What Would Checking an Application for AI Even Consist Of?

The honest answer is that you cannot know, because a checking step would be an internal process decision at one employer and no employer publishes it. That makes the detection framing a dead end. The more useful question is what such a step would have to consist of — because once you sketch it, it becomes clear why its output could not function as evidence, and why optimising against it is not available to you. What a reader actually notices in drafted prose, which is the answerable part, is covered in whether anyone can tell you used AI.

The three things a “check” could be

A tool run over submitted text. Someone pastes the document into a classifier and reads a number. This is the version people imagine, and it is the weakest, for the reason below.

A person forming an impression while reading. Far more common in practice, and not a check at all — it is ordinary reading, and it produces a hunch rather than a finding.

A question later in the process. The application proceeds and a conversation tests whether the person can talk about what the document claims. This is the only one of the three that produces information, and note that it does not test authorship at all.

Why a classifier’s output cannot be evidence

A classifier returns a likelihood about text, not a fact about a person. It has no access to how the document was made and cannot distinguish a drafting tool from a careful writer who edits for clarity, follows conventional structure, and avoids idiosyncrasy — which is precisely what resume advice tells everybody to do.

So a plain, well-organised, unadorned resume is the sort of document such a tool scores as suspicious, and it is also the sort of document that is good. That is not a calibration problem to be fixed; it is what happens when you measure style and infer origin. The same reasoning applies to the tells that circulate informally, which describe careful human writers just as well — set out in what a reviewer can actually infer from a resume’s prose.

An employer acting on a number like that would be acting on nothing, and would have nothing to say if challenged. Which is the practical point: a suspicion of this kind cannot be stated, so it does not get stated. It gets acted on silently, if at all, by moving on to the next application.

What this means for you, concretely

You cannot optimise against it. You do not know whether the employer runs anything, what it would be, or what it would flag. Writing to defeat an unknown classifier means degrading your prose on a guess.

Style is the wrong thing to fix anyway. The one property that does travel through every version of a check is whether the document’s claims correspond to things you did. Prose can be argued about; a claim can be tested. A reader with a hunch about your writing asks harder questions, and the questions are about content.

The real exposure is downstream, not at the document. If a bullet asserts something you cannot describe in detail, the conversation surfaces it — not because anyone detected anything, but because they asked. That is why the interview questions your resume invites is the more actionable version of this worry.

The one place a direct question does appear

Some application forms ask whether AI tools were used. Answer accurately. A truthful yes is very rarely disqualifying, and a false no is a written misstatement on an application form, which is the worse of the two by a wide margin. That is a policy question with a knowable answer, unlike everything else on this page.

The question worth substituting

Stop asking whether employers check and start asking whether your document survives being asked about. Rewrite anything you could not talk through with no notes in front of you; leave the plain, honest phrasing alone. The exact-match version of this question, and the older treatment of it on this site, is at do employers check for AI-generated content; the reader-side detail of what actually shows in drafted prose remains in whether anyone can tell.