How Do You Stop Hiring Managers Rejecting Candidates for 'Sounding Like AI'?

Why does a memo asking people to stop never work?

Because "sounds like AI" isn't a rule anyone chose to follow. It's a perception, and you can't revoke a perception by email. What you can change is what a rejection is allowed to contain. Make the reviewer write the specific claim in the work that's wrong, thin or unverifiable, and where they checked it. Style alone doesn't fill that field, so the instinct has nowhere to land.

The instinct is also the weakest instrument in the funnel. Across three studies of unstructured interviews, people who unknowingly ran interviews in which the answers were generated at random came away as confident in their impressions as people who ran real ones; having interviewed anyone made their predictions worse rather than better; and participants still reported they would rather have a random interview than no interview at all ². A confident read of a stranger's prose is that same machinery with less to work from.

A memo produces one of two outcomes, both bad. Reviewers stop saying it out loud and keep doing it, or the rejection note becomes "not a fit" and the reason leaves the record entirely. Neither is a screen anyone can audit six months later, when someone asks why one req family was rejected at four times the rate of another.

Settle the underlying policy in writing first. Whether an AI-written application is disqualifying at all is a decision the company makes once, and a rejection form with no stated policy behind it only moves the argument onto the form.

What does the rejection form have to make them write?

Three fields, and the rejection can't be submitted without them. One: the specific claim, number or decision in the submission that's wrong, thin or unverifiable. Two: what the reviewer did to check it. Three: what the check returned. A reviewer who fills all three has found something real. A reviewer who can only say the prose felt generated has found nothing anyone can act on.

Worked examples travel further than the rule, so put two of them in the form itself:

• Fills the fields. "The letter says they rebuilt the team's forecasting model. In screen I asked what the old model got wrong and what they changed; the answer named neither, and the two versions they described don't reconcile."
• Doesn't fill the fields. "Reads like ChatGPT. Every paragraph is the same length and it uses 'moreover' twice."

The second is a statement about sentences. The first is a statement about a claim, and it survives being read back to the candidate, which is the practical test, because an AI-related rejection you can't explain to the candidate is one you will end up explaining to somebody else.

One rule closes the loophole: anything a candidate could fix by rewriting sentences is not evidence about their judgment. Em dashes, tidy parallel structure, an unusually clean summary paragraph: all style, all coachable in an afternoon, and none of it tells you whether the person checked anything. Brief the panel on that distinction before the form ships rather than after the first dispute, since getting a panel to read AI use the same way mostly comes down to everyone holding one sheet of paper.

Why do engineering and commercial reqs flag different people?

Because the flag fires on writing conventions, and conventions vary by discipline and by first language. Even, low-variance prose is what the instinct reads as machine-written, and it is also what you get from a competent engineer writing in a second language, from anyone trained on a house style, and from a candidate who ran the letter past a template. One unexamined rule, two req families, two different rejection patterns.

The measurement of this exists and it is not close. Seven widely used GPT detectors were run over TOEFL essays written by non-native English speakers and over essays by US eighth-graders: the average false-positive rate was 61.22% on the TOEFL essays against 5.19% on the eighth-graders', and 97.80% of the TOEFL essays were flagged as AI-generated by at least one of the seven ¹. Prompting a language model to rewrite the same essays with richer vocabulary made most of the flags go away, which is the finding that matters for a hiring panel: the signal is linguistic range, not authorship.

Your reviewers are running a worse version of that classifier from memory, with no threshold and no record. Whether detectors work well enough to use in hiring is a separate question with the same answer: buying one doesn't convert a hunch into evidence, it converts it into a hunch with an invoice attached.

Which is why the by-req cut below is the first thing worth building. A single company-wide flag rate hides the exact pattern you need to see.

Measure the flag rate by req and by reviewer

Two cuts, monthly. Flags per hundred submissions by requisition family, and flags per reviewer. The first shows whether one discipline's writing conventions are being read as a defect; the second shows whether one person is generating most of them. Both are trivial to produce once the reason field exists, and neither is available at all while the reason lives in a reviewer's head.

There is a legal reason to want the record, separate from the accuracy one. The Uniform Guidelines define a selection procedure as any measure used as a basis for an employment decision, and define it expansively: "the full range of assessment techniques from traditional paper and pencil tests, performance tests, training programs, or probationary periods and physical, educational, and work experience requirements through informal or casual interviews and unscored application forms" ³. A resume reviewer's impression sits inside that definition. It is not outside the rules because nobody wrote it down.

The Guidelines also expect an employer to keep records disclosing the impact its selection procedures have on identifiable race, sex and ethnic groups, and treat a selection rate under four-fifths of the highest group's rate as evidence of adverse impact ⁴. Where a procedure does screen out a protected group, the EEOC's position on employment tests and selection procedures is that the employer has to show it is job-related and consistent with business necessity ⁵. "It sounded like AI" is not a showing anyone has managed to make. Where a tool does the screening, the duties get more specific still: New York City requires notice to a candidate living in the city at least 10 business days before an automated employment decision tool is used ⁶. A reviewer's hunch triggers no such notice, which is exactly why the form has to hold the record instead.

Set the threshold before you look at the numbers, so the review isn't a negotiation about whether a gap is large. One reviewer at twice the panel median, or one req family at twice the others, triggers a read of the actual notes. A read, not a retraining. Most months the notes explain themselves, and when they don't, the fix is usually one more worked example on the form.

What a written standard still can't tell you

Whether the person can do the work. The form fixes the reason, not the signal. A candidate whose letter names a claim and a source may have had every word of it produced for them, and a candidate whose prose reads flat may be the sharpest reader in the pile. Writing was never the measure of judgment; it was the part of the work that happened to be visible at screen.

So stop asking the application to carry weight it can't hold, and move the decision to something where the answer isn't stylistic. A short task with an assistant available, a confident claim inside it that happens to be wrong, and a record of what the candidate did about it will separate those two candidates in twenty minutes, and it returns a rejection reason nobody has to argue about. Whether a detector, an interview or a work sample is the right instrument is worth settling before you spend another quarter on the form.

Two things stay in place either way. Say what's allowed with AI in the posting, in one sentence: an unstated rule gets guessed at, and the guessing measures how well someone read your company rather than how well they work. And keep the reason field mandatory after the work sample lands, because the screen still exists and it is still the least examined thing in the funnel.

See how Olive measures this

See the benchmarks >