Six percent of candidates admit they have participated in interview fraud โ posing as someone else, or having someone else pose as them. Thirty-nine percent say they used AI during the application process, including to generate answers to assessment questions. Both numbers come from Gartner surveys of roughly 3,000 job candidates each, and they underwrite the projection that by 2028, one in four candidate profiles worldwide will be fake (Personnel Today, 2026).
Identity verification detects the first number. It does not touch the second.
That distinction is where most mid-market hiring budgets are about to go wrong. The tooling wave arriving this quarter โ document checks, liveness detection, deepfake screening layered onto the video interview โ is built against the impersonation vector. The impersonation vector is real, and it is the smaller one. The larger, faster-growing, and almost entirely undefended vector is the correctly-identified real person whose demonstrated competence was manufactured by a model.
Two Fraud Vectors, One Budget Line
Candidate fraud is routinely discussed as a single category. Operationally, it is two, and they fail in different places.
The impersonation vector asks: is this person who they claim to be? A proxy sits the technical screen. A synthetic identity clears onboarding. A deepfaced video feed carries an interview in real time โ Experian's 2026 fraud forecast singles out exactly this, generative tools producing hyper-tailored resumes and candidates capable of passing interviews live (Experian, 2026). This vector carries genuine security exposure, which is why it gets funded. As Gartner's Jamie Kohn put it, candidate fraud "creates cybersecurity risks that can be far more serious than making a bad hire" (Personnel Today, 2026).
The capability vector asks a different question: is the competence on display this person's own? Here the candidate is exactly who they say they are. Their ID is valid. Their face matches. Their CV, cover letter, work sample, and assessment answers were substantially produced by a model they will not have access to in the same way on the job โ or will have access to, but without the judgment to supervise it.
Identity verification resolves the first question completely and the second not at all. A verified identity attached to an unverified capability is a cleaner-looking bad hire, not a prevented one.
What Identity Verification Actually Proves
Identity verification is worth buying for what it does. It closes proxy interviewing, synthetic identities, and the sanctions and right-to-work exposure that sit behind them. If you are hiring remotely into roles with system access, that is not optional hygiene.
The failure is in the accounting. When a verification layer is installed, the fraud line on the risk register tends to get marked closed. It has not been closed; it has been halved. And the half that remains is the one growing fastest, because it requires no criminal intent, no technical sophistication, and no coordination โ just a candidate with a browser open in a second window.
Gartner's own read is that employers are finding it "harder to evaluate candidates' true abilities, and in some cases, their identities" (Personnel Today, 2026). Note the ordering. Abilities first, identity second and qualified. The ability problem is the general case.
The Capability Vector Is Bigger โ and Measurable
Self-report puts AI use in applications at 39%. Behavioral data from inside the interview puts the number higher.
Fabric analyzed 19,368 AI-conducted live interviews and found cheating signals in 38.5% of them, with cheating rates roughly tripling in late 2025 (Fabric, 2026). The distribution matters more than the headline: technical roles showed roughly four times the rate of sales roles, junior candidates roughly twice the rate of senior ones. Among repeat candidates, the behavior is bimodal โ a large group never cheats, a substantial group cheats in every interview they take.
This is vendor platform data, not peer-reviewed research, and it comes from a company selling detection. Discount it accordingly. But it is derived from observed behavior across nearly twenty thousand interviews rather than from asking people whether they cheated, and it moves in the same direction as Gartner's self-report from the opposite methodological side. When the honest-answer number and the observed-behavior number both land near four in ten, the range is the finding.
One detail in that dataset deserves to change a policy at most mid-market companies: take-home assessments showed higher cheating rates than live interviews, because unlimited time and privacy are exactly the conditions under which model assistance is undetectable (Fabric, 2026). The take-home exercise was the low-cost, candidate-friendly compromise most 50โ500 FTE teams adopted to avoid burning interviewer hours. It is now the least defensible step in the funnel.
The Counter-Argument: Isn't This Just the New Normal?
The strongest objection is that AI-assisted applications are not fraud at all. Candidates use the tools available; employers use screening models on the other side; a polished AI-drafted cover letter predicts nothing worse than a professionally-edited one did in 2019. Policing it is both futile and hypocritical.
Half of that is right. Using a model to tighten a CV is not fraud, and a company running AI screening while forbidding AI applications is in a weak ethical position. If the argument stopped there, the correct response would be to stop pretending the resume signal was ever strong and move on.
But the objection conflates drafting with demonstration. When an assessment exists to measure whether this person can do this reasoning, and the reasoning is done by a model, the artifact no longer measures the thing it was designed to measure. That is not a moral failure on the candidate's part. It is a measurement failure on yours โ and measurement failures are the hiring team's problem to fix, regardless of who caused them.
The practical consequence is not that you should catch and punish more candidates. It is that a large share of your funnel now consists of signals that no longer carry information, and you are still making decisions as if they do. Candidates already sense the asymmetry: only 26% believe AI will evaluate them fairly (Personnel Today, 2026). You are not defending a trusted process. You are rebuilding one.
Reorder the Funnel Before You Buy Another Layer
The standard funnel is resume โ recruiter screen โ video interview โ assessment โ onsite. Verification tooling gets inserted at the video interview, late and expensively.
Gartner's remedy is not primarily a tooling purchase. It is assessment with anti-cheating safeguards built in, including in-person components, plus system-level validation instead of individual surveillance (Personnel Today, 2026). And the candidate-side objection to that is weaker than most teams assume: 62% of candidates said they were more likely to apply to a position when the organization required an in-person interview.
The selection science supports the reordering independently of any fraud argument. Sackett and colleagues' revised meta-analytic estimates put structured interviews at an operational validity of about .42 and work samples at about .33 โ the latter a sharp downward revision from the long-cited .54 (Sackett et al., 2022). Neither is perfect. Both are measurable. The resume has never had a validity coefficient worth quoting, and it is now the single most AI-contaminated artifact in the process.
Put those together and the move is structural: run a proctored, capability-based assessment earlier in the funnel โ before the resume-and-video screen carries decision weight. One step then catches both vectors. An impostor cannot pass a live capability check they cannot delegate, and a real candidate with manufactured credentials cannot either. You are not adding a fraud control. You are replacing a low-validity step with a higher-validity one and getting the fraud control as a side effect.
What to Change This Quarter
Audit which of your steps are still measuring anything
Take your funnel and mark each stage with what it would cost a motivated candidate to fake it today. Resume: near zero. Cover letter: zero. Take-home exercise: near zero and quiet. Live structured assessment: high. Most mid-market funnels will find that three of five stages have moved to near-zero cost, and that no one has re-weighted the decision accordingly.
Move one assessment earlier, and proctor it
Not a new vendor. One existing step, relocated ahead of the screen, run live or proctored rather than as a take-home. Measure the change on two things: fraud incidents caught, and time-to-decision. The second usually improves, because a capability signal early kills bad pipelines before they consume interviewer hours.
Buy identity verification for what it is
Fund it, scope it honestly, and write on the risk register that it addresses impersonation only. The capability vector stays open until the assessment layer moves. A control that is documented as partial gets revisited. A control marked "done" does not.
One Decision
Look at where in your funnel a candidate first has to demonstrate something they cannot delegate. If that moment sits after the resume screen, the video interview, and the take-home โ and for most 50โ500 FTE teams it does โ then every decision you make before it is running on artifacts a model can produce for free.
Identity verification will tell you the person on the screen is real. Only assessment tells you the competence is. Move one step this quarter, and make it that one.
