draft and structure a letter is ordinary tool use, like a spellchecker or a template. What is not acceptable is letting it assert experience you do not have. Follow any policy the employer states, answer honestly if asked, and be able to discuss every claim in the letter.
Why do hiring managers say they can spot AI writing?
Because they are describing something real, just not authorship. Many AI-written letters are produced from the job posting alone, so they can only describe the job. A reader who sees fifty letters that say nothing about the applicant is detecting that emptiness, and attributing it to the tool.
Are AI detectors biased against non-native English writers?
Research says yes. A 2023 Patterns study ran seven detectors over 91 TOEFL essays written by humans and found an average false positive rate of 61.22 percent, while the same detectors were near-perfect on 88 essays by US schoolchildren. Careful, formal, limited-vocabulary English scores as machine-like.
What makes a cover letter look generic?
One test settles it: could this sentence be sent to a different employer, for a different job, unchanged? If yes, it is doing no work. Genericness is about missing content, not about style, and it happens to human-written letters and template letters exactly as often.
Should I tell an employer I used AI on my application?
You are not obliged to volunteer it, but never deny it. If the employer states a policy, follow it. If an interviewer asks directly, say so plainly and describe what you did: the drafting was assisted, the experience and the judgement in the letter are yours.
Next Steps
The question you arrived with turns out to be the wrong one to optimise against. Nobody can reliably identify the author of a page of application prose, and a great many people can identify a page that told them nothing. Only the second of those is under your control, and it is entirely under your control.
If you are assembling that material for the first time, working from the job description is the fastest way to decide which parts of your history are worth including at all.
Related resources: Tailor a Cover Letter to the Job Description | A Different Cover Letter for Every Job? | When You Do Not Meet the Requirements | Is Using AI for Interview Prep Cheating?