Most people ask the wrong question about AI cover letters. They want to know whether a recruiter can tell that ChatGPT wrote it. In practice, recruiters rarely think "this is AI"; they think "this could have gone to any company on the list", and they move on to the next application.
So the useful split isn't AI versus human. It's which parts of a cover letter a machine can handle, and which parts only you can supply.
What AI is actually good at
Structure is the easy win. A cover letter has a shape that hasn't changed much in twenty years: why this role, what you've done that maps onto it, what you want next, a closing line that isn't begging. Language models produce that shape reliably, which saves you the forty minutes of staring at a blank page that kills most applications before they start.
Vocabulary matching is the second win. Paste the job ad in and ask for the terms the employer uses for work you already do. If the ad says "revenue operations" and your CV says "sales admin", that gap costs you, and a model will spot it faster than you will after rereading your own CV two hundred times.
The third one gets underrated: writing in a second language. If you're applying in French or German at a solid B2, a model will fix the register problems that make a letter read as foreign. Not the content, the phrasing. That's a real advantage and there's no reason to feel odd about using it.
Where it falls apart
Ask a model to write about your experience and it produces claims that are true-shaped but empty. "I have a proven track record of improving team performance." Nobody has ever written that sentence about themselves and meant it. It's the average of ten thousand cover letters, which is precisely what a language model is built to produce.
What a hiring manager reads for is the opposite of an average:
- The number. Not "increased conversion" but "took checkout conversion from 2.1% to 3.4% over two quarters".
- The name. The project, the tool, the client, the market. "Migrated billing from Chargebee to Stripe" beats "led a payments migration".
- The reason it's this employer. Something you know about the company that isn't on the first screen of their careers page.
- The awkward part. The gap year, the pivot, the eleven-month stint. A model smooths these over, and smoothing over is worse than one honest sentence.
A model can't produce any of these, because it doesn't have them. It will happily invent something plausible if you let it, and that's how candidates end up in interviews defending a metric they made up in a hurry the night before.
A workflow that keeps both halves
- Extract before you generate. Read the job ad and pull out the five or six requirements that carry real weight, plus the exact words used for them. Do this yourself. It takes four minutes and it decides everything that follows.
- Ask for a skeleton, not a letter. Have the model give you paragraph purposes and topic sentences against those requirements. If you ask for a finished letter you will get one, and you will be tempted to send it.
- Fill in the facts. Go through the draft and mark every claim with no number, no name, no date. Replace each with something from your actual history. If you can't replace it, delete it. A short letter with three real facts beats a full page of well-formed nothing.
- Strip the tells. Separate pass, separate mindset. It goes faster than you'd expect once you know the list.
- Read it out loud. Every sentence you stumble over is a sentence you didn't write. Your ear catches what your eye skims, especially in your own language.
That last step sounds soft. It's the highest-yield one on the list.
What gets a machine draft binned
The phrases that give it away
English drafts share a small set of habits that read as generated to anyone who screens applications weekly:
- "I am writing to express my strong interest in the position of..."
- "I am particularly drawn to your company's commitment to innovation."
- "This role aligns perfectly with my career trajectory."
- "proven track record", "passionate about", "dynamic environment", "leverage my skills"
- Sentences built on "Not only... but also..."
- Perfectly balanced triples: "creativity, collaboration and results".
None of these are forbidden words. The problem is density. One of them is a bland sentence. Five of them across three paragraphs is a pattern, and patterns are what a screener notices on the two-hundredth letter of the week.
Applying in a second language? The same thing happens there with a different phrase list. Ask a native speaker to read the draft once, not for grammar but for whether it sounds like a person.
Copying the job ad back at the employer
There's a specific failure mode when you paste the job ad into a model and ask it to match the language closely. You get a letter that reproduces the ad's phrasing almost verbatim, sometimes in the same order it appeared.
That reads badly to a human, and it can work against you in systems that score documents for relevance. Repeating a phrase five times doesn't make you five times more relevant, it makes the document look engineered. Use the employer's vocabulary once, where it's accurate, then write the rest in your own words. If a sentence in your letter appears word for word in the posting, rewrite it.
Keyword coverage belongs on your CV anyway. The letter is where you explain the parts of the CV that don't explain themselves, which is also why a tool like Postulit handles the structured half of an application from your LinkedIn profile and leaves the argument to you.
What about disclosure
Policies vary and they keep changing. Some employers ask candidates to declare AI use, some forbid it in the application instructions, some run their own AI screening and would find your objection awkward. A number of public-sector and graduate schemes now say something explicit either way.
So read the instructions. If there's a rule, follow it, including when it strikes you as silly. If there's no rule, the standard most people apply in practice is that help with drafting and language is fine while inventing experience is not, which is the same line that applied back when candidates paid a friend to edit their letters.
Where to start tonight
Open the last cover letter you sent and count the sentences containing a number, a proper noun or a date. Fewer than three? That's the problem, and no model will fix it for you. Rewrite those sentences with facts from your own history first, then let AI clean up whatever is left.