How to Make a Cover Letter Not Sound AI Generated

You have a draft open. It is grammatical, it is polite, and something about it makes you uneasy. You are not really asking whether to use AI on a cover letter, that question is settled for you. You want to know what to cut before you send it.

Here is the short version, and it will save you an hour. There is no good public evidence that employers routinely run cover letters through AI detectors, and the detectors themselves are unreliable. What loses you the job is more often that the letter could have been sent to any company, unchanged. That is a content problem, not a vocabulary problem, and it is fixable.

Do employers actually run cover letters through AI detectors?

Probably not often, and the tools would not be trustworthy if they did. Stanford researchers (Liang and colleagues, published in Patterns in 2023) ran seven detectors over TOEFL essays written by non-native English speakers and found an average false positive rate above 61 percent, while the same detectors were close to perfect on essays by US eighth graders. OpenAI withdrew its own AI classifier in 2023 over its low accuracy. Detectors key on low perplexity, meaning predictable word choice, which measures writing variety rather than authorship. Detection vendors do market accuracy figures to recruiters, but the vendor pages we checked publish no cover-letter-specific benchmark and no evidence that employers are screening this way.

Humans are not obviously better at spotting authorship either. ResumeBuilder, a resume-industry company, reported in 2023 that in its own survey of 1,000 hiring managers, 82 percent failed to correctly identify which cover letter introductions ChatGPT had written. Treat a single company-run survey lightly, but the direction is consistent with the rest of the evidence. Editing until a detector calls you human is optimizing for the wrong reader, and it is unfair advice if English is your second language, because your own unassisted writing is exactly the kind most likely to get flagged.

The words worth deleting, and why the list is not enough

The vocabulary fingerprint is real and it has been measured. Kobak and colleagues (Science Advances, 2025) compared more than 15 million PubMed abstracts before and after ChatGPT and found roughly 900 words whose use suddenly spiked. Before 2024 the excess words were mostly nouns; in 2024 they were mostly verbs and adjectives, which is part of why AI prose can read as ornamental rather than informative. Words worth searching for and cutting:

• delve and delves
• showcasing, underscores, compelling
• pivotal, testament, tapestry, realm, landscape
• unwavering and meticulous, when no example follows them
• the opener: I am writing to express my strong interest in the [Role] position at your esteemed organization

Delete all of it. Then accept that a letter can pass every vocabulary check and still be interchangeable. The word list is the smoke. The next section is the fire.

The swap test

Replace the company name and role title with a competitor's. If the letter still reads fine, it is dead, no matter who or what wrote it. Then go line by line and mark each sentence as one of three things: derivable from the job posting alone, derivable from your resume alone, or something only you could have written. Plenty of rejected letters are entirely the first two. Aim for at least two sentences of the third kind, and put one of them first.

Harvard's career services advises that generative AI should not be the primary author of your application materials, and suggests treating AI text as suggested edits rather than a finished draft. Ohio State's engineering career services is blunter about the raw output, calling it generic and robotic. Tufts frames AI as a starting point rather than the finish line. Harvard's two checks are better than any detector: read the letter aloud, and confirm you could defend every line in an interview.

Fix the input, not the edit pass

A common reason drafts come out generic is that the prompt contained only the job description, the one document every other applicant also has. The repair is upstream. Before you generate anything, gather:

• your own rough notes on one project, including what went wrong and what you decided
• one specific thing about this company you actually touched: a changelog, a pricing page you compared, a support reply, a talk you watched
• something you wrote unprompted in your normal register, such as a long email, pasted in as a voice sample (MIT's career advising team recommends using your own writing this way)
• if your work does not come with percentages, the scale, the constraint, and the stakes instead, which is the honest version for teachers, care workers, admins and career changers

That is roughly the principle behind Tailor CV in mbason-ai: it works from your own history against a specific posting, not from the posting alone. The defend-every-line test is the useful one here: if you could not say a sentence out loud in the interview without wincing, it does not belong in the letter. And because volume is part of what pushes people toward generic letters in the first place, the Application Tracker gives each application a forwarding email so its status updates itself when a reply lands, with one-click calendar export for the interviews that follow. If you want the fuller picture, here is what we are building.

Who it is not for

If you are applying to five carefully chosen roles, skip the tooling and write the letters yourself, slowly. If you have no work you can describe concretely, no prompt and no product fixes that, and the letter is not where you should be spending the evening. If the application explicitly asks you not to use AI, or the role is effectively a writing sample, follow the instructions in front of you; this post is about quality, not permission. And if what you wanted was a humanizer that scrambles text until a detector relents, we do not sell one and would not point you at one: in our view it degrades clear writing, and on the evidence above the detectors it games misfire hardest on the people such tools claim to protect.

Should you tell them you used AI?

Employer and application policies vary, and most applications do not ask about drafting assistance at all. Where an application asks, answer truthfully. Where it asks for original writing, write it. The line that actually matters is simpler: never claim work you did not do. Rules around AI in hiring are also shifting on the employer side, with state-level notice and anti-discrimination provisions arriving, so read the specific application in front of you rather than relying on a general rule.

So run the swap test one more time. Would this letter still make sense if it arrived at a different company? If yes, rewrite it, not because a machine will catch you, but because a person will not care.