Should Recruiters Reject AI-Written Job Applications? A Balanced Hiring Perspective

A candidate sends in a cover letter that reads beautifully. Clean structure, no typos, every paragraph lands. Then a tool flags it as 92% AI-generated. Do you toss it?

That question is landing on recruiters’ desks every day now, and the knee-jerk answer (“reject it”) is the wrong one more often than people think. Here’s a balanced look at what an AI-written job application actually tells you, what it doesn’t, and how to screen candidates fairly when nearly everyone has a writing assistant in their browser.

Key takeaways

• AI-written applications are common now; a reject-on-detection rule throws out strong candidates alongside lazy ones.
• AI detectors output probabilities, not proof, and they misfire on non-native English writers, short text, and lightly edited drafts.
• What you actually hire for is fit, judgment, and truthfulness. An AI score measures none of those directly.
• A better approach: judge the substance, verify claims in interviews and work samples, and ask whether AI was used well or carelessly.
• If you screen for AI, make it a disclosed expectation, not a hidden disqualifier, and never let a detector score decide alone.

The real question isn’t “did they use AI”

It’s “can they do the job, and are they telling the truth about themselves?”

Those are the two things a hiring process exists to answer. AI use is a proxy at best, and a noisy one. A candidate who used ChatGPT to tighten three rambling paragraphs into a clear pitch may be a sharper communicator than one who submitted unaided mush. The tool isn’t the signal. The thinking behind it is.

There’s a useful parallel here. When spell-check arrived, nobody argued that a typo-free resume meant the applicant was hiding something. When Grammarly went mainstream, recruiters didn’t start penalizing clean grammar. AI writing assistants are further along the same line. The interesting question was never the tool. It was always: does this person have the experience they claim, and can they reason on their feet?

What an AI-written application can and can’t hide

AI is genuinely good at the surface layer of an application:

• Smoothing awkward phrasing
• Matching a professional tone
• Restructuring a wall of text into readable sections
• Tailoring generic language to a job description’s keywords

That last one matters for recruiters running keyword-based screening, because a candidate can now mirror your posting almost perfectly. But notice what AI still can’t do: it can’t invent a real project the candidate shipped, recall a specific metric they moved, or answer a follow-up about a decision they made under pressure. The substance still has to come from a human who lived it.

So the polish is cheap now. The substance isn’t. Smart screening shifts weight away from the polish.

A quick mini-scenario

Two candidates apply for a marketing role.

Candidate A submits a flawless, AI-polished cover letter full of phrases like “results-driven professional passionate about driving impact.” Lovely sentences. Zero specifics.

Candidate B submits a slightly clunkier letter, clearly also AI-assisted, that says: “Last year I ran a 6-week email reactivation campaign for 40,000 lapsed users and recovered about 8% of them. I’d want to try the same playbook on your churned trial accounts.” (Numbers like these are the candidate’s own claim, so verify them later.)

If you auto-rejected both for “AI tells,” you’d lose Candidate B, the one who actually told you something. The detector can’t see the difference between empty polish and grounded specifics. You can. That’s the whole argument for reading before you reject.

Why detector-based auto-rejection backfires

Plenty of recruiting teams have bolted an AI detector onto their applicant tracking system and set a threshold: above X%, the application gets filtered out. It feels efficient. It’s also where the trouble starts.

Detectors don’t catch AI. They estimate the statistical likelihood that text looks AI-generated, mostly by measuring how predictable the word choices are and how much the sentence lengths vary. If you want the mechanics, our breakdown of perplexity and burstiness walks through exactly what these tools measure and why predictable-sounding human writing trips them up.

A few facts worth keeping in front of you:

• They produce false positives. Polished, plainly-written human text often scores high. A careful writer who avoids slang and keeps grammar clean can look “AI” to a detector.
• They’re biased against non-native English writers. A 2023 study led by Weixin Liang and colleagues, including Stanford’s James Zou, titled “GPT detectors are biased against non-native English writers,” found detectors flagged essays by non-native English speakers as AI far more often than native-speaker essays. Auto-rejecting on score can quietly become a discrimination problem in your funnel.
• They’re weakest on short text. Cover letters and resume summaries are short. That’s precisely where detectors are least reliable, because they have too few words to judge.
• The vendors themselves have backed off. OpenAI retired its own AI-text classifier in July 2023, citing low accuracy. When the people who build these models can’t make a reliable detector, a third-party score deserves real skepticism.

There’s also a compliance angle worth flagging. The EEOC has warned that algorithmic and AI-based hiring tools can create adverse impact against protected groups, and that anti-discrimination law applies whether a human or a machine made the call. A detector that disproportionately flags non-native English speakers is exactly the kind of automated screen that can draw scrutiny. “The software did it” is not a defense.

None of this means AI use is irrelevant. It means a score is a starting point for a question, never the end of a decision.

A balanced screening framework

Here’s a practical way to handle AI-written applications without either ignoring the issue or over-policing it.

1. Decide your actual expectation, and say it out loud

If you genuinely want to assess unaided writing, build a stage for that: a timed, in-person or proctored writing prompt. Don’t try to reverse-engineer it from a polished application a candidate wrote at home with every tool available. And put your expectation in the posting. “We use AI tools here and don’t mind if you do too” or “the first-round exercise must be your own unaided writing” both work. Hidden rules just punish honest people who guessed wrong.

2. Read for substance before you screen for style

Train screeners to look for specifics: numbers, named projects, decisions, tradeoffs. These are hard to fake and easy to verify later. A beautifully written application with no concrete claims is weaker than a rougher one full of them, AI or not.

3. Verify in the stages AI can’t reach

The interview is your real detector. Ask a candidate to walk through a project from their application. Ask the follow-up question. Ask what they’d do differently. Someone who lived the experience answers fluidly; someone who outsourced their whole identity to a chatbot stalls. A short, relevant work sample does the same job. AI can write about doing the work. It can’t reliably perform a coherent conversation about work the candidate never did.

4. Treat AI use as a skill signal, not a red flag

In a lot of roles now, “uses AI tools effectively” is a plus, not a minus. A candidate who used AI to produce a tight, specific, well-targeted application demonstrated exactly the kind of leverage you’d want them using on the job. The failure mode isn’t “used AI.” It’s “used AI to mask emptiness,” and substance-first reading catches that.

Where AI use should still concern you

To be clear, this isn’t a free pass. Two things genuinely matter:

Honesty of claims. Lying is lying whether a human or a model typed it. A fabricated degree, a fake job, an inflated title, those are integrity problems, and they’re worth catching. But you catch them by verifying facts, not by measuring writing style.

Generic, untailored slop. An application that’s clearly mass-generated and blasted to a hundred companies signals low effort and low interest. You don’t need a detector for that; you can feel it in the absence of anything specific to your role.

Both of those are about content. Neither requires you to play forensic linguist with a probability score.

The bottom line for recruiters

Rejecting every application a tool flags as AI is fast, defensible-sounding, and mostly wrong. You’ll lose strong candidates, you’ll lean hardest on the people whose first language isn’t English, and you’ll still let polished-but-empty applications through, because the detector can’t tell substance from style. AI in applications is now the water everyone’s swimming in. The recruiters who win will be the ones who stop hunting for the tool and start evaluating the human using it.

If you’re a job seeker on the other side of this, the same logic applies in reverse: use AI to express your real experience clearly, not to manufacture experience you don’t have, and make sure your writing still sounds like you. You can paste a draft into PaperBleach and try it free to see how natural and human your application reads before you send it. Want more on how detection actually works? Browse more on the blog or see the pricing.

Frequently asked questions

The Q&As below cover the questions recruiters and candidates ask most: whether AI writing is detectable, whether rejecting on a flag is fair, when to disclose AI use, what to screen for instead, and whether AI use signals dishonesty.