Is Using AI During a Job Interview Cheating? What Hiring Managers Actually Catch

How Common Is It, Really?

Resume Genius's 2026 Job Seeker Insights Report, based on a survey of 1,000 active U.S. job seekers published May 4, 2026, found that 78% of job seekers use AI at some point during their job search — resumes, cover letters, company research, and practice questions among them. That figure alone isn't controversial; using AI to prepare is now closer to the norm than the exception.

A narrower, more specific figure from the same report is the one that actually matters here: 22% of job seekers admitted to using AI during the interview itself, in real time — not beforehand, during.

• 78% use AI at some point during their job search (Resume Genius, 2026 Job Seeker Insights Report, n=1,000 active U.S. job seekers, published May 4, 2026)
• 22% specifically admit to using AI live, during the interview, to help generate an answer
• HR consultant Bryan Driscoll told the outlet he suspects the real figure is higher: "That's probably low"

Driscoll's fuller comment cuts both ways: "If your interview can be passed by ChatGPT in real time, you're not interviewing effectively." That's a criticism of lazy interview design — but it's also a warning for candidates. If a question really can be fully answered by a generic AI response with no connection to your own background, a careful interviewer is going to notice that your answer sounds like everyone else's.

“22% of job seekers admit to using AI live, during the interview itself — not just to prepare for it.”

What Hiring Managers Say They've Actually Caught

The employer side has its own 2026 Resume Genius report, fielded through the survey platform Pollfish among hiring managers and hiring team members. Three numbers from it are worth knowing: 46% say they've personally seen a candidate use AI to answer interview questions; 82% say they're concerned about candidates using AI tools somewhere in the application process; and 17% say they've specifically caught a candidate using deepfake technology during a video interview.

A fourth figure says something broader about where this is heading: 86% of hiring managers agree that AI is going to make verifying a candidate's actual identity and abilities harder over the next year — not just in interviews, across the whole hiring process.

Worth flagging plainly: this is a single vendor-run survey, not an independently replicated or peer-reviewed study, and the published write-up isn't even fully consistent about its own sample — the article's introduction cites 1,500 U.S. hiring managers and team members, while its methodology section separately states a target sample of 1,000. Resume Genius is a resume- and career-tools company, which gives it a commercial interest in coverage about AI anxiety in hiring. None of that makes the directional finding implausible — it does mean these are self-reported, vendor-published numbers worth treating as a signal, not a precise count.

The same authenticity worry shows up in a related but separate problem this blog has covered before: whether what a camera shows during an AI-moderated interview can itself be gamed — a different mechanism, the same underlying question: can anyone fully trust what a screen is showing them.

The Real Distinction: Preparing With AI vs. Performing With It

Neither survey distinguishes between two very different behaviors that get flattened into "using AI in an interview." The first is using AI beforehand — researching the company, drafting and refining talking points, or stress-testing an answer against likely follow-up questions, which is the subject of this blog's own piece on answering the "how do you use AI" interview question well — and it's unambiguously normal, expected, and arguably something interviewers are now actively screening candidates to be good at, not avoid.

The second is using AI during the conversation itself — reading, paraphrasing, or lightly editing an answer a model generated seconds earlier, in response to a question you haven't actually thought through yourself. That's the behavior the 46%/22% figures above are describing, and it's the one that creates real risk for a reason that has nothing to do with getting caught: it answers a question nobody actually asked you to answer.

A hiring manager isn't scoring whether an answer sounds polished in isolation — they're scoring whether you, specifically, can produce that judgment under real conditions. This blog's own argument about AI and careers is that the gap that matters is between people who use AI well and people who don't, and reading a model's output verbatim in real time is close to the opposite of using it well — it skips the judgment entirely and keeps only the words.

The line isn't whether AI touched an answer at any point before the interview. It's whose thinking is actually in the room right now.

What Actually Gives It Away

Neither Resume Genius report describes the mechanics of how hiring managers notice. For that, the clearest account comes from a different and well-established source: Alison Green's long-running workplace-advice column, Ask a Manager, where a hiring manager wrote in on November 19, 2025 describing an interview that ended early once a pattern became obvious.

• Camera avoidance — the candidate claimed "technical difficulties" and couldn't turn her camera on, despite a policy that interviews be conducted on camera
• Restating the question — answers repeated the question back rather than actually addressing it
• Buzzwords without substance — plenty of jargon, with no specific detail attached to it
• A flat, scripted cadence — speech that sounded like it was being read rather than spoken
• Generic answers to simple prompts — even a reply to "do you have any questions for us?" sounded AI-generated

Readers who responded to that post added two more patterns worth knowing: eyes visibly tracking across the screen before and during an answer, with a noticeable delay between the question ending and the answer starting; and uneven performance by question type — doing fine on direct, factual, technical questions, then shifting into an obviously different register for open-ended "tell me about a time" questions, which are exactly the ones a generic model fills with the most plausible-sounding, least specific material.

That last pattern tracks closely with something else this blog has written about rehearsed answers generally: genuine, lived experience behaves differently under a follow-up question than a rehearsed or generated one does — the tell usually isn't the first answer, it's what happens when someone asks you to go one layer deeper.

Green's advice to interviewers is direct: name it in the moment ("it sounds like you may be looking up answers and reading them"), give the candidate one chance to answer in their own words, and end the interview if the pattern continues. Her read, echoed by multiple commenters describing similar calls of their own, is that even a candidate who corrects course after being called out typically still takes a real hit to how they're evaluated — recovery is possible, but reportedly rare.

“The tell usually isn't the first answer. It's what happens when a follow-up question asks for one layer deeper.”

Why It Backfires Even When You're Not Caught

Set the risk of getting caught aside entirely for a moment. Driscoll's framing above is worth sitting with: if a question can be fully answered by a model with zero knowledge of your actual background, that's informative on its own — either the question was too generic to be worth asking, or the answer wasn't actually about you.

There's a second cost that has nothing to do with the interview you're in and everything to do with every interview after it. The entire premise behind structured, repeated interview practice is that the skill being tested is a skill, built the same way any other skill is: through reps, feedback, and adjustment, not through access to a tool in the moment. Leaning on AI to generate an answer live doesn't just risk that one conversation. It substitutes a one-time patch for the practice that would have made the next interview, and the one after that, genuinely easier without a laptop open.

This blog has made the same argument about an earlier stage of the funnel — using AI to improve a resume is normal, but outsourcing it entirely tends to produce something generic rather than specific — for exactly the same underlying reason good interview answers fail when they're AI-generated in the room: specificity is the thing being scored, and specificity is the thing a generic model can't supply about a life it doesn't know.

What To Do Instead

The practical version of all of this is simpler than it sounds:

• Use AI before the interview, not during it — to research the company, draft and refine two or three detailed stories, and pressure-test them against likely follow-up questions.
• If you blank on a question, pause and think out loud rather than reach for a device. A visible pause reads as normal; a sudden shift into polished, generic phrasing right after one does not.
• Keep your examples specific enough that a generic model genuinely couldn't have produced them — names, numbers, what you personally did versus what the team did, what went wrong before it went right.
• If the format is remote or one-way recorded rather than live, the same logic still applies: a recorded answer that sounds written rather than spoken reads the same way to whoever reviews it later.

None of this requires performing confidence you don't actually have. It requires having gone through the reps enough times beforehand that the answer in the room is actually yours — which turns out to be a harder thing to fake than it is to build.

Key Takeaways

If you remember only a few things, remember these:

• 22% of job seekers admit to using AI live, during an interview, to help generate an answer, and 46% of hiring managers say they've personally caught it (Resume Genius, 2026 reports) — treat both as vendor-published, self-reported figures, not precise counts.
• Using AI to prepare beforehand is normal and increasingly expected. Using it live, in the room, to generate your actual answer is the specific behavior that creates real risk.
• Reported tells have nothing to do with sophisticated detection software: a flat or scripted cadence, answers that restate the question, buzzwords without specifics, a delayed start paired with eyes tracking across a screen, and strong factual answers that suddenly go generic on open-ended questions.
• Even candidates who correct course after being called out typically still take a real hit — recovery is possible but reportedly rare.
• The safer move isn't a better way to hide it. It's building specific, rehearsed-but-genuine answers beforehand so you never need to reach for one in the room.