Should You Reject a Candidate Who Used AI to Write Their Resume?

Is using AI to write a resume a reason to reject?

No, not by itself. In a field experiment with roughly 500,000 jobseekers in an online labor market, applicants given algorithmic writing assistance were hired 8% more often, and the researchers found no evidence that employers ended up less satisfied ¹. The assistance made candidates easier to read rather than harder to judge. Rejecting on authorship gives up that legibility and buys nothing you can verify.

Nobody can reliably tell the difference anyway. Text detectors fail in the direction that costs most: one study found detectors classified writing by non-native English speakers as machine-generated at high rates while rarely flagging native writing, and that simple prompting defeated them ². Run that over an application stack and the errors land on people who learned English later, which is an adverse-impact problem you built yourself. Whether detectors work at all in hiring has a short answer.

Human judgment isn't the fallback. A recruiter who believes they can spot model prose is reacting to something real (flat, generic, template-shaped writing), but career-services templates have produced exactly that for decades, and the tell doesn't separate a model from a well-coached candidate. Guessing at AI writing is a coin flip your process then treats as a finding.

A reject-on-suspicion policy also has a selection effect nobody wants. Candidates who are careless about the tool get caught; candidates who are deliberate about hiding it do not. You end up filtering for concealment skill, and the group you most wanted to identify is the group best equipped to route around you.

What was the resume ever evidence of?

Nothing you could check. A resume is a self-report: a list of things a candidate says they did, in an order they chose, with the failures left out. Nobody ever hired on the strength of the document; the document earned a conversation and the conversation did the work. AI lowered the cost of producing a polished claim. It did not change what the claim was worth, because the claim was never worth much.

What actually changed is the spread. When every application is fluent, the resume stops sorting anyone. The gap it used to carry between careless and careful writing collapsed, and that gap was doing most of the screening. This is a capacity problem rather than an ethics problem, and it arrives in the same quarter as keyword screening quietly failing for the same reason: the artifact got cheap to optimize.

So treat the resume as routing rather than evidence. It tells you what the candidate wants to be considered for and what vocabulary they are comfortable in. Everything you used to infer from prose quality (care, precision, judgment) has to move to a step where the candidate does something instead of describing something, which is the actual argument for replacing the cover-letter screen rather than trying to authenticate it.

One property is still worth reading closely: specificity. A model will produce "drove cross-functional alignment" all day. It will not invent that a candidate re-added three supplier quotes onto the same shipping terms and cut a landed cost by four points. Specific claims are checkable, and checkable is the only quality a resume line can have that survives into the next round.

Which jobs make AI-assisted writing a competence?

The ones where drafting and editing already happen with a model. Among US workers who have used an AI chatbot for work, the most common uses are research, editing written content and drafting it ³, so a candidate who drafted their resume with a model performed the most ordinary task in the tool's repertoire. In those roles, not using it is the stranger signal.

It is not universal, and that changes how the signal reads. Daily use is still minority behavior: 23% of employed US adults had used generative AI for work in the previous week and 9% used it every workday ⁴, and 55% of workers say they rarely or never use a chatbot at work ³. "Everyone does this now" is false. "This is normal in the jobs where it's normal" is the accurate version, and the job is yours to name.

The distinction that predicts trouble is not AI or no AI. It is supervised or unsupervised. A marketer who drafts with a model and cuts the two lines that overstate the case is doing the job. An analyst who forwards a model's summary of a rule without opening the rule is a liability, and would have been one with a junior's summary too. Ask which of those the candidate is describing, because what the role actually does with AI is the question that makes the resume readable again.

Reject on the claim, not on the author

Two things justify a rejection at this stage, and neither is authorship. First, a specific claim that fails a check: a title, a date, a headcount, a named system, a certification. Second, an account of working the role cannot carry, like an analyst describing model output sent to a regulator with nobody else reading it. Both are about content. Both survive being explained to the candidate.

• A claim that fails verification. These were always the grounds and they still are. Checking is cheap, the candidate can answer for it, and the finding is written down in words anyone can review.
• A working style the role can't afford. The tell is not "used a chatbot." It is output shipped untouched. Ask what the candidate refused and what they checked outside the conversation. Verifying an AI claim on a resume takes about ten minutes on a call.
• Nothing else at this stage. Tone, polish, formatting, a fondness for the word "robust": style tells that a $19 template produces too, and none of them predicts performance.

Whatever the rule is, apply it identically to everyone. A step that decides who advances is a selection procedure, and the standard questions asked of one are whether it was applied consistently and whether it relates to the job ⁵. "It read like AI to me" fails both halves, and it fails them in writing, in a file, months later.

Write the rule down before the stack arrives. Two lines is enough: AI assistance in the application is permitted, and every factual claim will be verified. Stating it removes the incentive to hide the tool, and it stops the rule being re-invented per resume by whoever happens to be screening on Thursday.

Where does a resume screen stop?

At the claim. A resume screen, with or without AI in the picture, can tell you what a candidate says they did and whether the checkable parts hold up. It cannot show you the candidate deciding anything: what they framed before generating, what evidence they demanded, what they refused. Those happen in the middle of the work, and the resume is the artifact left over afterward.

The move is to spend less on the document and more on the step after it. A short job simulation or work sample produces the thing a resume cannot: a record of the candidate working, which is the only place AI collaboration is visible at all. Run it identically for every finalist, grade it against something written down before the first submission arrived, and authorship stops being a question, because the candidate's AI use sits inside the evidence rather than behind it.

The cost is real, and worth saying out loud. Every added step drops completion, and 400 applications do not survive a work sample at the top of the funnel. The realistic shape is resume as routing, one cheap verification pass on the claims that matter, and a work sample for the short list. That costs more than a detector subscription and it is the only version that produces something you can defend to a candidate, a hiring manager, or a lawyer.

Olive is one instrument for that last step and not the only one; multiple-choice AI literacy tests, code-collaboration graders and unwatched take-homes are all real approaches with different trade-offs. What matters more than the vendor is that the evidence comes from work, and that the candidate can read what was written about them.

See a sample report >