Do Recruiters Notice AI-Written Applications?

The question people actually mean is "will I be caught", and the answer is that detection is unreliable, most recruiters are not attempting it, and the thing that costs you is not detection. For background on less visible forms of workplace tracking, consult stealth monitoring software.

Detection does not work well

AI detection tools produce both false positives and false negatives at rates that make them unusable for decisions. They flag non-native English speakers disproportionately, they flag careful formal writing, and they miss text that has been edited. For an independent public reference, consult EEOC artificial-intelligence resources.

Some employers use them anyway, which is a risk that falls unevenly and unfairly — particularly on people writing in a second language, whose careful, correct prose is exactly what these tools flag.

And a growing number of employers do not care, because they assume everyone is using these tools and have moved the assessment elsewhere.

So "will I be caught" is the wrong question. The right one is "what does this application look like to someone reading four hundred".

What recruiters actually notice

Not the model. The result.

Generic content. A letter that could go to any company, because nothing in it is specific to this one. This is the strongest signal and it predates these tools by decades — the difference is that generic text is now much faster to produce, so there is more of it.

Uniform structure across applicants. Recruiters reading many applications see the same shape repeatedly: same length, same three-paragraph rhythm, same closing sentence. Any individual one passes; the pattern is visible.

Register mismatch. Polished formal prose in the letter, and a different person on the call.

Claims that collapse. An impressive project described in the letter that the candidate cannot discuss for two minutes. This is the one that actually costs offers, and it is not detected by software — it is detected by a follow-up question.

Errors nobody read. The wrong company name, an inference about the role that is wrong, enthusiasm for something the company does not do.

And volume behaviour. Fifty lightly varied applications from one person in a day is visible in the system.

What actually gets penalised

Being precise, because the anxiety and the risk are not in the same place.

Using a tool to phrase your real experience: essentially never penalised. Nobody is assessing your typing.

Producing something generic: penalised heavily, and it would have been penalised if you wrote it by hand.

Fabricating experience: penalised severely, and this is misrepresentation regardless of what produced the text. See where the line is.

Ignoring an explicit instruction not to use AI, particularly on a task: treated as an integrity issue and it ends processes.

Using it during a live interview: increasingly detected, and it ends the process where it is found.

What to do about it

Make it specific. One sentence that could only have been written about this company and this role does more than any amount of polish. That is also the fix for the generic problem whether or not a model was involved.

Rewrite in your voice. Read it aloud. If you would not say it, change it.

Verify every specific. Generated impressiveness produces plausible numbers that are not yours. See quantifying achievements.

Rehearse everything you send. Two minutes on each claim. This is the step that converts a written application into something you can defend, and it is where the actual risk is removed. See preparing for an interview.

Read the instructions. Where an employer states a position, follow it. Where they permit it with disclosure, disclose.

If you are asked directly

It happens, and increasingly.

Answer honestly. "I used a tool to help structure it, and the content is mine" is a reasonable and increasingly ordinary answer.

Denying it and then being unable to discuss what you wrote is much worse than the disclosure would have been.

And if an employer treats ordinary assistance as disqualifying, that is information about them — though it does not help you if you wanted the job.

The unfair part

Worth naming.

Detection tools disproportionately flag people writing in a second language. Careful, formal, grammatically correct prose is exactly the profile these tools score as machine-written, and it is also the profile of someone who has learned the language formally.

If this affects you, the protections are the same as the general advice and matter more: specificity that only you could have written, your own voice, and the ability to discuss everything in the document. A candidate who is fluent about their own work in conversation is difficult to dismiss on a detector's output.

The short version

Detection is unreliable and many employers have stopped trying.

What they notice is generic content, register mismatch, and claims that collapse under one follow-up question.

Ordinary assistance is essentially never penalised. Fabrication and ignoring explicit instructions are.

Specificity is the fix — one sentence that could only be about this role does more than any polish.

And rehearse everything you send, because the real check happens in conversation rather than in software.