You have the application open in one tab and a chatbot in the other. Some of the sentences started as yours and some started as its, and after four rounds of editing you could not reliably say which is which any more. The worry is not really whether that was allowed. It is whether someone at the other end can tell, and what they will do about it if they think they can.
Three things are worth knowing before you send it. The tools built to detect machine-written text are wrong often, and wrong in a direction that should worry some people far more than others. When one of the larger hiring platforms built something for the AI problem, it built a thing that never asks who wrote the words. And the employers who have bothered to publish a policy are asking for something no detector could measure.
The detectors fail in one particular direction
In 2023 a Stanford group ran seven commercial GPT detectors — Originality.AI, Quil.org, Sapling, OpenAI's own classifier, Crossplag, GPTZero and ZeroGPT — across two sets of writing that were definitely produced by humans.1 One set was 91 TOEFL essays by non-native English writers, taken from a Chinese educational forum. The other was 88 essays by US eighth-graders, from the Hewlett Foundation's ASAP dataset.
The detectors did well on the eighth-graders: an average false positive rate of 5.19%. On the TOEFL essays the average false positive rate was 61.22%. All seven detectors agreed that 18 of the 91 essays were machine-written, and 89 of the 91 were flagged by at least one of them.1
Then the researchers ran the experiment that turns an embarrassing result into a damning one. They had ChatGPT rewrite the TOEFL essays with the instruction to "Enhance the word choices to sound more like that of a native speaker". The share flagged as machine-written fell from 61.22% to 11.77%. Run the other way, with the model asked to simplify the eighth-graders' language instead, the flagged share of those essays rose from 5.19% to 56.65%.1
That is the whole finding, and it is worth reading twice. Take each corpus on its own terms, because the two are different bodies of writing and comparing across them controls nothing. Within the TOEFL essays, the same source texts were flagged as machine-written 61.22% of the time while they were purely human work, and 11.77% of the time once a model had rewritten them. Within the eighth-grade essays, the same source texts went from 5.19% flagged to 56.65% once a model had rewritten them. In each corpus a machine had demonstrably touched the text, and in each the flag rate followed the register rather than the machine — downwards, in the case where the rewrite made the prose more elaborate. These detectors are not responding to evidence of a machine. They are responding to plain, restricted, unornamented English. Someone writing carefully in their third language produces exactly that. So does a model on a neutral setting.
The limits matter. This is one study from 2023, on 179 essays, testing the detector versions on sale at that moment rather than whatever those products have become since. Essays are not cover letters, and no equivalent study exists on hiring documents. The paper was posted as a preprint in April 2023 and published in Patterns that July.2 What it establishes is that the failure is systematic rather than random, and there is nothing about an application that would make that failure less likely.
An institution that did the arithmetic in public
Universities had far stronger reasons than employers to make detection work, and far more of them tried it at scale. Vanderbilt turned it off.
In August 2023 the university disabled Turnitin's AI detector across its courses. The reasoning is the useful part. Turnitin claimed a 1% false positive rate. Vanderbilt had submitted 75,000 papers the previous year, so on the vendor's own figure "around 750 student papers could have been incorrectly labeled" as containing AI writing. It added that "Turnitin gives no detailed information as to how it determines if a piece of writing is AI-generated", and that detectors "have been found to be more likely to label text written by non-native English speakers as AI-written".3
That is a decision about coursework rather than hiring, and one university's decision at that. The arithmetic transfers anyway. One per cent is the optimistic number, supplied by the party selling the tool, and it is still a poor trade when the volume is large and the entire cost of an error lands on the person being accused.
What hiring bought instead
Here is the part that almost never appears in the advice, and it reframes the question.
In June 2025 the applicant tracking system Greenhouse announced a partnership with the identity company CLEAR. Candidates verify themselves inside the hiring system by matching "a candidate's selfie to their government-issued ID", corroborated against other sources, alongside fraud detection aimed at "identity misrepresentation".4 The product page for the feature describes fraud signals, a blocklist for known malicious sources, and matching résumés against criteria a recruiter has set. It claims nothing about detecting who wrote the text.5 The announcement quotes a Gartner forecast that by 2028 up to a quarter of applicants could be fraudulent — a prediction rather than a measurement, but a fair indication of what the buyers were worried about.4
The question that product exists to answer is whether there is a real person behind the application. Not who composed the third paragraph. One vendor's roadmap proves nothing about the rest of the market: it does not show that nobody anywhere runs a text detector over a cover letter, and this article cannot tell you what any particular employer does. What it is is one dated, concrete example of what a large hiring platform chose to build when it addressed the AI problem head on, and what it built checks identity.
The employers who wrote it down are asking something else
A small number of employers publish what they expect from candidates. Those documents are far more informative than speculation about detection, because they say out loud what is actually being judged.
The UK Civil Service publishes a candidate's guide to AI in recruitment. Its acceptable uses include using AI to "help you refine and clarify your ideas and thoughts", to research public information about the organisation, and to "check the spelling, grammar and clarity of what you have written". Its unacceptable uses include "inflate or invent your skills and experiences", "create generic responses and copy these into your application", and using AI to complete situational judgement or numerical assessments. The guidance puts the principle in one sentence: "The goal of using AI in your job application is to help you show us who you are and what you are good at. It's not about creating a persona that isn't you."6
Anthropic, a company whose product is a chatbot, publishes a version broken down by stage. For a résumé and application questions: "Please create your first draft yourself, then use Claude to refine it." For take-home assessments: "Complete these without Claude unless we indicate otherwise." For interview preparation it actively encourages the tool. For the live interview: "This is all you–no AI assistance unless we indicate otherwise."7
Two organisations with almost nothing else in common, drawing the same three lines. Refining your own material is fine. Inventing experience is not. Live assessment is where it stops.
Notice that none of those lines is about text, and none of them could be enforced by reading the prose. A detector cannot tell whether an achievement happened. It cannot tell whether a paragraph began as your thought or the model's. What these employers have written down are rules about honesty and provenance, and the only person with access to that information is you.
What that leaves worth doing
Look for a stated policy first, and follow it. Where one exists it tends to sit on the careers site rather than in the advert, and following an explicit instruction puts you in a far better position than guessing at one. Where none exists, the Civil Service and Anthropic lines are a reasonable default: they are what two organisations concluded independently when they had to write it down.
Start from your own material, not from a blank prompt. This is the practical form of every published rule. Ask for a draft with no input and you get generic responses, which is the thing the Civil Service names as unacceptable and which a human reader spots without any software at all. Ask for help shaping your own examples and the result stays yours. Where an employer has asked for your own first draft, as Anthropic does, that instruction settles it and nothing here overrides it. Where none has said anything, JobCraftly drafts a cover letter from your own résumé and the job you are applying for, in a tone you choose — built out of material you supplied rather than material it made up, and still yours to check claim by claim before it goes anywhere.
Check every specific it produces. Models fill gaps confidently, and the gap they fill most readily is a number. A metric you did not achieve or a system you never used is not an AI problem; it is a false claim with your name at the bottom, and the interview is where it surfaces.
Treat the live stages as a different category. Both policies quoted here permit AI in preparation and forbid it in the room. That is the one line drawn consistently, and it is the one worth assuming even where nothing is written down.
If English is not your first language, know which way the error runs. The evidence above says the misclassification risk falls on you disproportionately, and that the thing which reduces it is more elaborate phrasing. Do not read that as advice to write ornately for a machine. Read it as a reason not to accept a detector's verdict about yourself, and, if you are ever told your application was flagged, to ask what tool produced the number and what its false positive rate is on writing like yours.
What is actually being assessed
The anxiety about detection assumes the wrong thing is being measured. Some employers do have a rule about origin — Anthropic's asks for a first draft that is yours — but it is a rule you follow or do not, not a property anyone reads off the page. What is being assessed is whether what those sentences claim is true, and whether you can do the work.
A tool that helps you say something true more clearly does not interfere with either of those. A tool used to manufacture something that is not there fails at the first interview question, and would have failed it in 2015 too. The technology is new. What is being asked of you is not.
