How to Tell If a Resume Is AI-Generated, and What to Do

In August the head of hiring at one of our pilot customers asked us for an AI check on resumes, with one condition attached: “I don’t want it to reject, I want it to flag.” A resume that reads like ChatGPT wrote it and has nothing real underneath was their call to reject. A resume that reads like ChatGPT wrote it and describes real, specific work was fine.

That distinction is the whole answer to “how do I tell if a resume is AI-generated”, and most of the advice ranking for that search skips it. So this post covers both halves: the signs worth noticing, and what to do once you notice them.

How can you tell if a resume is AI-generated?

You usually cannot tell for certain, and you do not need to. The useful signs are generic phrasing, uniform bullet points, achievements with no numbers or names, and a level of polish the rest of the application does not match. Treat them as a reason to check the claims, not as proof, because a person can write that way and a good AI edit can hide it.

The reason certainty is out of reach is simple. Resumes were formulaic long before AI: action verb, task, result, repeat. A model trained on millions of them produces exactly that pattern, and so does a careful human using a template. What separates a weak resume from a strong one was never the prose. It was whether the claims hold up.

8 signs a resume was written by AI

None of these proves anything alone. Two or three together, on a resume with thin substance, is worth a closer look.

1. Interchangeable verbs. “Spearheaded”, “leveraged”, “orchestrated”, “drove cross-functional synergies”. Each bullet could sit on any resume in any industry.
2. Bullets of identical length and rhythm. Real resumes are uneven because real jobs are uneven. Eight bullets of exactly two lines each, all built as verb, task, impact, read like one generation pass.
3. Achievements without anchors. “Improved customer satisfaction significantly” with no number, no product, no customer, no date. AI fills space with outcomes it cannot know.
4. Seniority the detail does not support. A two-year analyst who “defined the data strategy” and “aligned executive stakeholders” but cannot name a single dashboard, table or tool.
5. The job description, rephrased. Every requirement in your posting appears in the resume, in the same order, in slightly different words. We wrote about why keyword match is not fit; AI makes this trivial to produce.
6. A summary that says nothing. “Results-driven professional passionate about delivering value in dynamic environments.” It is grammatical, confident and empty.
7. Hidden text. Paste the resume into a plain text editor. White-on-white keywords, or a line addressed to an AI screener, show up immediately. This one is not about AI writing at all; it is an attempt to game the screen, and it is the only sign here that justifies a reject on its own.
8. A mismatch with everything else they wrote. A flawless resume next to a one-line application answer or a cover note full of errors. People write differently under different conditions, so this is a nudge, not a finding.

Notice what is missing from the list: perfect grammar. Clean writing is not a sign of AI. It is a sign of someone who cared, or who asked a friend, or who used a tool the way people have used spell-checkers for thirty years.

Why AI detectors are the wrong tool for resumes

The obvious shortcut is to paste each resume into an AI detector. Two pieces of evidence say not to.

The company that built the best-known model could not make detection work. OpenAI launched an AI text classifier in January 2023 and withdrew it on July 20, 2023, citing its low rate of accuracy. In its own evaluation it identified only 26% of AI-written text as likely AI, and it wrongly labeled human-written text as AI 9% of the time. OpenAI also said it was unreliable on short inputs. A resume is a short input.

Detectors penalize people who write in a second language. A Stanford team (Liang and colleagues, published in Patterns, 2023) ran seven widely used detectors on 91 TOEFL essays written by non-native English speakers. On average the detectors flagged 61% of them as AI-generated, and 89 of the 91 essays were flagged by at least one detector. US students’ essays were classified almost perfectly.

The detail that should end the debate: when the researchers ran the same non-native essays through ChatGPT to make them sound more native, the false positive rate fell to about 12%. The detector punished people for not using AI.

For any team hiring in India, or hiring anyone who learned English as a second language, that is not an edge case. It is most of the applicant pool. A detector-based reject would quietly filter out careful, plain writers and let through anyone who asked a model to polish their resume. That is the opposite of what you wanted.

Can an ATS detect AI-written resumes?

Mostly no, and that is by design. An applicant tracking system parses a resume into fields (name, employers, dates, skills) and matches those against the job. A tidy AI-formatted resume parses as well as a tidy human one, often better. Some vendors now show an AI-likelihood score inside the ATS; it inherits every problem above, so read it the way you would read a detector: as a reason to look, never as a reason to reject.

Should you reject AI-generated resumes?

Not for being AI-generated. Reject for what the AI was used to do. It helps to separate three cases, because they need different responses:

The SHRM framing of this is a fine line between enhancement and dishonesty, and the line is drawn in the conversation, not on the page. That also gives you a defensible record: “could not describe the project listed as their main achievement” is a reason you can stand behind. “Our detector said 74%” is not.

How to check the claims behind a polished resume

The point of noticing the signs is to choose what to ask. A short routine that works in a 10 to 20 minute first round:

1. Pick the two or three biggest claims on the resume. The ones that, if true, would make this person a strong hire.
2. Ask about the first step, not the result. “Your resume says you cut onboarding time by 40%. What did you do first?” Someone who did the work can name the first step, the person they spoke to and what changed next. A rehearsed or generated claim usually has a result and no first step. Our list of interview questions to ask candidates has thirty more with the follow-up to use.
3. Ask for one artifact. A query they wrote, a deck, a ticket, a design, a number from a report. Not to keep it, just to hear them describe it.
4. Compare how they talk with how they wrote. Fluent bullet points and a hesitant, general spoken answer are worth a second question. A plain resume and a vivid, specific answer are a candidate the resume undersold.
5. Ask about AI openly. “Did you use any AI tools on your resume?” Many will say yes and explain how. The answer tells you about their judgment, and it removes the guessing.

Steps 2 and 4 are where AI help stops helping. A model can write the bullet; it cannot sit in the call and explain what happened on the Tuesday the migration failed. That is also why candidates who lean on ChatGPT during a live interview give themselves away through timing and delivery rather than through their wording: a long pause, then a fluent answer.

If you are setting up the first round itself, our guide to pre-screening interview questions covers the knock-out questions and the scoring that sit around these follow-ups.

What we built after that request for a flag

Back to the hiring lead who wanted a flag and not a reject. Here is what we shipped that week, and the rules we gave it.

When HireQwik’s resume screening scores a candidate against the job, it also estimates how much of the resume’s presentation looks AI-generated, on a 0 to 100 scale, with up to four short observations of what it noticed: stock phrasing, uniform bullets, claims with no numbers or dates. In the review queue, resumes at 60 or above get an “AI-written” chip, and the chip is highlighted at 80 or above.

The rule that matters: the flag never changes the resume score and never rejects anyone. The score comes from the evidence of fit in the resume, which is where an inflated resume already loses points, because vague claims carry little evidence. The flag only tells the recruiter where to look harder. A high flag on a resume with real substance is fine; a high flag on a resume with nothing underneath is the recruiter’s call, which is exactly what they asked for.

The second half is the conversation. A job can be set so that our AI interviewer asks each candidate three questions built from their own resume, such as how they implemented the project they listed, scored against a rubric written for those answers. That is step 2 from the routine above, run for every applicant instead of the few a recruiter has time to call.

The pattern behind both is the one from our earlier post on AI resumes: when the document stops being a reliable signal, move the check to the part a model cannot do for the candidate. Stop asking “did AI write this?” and start asking “can this person explain what they wrote?”