The advice circulating right now is that a quick pass through ChatGPT will sharpen your resume: tighten the phrasing, fix the grammar, make it sound more polished before you send it out. On the surface this seems harmless. It's just editing help, the same way spell-check is editing help.
At director level and above, it isn't harmless, and it isn't just editing help. A generic AI pass doesn't just risk detection, it erases the one thing your resume is supposed to prove: that nobody else could have written this specific document about this specific career. Whether a recruiter can technically "tell" is the wrong question. The real cost is what the AI polish does to the substance underneath, not whether a tool catches it.
The percentage everyone quotes doesn't mean what it sounds like
You'll see this stat everywhere: recruiters can spot an AI-written resume. Depending on which survey you read, the number is 33.5%, 74%, or 88%. Those are all real, published figures from three different US surveys, and they contradict each other because they're measuring different things.
Insight Global asked hiring executives if they believe they can tell. TopResume ran an actual identification test on real resumes and got a lower success rate. Resume Genius asked whether hiring managers had encountered AI content at all, which is a different question entirely from detection accuracy.
Two of these are confidence surveys; TopResume's is a small real-world identification test, and even that isn't a validated detection benchmark. When OpenAI built an actual classifier to detect AI-written text in 2023, it correctly flagged only 26% of AI text and mislabeled 9% of human writing as AI. OpenAI pulled the tool later that year because the accuracy was too low to be useful.
That feeling recruiters describe is still worth taking seriously. Independent tests of AI detection tools show wide swings in accuracy, from roughly 52% to 91% depending on the tool and dataset, and no single detector has been established as a validated, resume-specific standard. None of the figures above should be read as a fixed, agreed-upon percentage, and it's unclear whether they'd hold outside the US, since no comparable large-sample survey there has been identified. The real question is what triggers that feeling in a human reader, because that's the thing you can actually control.
What triggers the feeling isn't a word, it's a rhythm
Most advice on this topic hands you a list of words to avoid: delve, leverage, robust, dynamic. Cutting those words is fine, but it treats the symptom, not the mechanism. A resume doesn't read as AI-written because of one flagged word. It reads that way because every sentence is the same length, every bullet follows the identical structure, and every achievement is described in the same confident, context-free register regardless of whether it was a minor project or the defining move of someone's career.
That uniformity is the actual tell, because human writing doesn't behave that way. A person describing a major product launch writes about it differently than they write about a routine budget cycle, because they remember one and are reciting the other from a template. AI-generated text doesn't have that variance. Everything gets the same weight, the same polish, the same tone.
Take this bullet, run through a generic AI polish prompt: "Spearheaded a strategic initiative to drive operational efficiency, resulting in significant improvements to team performance and stakeholder satisfaction." It's grammatically smooth and says almost nothing. It could describe literally any manager at any company in any industry.
Compare it to something a person who actually did the work would write, because they remember the specific decision that mattered: "Cut order-to-ship time from 11 days to 4 by renegotiating two supplier contracts and killing a manual approval step nobody could justify." The second version has a number, a mechanism, and a decision. It couldn't have been written about a different person's job.
The tell isn't a banned word list, it's a resume where every bullet sounds like it happened to the same generic person.
Where the common advice is right, and where it stops helping you
The advice to use AI as an editing pass isn't wrong on its face. Grammar checking, catching passive voice, restructuring an overlong paragraph: these are legitimate, low-risk uses, and refusing them on principle wastes time. The failure mode isn't using the tool. It's feeding it a thin input and accepting whatever confident-sounding output comes back without correcting it.
The mechanism is straightforward. Large language models generate the statistically most likely next words given a prompt. If your prompt is "make this bullet sound more impressive," the model has no access to your actual budget, your actual team size, or what specifically you decided that other people wouldn't have.
It fills that gap with the most generic phrasing that fits the shape of an achievement bullet. The output sounds impressive and proves nothing, because the model was never given anything specific to prove.
The fix isn't avoiding the tool. It's controlling what goes into it. If you give a model your actual numbers, your actual decisions, and ask it only to tighten sentence structure and cut redundant words, the scope marker and the outcome survive because you supplied them. If you ask it to "sound more senior" or "make this more impactful," you get vocabulary inflation with nothing underneath it, and that's the version that reads as hollow to anyone who's reviewed a lot of resumes.
The interview is where the real cost shows up, not the screening stage
This is the part almost nobody writing about AI resumes addresses, and it matters more at senior level than the detection question does. Suppose the AI-smoothed resume gets through. It's fluent, it's polished, and a recruiter or hiring manager moves you forward.
The resume did its job. Now you're in a room, and the interviewer asks you to walk through the "strategic initiative that drove operational efficiency."
You can't, because you didn't write it. You approved language that sounded good and never unpacked the underlying decision yourself, so when someone asks a second-order question, there's nothing behind the sentence. This is the credibility gap that can cost senior candidates offers: not getting caught by a tool at the screening stage, but going quiet or vague three questions into a follow-up at the stage that actually determines the hire.
A generic AI pass doesn't just risk looking automated. It risks leaving you unable to defend your own resume out loud.
At manager level, an interviewer might let a vague answer slide. At VP or C-level, they're far less likely to, because the entire premise of hiring someone senior is that they can articulate exactly what they did and why it worked, in real time, without a script. A resume that oversells and an interview that can't back it up is a worse outcome than a resume that undersold you and got a smaller first impression.
A decision rule for using AI on your resume without hollowing it out
Before you paste a draft into any AI tool, run it against this:
- Does every bullet already contain a number, a specific decision, or a named mechanism, before you ask AI to touch the wording?
- Are you asking the tool to restructure a sentence, or asking it to invent the substance of the sentence?
- Could you defend every claim in this document, out loud, with a follow-up question you didn't expect?
- If you removed every adjective the tool added, would the bullet still say something true and specific?
- Would two different bullets on this resume, if unlabeled, still sound like they were written by two different levels of seniority and scope?
If the answer to any of these is no, the resume needs more of your own specifics in it before it needs more editing help.
Before: "Led cross-functional teams to deliver innovative solutions that enhanced overall business performance and drove measurable results."
After: "Rebuilt the vendor onboarding process across three teams, cutting time-to-first-invoice from 6 weeks to 9 days and removing $1.2M in annual contract leakage."
Both describe the same underlying work. One reads like it was generated to fill space. The other reads like it happened, because it's specific enough that it couldn't have happened to someone else.
What actually protects you isn't avoiding AI, it's ownership of the substance
The double standard here is worth naming plainly. The same companies penalizing candidates for AI-polished resumes are running AI-driven applicant tracking systems and screening tools on their own side of the process. Nobody is asking recruiters to disclose which parts of their screening are automated. The asymmetry isn't going away, and pretending otherwise doesn't help you.
What does help is treating AI the way you'd treat any other drafting tool: useful for structure, useless for substance. The resume itself is not the deliverable that matters. It's a proxy for a conversation you're going to have to carry, unaided, across multiple interview rounds. If the resume is more polished than you are prepared to be in that room, you've built a liability, not an advantage.
Building a resume you can defend, not just one that reads well
Most people trying to fix this go word by word, hunting for the phrase that sounds "too AI." That's backwards. The fix starts with the underlying material: the actual decisions, numbers, and scope that a generic prompt can't invent because you never gave it to the tool in the first place. Get that right and the phrasing question mostly takes care of itself.
If you want general background on Sanctum before deciding anything, our about page gives a broad outline of the service and who runs it.
If you'd rather talk through your own resume than run it through another checklist, you can book a call and we'll look at what's actually on the page versus what you'd need to defend in a room.
