Can recruiters tell a cover letter was written by AI?

The patterns that actually stand out

The first is an opening that asserts enthusiasm without naming an occasion. "I read your advertisement with great interest and feel particularly drawn to your innovative company" says nothing that would not fit every advert ever written. A person genuinely interested in this role has a reason, and that reason is concrete.

The second is qualities with no evidence. "A team player, resilient and communicative" is a list of assertions. It is suspicious precisely because someone who has lived through something tends to recount it — and a language model that has lived through nothing produces exactly that list.

The third is the inserted company name. Sentences like "your position as a leading company in X inspires me" contain the name but no information about the company. Delete the name and the sentence fits anywhere — that is the actual finding.

The fourth is length. A text that fills exactly one page without needing to, built from four paragraphs of equal length, reads as formatted rather than written. People write unevenly; a paragraph is short because there was no more to say.

And the fifth is register. Conspicuously elevated phrasing that appears nowhere else in the file creates a gap between the letter and the person who later sits in the interview.

Why the problem predates the tools

The numbers show interchangeability is not a new phenomenon. For every two Swiss searches for "Motivationsschreiben" there is a good one for "Motivationsschreiben Vorlage". For the English "cover letter", four in five also search for a template. Across all languages, roughly half of cover-letter searches go to templates, samples and examples.

Which means: long before language models existed, a large share of applicants started from someone else's text and adapted it. Recruiters have known the result for decades — the standard letter that could equally well have gone to an accountancy office, a manufacturer or a care home.

The outlier is interesting: in French the template share is markedly lower. That is not evidence of greater diligence but of different search habits, and of "exemple" being sought instead of "modèle" — but it shows how culturally rather than technically determined these patterns are.

The conclusion is both reassuring and uncomfortable: nobody is rejected for using a tool. People are rejected for submitting an interchangeable document — and that was perfectly possible without software.

What recruiters actually check

Almost nobody in day-to-day Swiss recruiting runs cover letters through a detector. Those tools are unreliable, they produce false positives disproportionately often for non-native writers, and basing a rejection on one would be indefensible.

What gets checked instead is the connection. Does the letter contain anything only someone who read the advert could have written? Is one of the stated requirements picked up and evidenced with an experience? Does the explanation match the anomalies in the CV — the gap, the move, the relocation?

That check is indifferent to how the text came about. A hand-written letter with no connection fails it exactly as a machine-generated one does. And a letter produced with a tool's help that evidences two concrete connections passes it.

The second filter matters: the interview. Submit a letter whose register and self-presentation you cannot sustain in conversation, and you create precisely the gap that gets remembered. That is the real reason an inflated cover letter hurts.

The test that settles it in ten seconds

Delete the company name and the job title from your letter. Read it again. If it still fits any role afterwards, you have not written a cover letter — you have filled in a template, regardless of whether software was involved.

The second test: count the checkable statements. A checkable statement names a situation, an action and a result. If your letter contains not one, it consists of assertions, and assertions are the material interchangeable texts are made of.

The third is the read-aloud test. Sentences you would not say out loud do not belong in your letter. They are the places where a tool spoke on your behalf.

Pass all three and the origin of the first draft stops mattering. That is exactly what these tools are good for: structure, a first pass, checking for repetition — not the part that counts.

How to use a tool properly

Give the tool material rather than a task. Type "write me a cover letter for this job" and you get the average of all cover letters. Supply the advert, two of your own projects with their results, and your reason for moving, and you get a draft made of your own material.

Always produce the two middle connections yourself. Take the two heaviest requirements from the advert and evidence each with a concrete experience. That is the real work, it rarely takes more than a quarter of an hour, and it is the part no tool can invent for you because it does not know your experience.

Then use the tool to cut, not to extend. Language models tend to pad; the Swiss convention is one page. An instruction like "delete everything that contains no checkable statement" is the single most useful application there is.

And write the first and last sentence yourself. Those two are the most likely to be read, and they are also where the machine register shows through most clearly.

What not to do

Do not hide your use of tools by inserting deliberate mistakes. That is the forum advice which reliably backfires: a letter with planted typos does not read as more human, it reads as careless.

Do not overshoot on register. If your German is B2 and your letter is C2, the interview opens a gap that weighs more than the level itself. Write at the level you actually speak — which in Switzerland is no weakness, since four in five stated language requirements ask for professional rather than native level anyway.

And do not reuse the same letter within one industry. Recruiters in a field know each other, and in a market the size of Switzerland the odds of the same document landing twice on the same desk are higher than elsewhere.

The calmest route remains the simplest: a fixed base structure, two evidenced connections per role, one page. That makes the question of whether a tool shows entirely moot.