"Every candidate is cheating with AI" and "it's just people using tools" are both wrong, in instructive ways. Somewhere between a job seeker asking ChatGPT to tighten a CV bullet and a state-sponsored operative interviewing behind a stolen identity, a line gets crossed, and most of the panic in hiring right now comes from refusing to say where. (We build an AI screening tool, so we have a stake in this conversation; read us skeptically too.) Everything below is sourced and dated; the picture is as of July 2026 and moves fast.
The useful skill isn't outrage, it's classification. So before any numbers: a taxonomy.
Three kinds of claims, three levels of trust
Nearly everything written about fake candidates belongs to one of three categories, and they deserve very different weight. Documented incidents are cases with names on them: prosecutions, agency advisories, companies describing what happened to them. They prove existence, not prevalence. Survey self-reports measure what candidates and employers say they do or see; they're evidence of scale but inherit every bias of self-reporting, and "using AI" can mean anything from spell-check to fabrication. Projections are analyst forecasts about the future; they are directionally useful and numerically unfalsifiable, and they should never be quoted as if someone counted something. Most scare content collapses these three into one number. Keeping them apart is the whole discipline.
What's documented: incidents with names on them
The floor of the problem is real, criminal at its extreme, and concentrated in remote technical hiring. The FBI's Internet Crime Complaint Center warned as far back as June 2022 of complaints involving deepfakes and stolen personal information used to apply for remote IT, programming and database roles, including video interviews where lip movement and audio didn't align. The FBI noted why those roles specifically: many of the targeted positions came with access to customer data, financial records and corporate IT systems, which is what separates this category from ordinary CV embellishment. The warning aged well. In July 2025, the US Department of Justice announced the sentencing of an Arizona woman to 102 months in prison for running a "laptop farm" that helped North Korean IT workers pose as US-based employees: the scheme touched 309 US companies, used 68 stolen identities, and generated over $17 million, with the workers' company laptops physically hosted in her home to fake a US location.
The most instructive single case is the security-awareness company KnowBe4, which disclosed in July 2024 that it had itself hired a North Korean operative: a stolen US identity, an AI-enhanced stock photo, multiple video interviews and a background check all passed, and the deception surfaced only when the new "employee" loaded malware onto the company laptop and was caught by monitoring software on day one. The lesson KnowBe4 drew publicly is the right one: a polished hiring funnel verified documents and appearances, and none of those were the thing that needed verifying.
A different kind of documentation sits closer to everyday hiring: the tooling. Interview Coder, an overlay that feeds answers to candidates during live technical interviews, got its student creator suspended from Columbia; he went on to co-found Cluely, which raised a $5.3 million seed round in April 2025 and $15 million from Andreessen Horowitz in June 2025, marketing itself with the tagline "cheat on everything" (funding and origin as reported by TechCrunch). Whatever one thinks of the pitch, real-time answer assistance in live interviews is now a venture-funded product category, not a hypothetical.
What surveys self-report, and what they can't tell you
The survey layer says the ordinary end of the spectrum is now mainstream. In a Gartner survey of 3,000 job candidates (published July 2025), 4 in 10 said they use AI during the application process, mostly to write CV text, cover letters, writing samples or assessment answers, and 6% admitted to genuine interview fraud: posing as someone else, or having someone else interview as them. The same research found the distrust runs both ways: only about a quarter of candidates trust AI to evaluate them fairly, and only half believed the jobs they were applying to were even legitimate. Fakery, it turns out, is a two-sided market.
On the employer side, Robert Half's survey of US HR leaders (fielded November 2025, published March 2026) found 67% say reviewing AI-generated applications has slowed their hiring, 20% report delays of more than two weeks, 84% say their HR teams are overworked by the added review load, and 65% of hiring managers say the surge of AI-enhanced applications has made verifying candidate skills harder. Note the mechanism: AI didn't just change what applications say, it changed how many there are, and a screening process built on human reading time degrades on both axes at once.
Read all of these with the self-report caveat attached. The 6% who admit fraud is a floor, not a ceiling; the 4 in 10 using AI spans everything from grammar cleanup to invented experience; and employer perceptions of "AI-generated" applications are guesses, since text detectors are unreliable and polished writing is not evidence of anything except polish.
What's projection, not measurement
The most-quoted number in this space is Gartner's forecast that by 2028, one in four candidate profiles worldwide will be fake (July 2025). It's a projection: a scenario extrapolated by an analyst firm, not a count of anything that has happened. Quote it as what it is. If it's directionally right, identity assurance becomes a standard hiring layer within a few years; if it's wrong, it will be because employers adapted, which is the same reason to adapt either way. What it is not is evidence that a quarter of your current applicant pile is fictitious.
Polish is not fraud: drawing the line that matters
A candidate who uses AI to fix grammar, tighten phrasing, or tailor a CV to your posting has not cheated; they've used a writing tool, and per the survey data above, so did much of your pipeline. Treating the AI-generated resume as disqualifying by itself is both unenforceable (detectors misfire, and fluent human writers look "AI-generated") and beside the point. The CV was always a marketing document; AI just made good marketing free.
What actually broke is any screening method that graded the marketing. Keyword filters were trivially gamed before generative AI; now the gaming is automated, and a filter that matches vocabulary is simply selecting for the best text generator. The same logic applies to interview answers: candidates using ChatGPT in interviews, whether through a hidden overlay or a rehearsed script, produce fluent, structured, general-knowledge answers on demand. What the tools cannot supply is what they've never seen: the specific system the candidate built, the real tradeoff they made, the number that only exists if the project did.
So the line worth enforcing is not "did AI touch this text" but "do the claims survive contact with evidence." Polish is fine. Fabricated experience is misrepresentation. Impersonation is a crime with sentencing guidelines, as the documented cases above show. Say this to candidates explicitly: state what AI use you consider acceptable in your process, and you convert an ambiguous norm into an honest one, which is also Gartner's own first recommendation to employers. A workable policy fits in three sentences:
You're welcome to use AI tools to edit, structure and prepare your application materials; everything in them must describe work you actually did. During interviews, answer from your own knowledge, without real-time assistance from any tool or person. Every part of the process must be completed by you.
Publish something like that in the posting or the interview invitation, and the honest majority knows the line while the dishonest minority loses the "everyone does it" cover.
What employers are doing about AI interview cheating
The visible response is a partial retreat to physical presence. Google now ensures at least one in-person interview round for candidates, as CEO Sundar Pichai confirmed in June 2025 on Lex Fridman's podcast, after internal pressure from engineers reported by CNBC that spring. Interestingly, candidates approve: in the same Gartner research, 62% said they'd be more likely to apply to a role that required in-person interviews, presumably because a company that verifies candidates is also more likely to be offering a real job.
In-person rounds are honest evidence-gathering, but note what they cost: they re-import every scheduling, geography and throughput constraint that remote hiring removed, which is why they've returned as a final round, not a first one. The other documented responses are layered rather than theatrical: identity and background verification for sensitive roles (the DOJ and KnowBe4 cases were both caught by verification and monitoring layers, not by interviewers' intuition), work samples and probing follow-ups over rehearsable question banks, and explicit AI-use policies communicated up front.
One response to avoid: running applications through AI-text detectors and rejecting on the result. Beyond their general unreliability, a Stanford-led study found GPT detectors systematically misclassify writing by non-native English speakers as AI-generated (Liang and colleagues, Patterns, 2023), which turns an authenticity check into an accent filter in text form. A crackdown that can't distinguish fraud from fluency creates the discrimination problem your bias controls exist to prevent.
Can you screen for authenticity without a lie detector?
Yes, because authenticity has a testable signature: specificity under probing. A candidate who did the work can name the system, the constraint, what broke, what they'd change; an answer assembled by a language model or a coach stays general one level below the surface. You don't need to detect the tool. You need an interview that asks for evidence and follows up on what it hears, and a scoring method that records where the evidence ran out.
Where rubric scoring fits, and where it honestly doesn't
Our angle, stated with our interest on the table: this is what rubric-first screening is structurally good at. In Rubrily, every answer is scored 0–10 per criterion against your rubric's evidence definitions, with a written justification citing what the candidate actually said. Generic answers are thin on exactly the thing the rubric pays for: a fluent paragraph with no named system, no real tradeoff and no checkable detail scores middling at best, and the justification says why in plain text ("no specific example provided; described general best practices"). Adaptive follow-ups push on the candidate's own claims, which is where assisted answers are weakest, and when there's no evidence to score, the system returns "Cannot evaluate" rather than inventing a number. The whole chain stays auditable, so when a score is challenged, you can show exactly where substance was and wasn't.
Async interviews also run monitored, with the candidate's informed consent: tab switching, extended absence from camera, and AI-generated answers can be detected and surfaced as flags on the report. Flags are context for a human reviewer, never auto-rejection, and that design is deliberate: integrity signals can misfire, and oversight only means something when a person weighs the evidence instead of rubber-stamping the machine.
And the honest boundary: Rubrily does not do lie detection, facial or emotion analysis, deepfake detection, or identity verification, and no scoring system can tell you whether a claim is true. A rubric score measures the quality and specificity of evidence a candidate presents; verifying that the person is who they say they are, and that the named project existed, remains the job of humans and process: references, background checks, verification layers for sensitive roles, and a final conversation with a person. What explainable scoring contributes is knowing precisely which claims your decision rests on, so you know what to verify before you extend the offer.
FAQ
How do recruiters detect AI-generated resumes? Reliably, they don't: text detectors misfire in both directions, and polished writing proves nothing. The workable approach ignores authorship and tests substance: probe the CV's specific claims in a structured interview with follow-ups, and score answers against evidence-based criteria rather than fluency.
Can candidates cheat in AI interviews? Assist tools exist and are documented, so no format is cheat-proof. Structured probing raises the cost: follow-ups on the candidate's own claims expose generic answers, evidence-anchored scoring grades specificity, and consent-based monitoring can flag tab switches, camera absence and AI-generated answers for human review.
How common are fake job applicants? Separate the claim types: documented impersonation cases are real but concentrated in remote IT hiring (one DOJ-prosecuted scheme touched 309 US companies); in surveys, 6% of candidates admitted interview fraud (Gartner, 2025); the "1 in 4 profiles fake by 2028" figure is a Gartner projection, not a measurement.
Is it wrong for candidates to use AI in applications? Using AI to polish or tailor honest content is now mainstream (about 4 in 10 candidates, per Gartner) and isn't cheating. Fabricating experience is misrepresentation; interviewing behind someone else's identity is fraud. State your acceptable-use policy openly so candidates know exactly where your line sits.
Screening that grades evidence, explains every score, and says "Cannot evaluate" instead of guessing. Join the waitlist →