How candidates use AI to cheat hiring assessments

MatchCard · 19 July 2026 · 7 min read

Quick answerAI cheating in hiring assessments means candidates using ChatGPT or similar tools to generate, polish, or feed answers during screening. In a 2026 sample of 19,368 real interviews, 38.5% showed clear signals of AI-assisted cheating, and most still passed because the process wasn't looking for it.

In a sample of 19,368 real interviews analysed between July 2025 and January 2026, 38.5% showed clear signals of AI-assisted cheating. Not suspected: flagged, by systems built to look for it. And most of those candidates still passed, because the processes screening them weren't looking.

This isn't a fringe behaviour that better job ads will fix. It's a structural change in how people apply for jobs, and it demands a structural change in how employers screen. Here's what the cheating actually looks like, why traditional detection fails, and what catching it genuinely requires.

The four levels of AI-assisted applying

"Cheating with AI" covers behaviours with very different levels of intent, and it's worth being precise about which level you're dealing with, because they demand different detection strategies and carry different risks:

1. Polish. A candidate drafts their own answer and asks AI to tidy the grammar. It's nearly universal now: a Gartner survey found 39% of candidates already using AI somewhere in the application process back in late 2024, and the number has only moved one way since.
2. Generation. The candidate pastes the question into a chatbot and submits what comes back, lightly edited or not at all. This is where authenticity genuinely breaks: the answer no longer contains information about the person.
3. Live assistance. A second screen feeding answers during a live or recorded interview. In one survey of 4,136 people reported in the same research, 91% of hiring professionals said they'd encountered or suspected AI-generated answers in online meetings.
4. Identity fraud. The far end: someone else, or something else, takes the interview. 6% of candidates surveyed by Gartner admitted to interview fraud, posing as someone else or having someone pose as them, and Gartner predicts that by 2028, one in four candidate profiles worldwide will be fake outright.

Why the obvious countermeasures don't work

"AI detectors" don't solve it. Statistical AI-text detectors are notoriously unreliable on short answers, penalise non-native English speakers, and are trivially defeated by asking the model to "write casually." Running answers through a detector and trusting the percentage is closer to superstition than screening.

Proctoring escalates, but doesn't discriminate. Lockdown browsers and webcam monitoring treat every candidate as a suspect, wreck the candidate experience (remember that only 26% of applicants trust AI screening to be fair as it is), and still miss the second phone sitting next to the laptop.

Reverting to in-person doesn't scale. In-person interview requests jumped from 5% to 30% in a year, largely as a fraud response. But as a first screen, it means paying interview-stage costs for every applicant, which is exactly what screening exists to avoid.

What detection actually requires: reading, not scanning

The approaches that hold up share a principle: stop trying to detect the tool, and start evaluating the evidence. A generated answer fails in characteristic ways that have nothing to do with statistical text patterns:

• It's specific to nothing. Real answers contain the residue of real events: names withheld but shapes visible, numbers that are oddly exact, details nobody would invent. Generated answers are smooth, structured, and could belong to anyone.
• It doesn't know what the question was for. A good screening question is engineered to force a disclosure: a trade-off, a failure, a conviction. Generated answers routinely deliver a competent essay on the topic while dodging the disclosure entirely, because the model doesn't know the question had a target.
• It can't stay consistent under cross-examination. One answer can be faked. Ten answers, a CV, a cover letter and a salary expectation, all checked against each other, is a different problem: voice shifts mid-assessment, the self-rating contradicts the story, the CV timeline disagrees with the anecdote. Consistency is expensive to fake and cheap to check, if the system is actually checking.

This is the design behind MatchCard's authenticity read: every answer is evaluated for what the question was engineered to elicit versus what came back, quoted evidence is required for every judgement, and everything a candidate submits is cross-examined against everything else. Not a percentage from a black-box detector, but a documented case, defensible to the candidate themselves. Most of the detection tooling on the market is built specifically for live coding interviews; see how the same problem plays out in scenario and behavioural interviews, the format most non-technical roles actually use.

The part most vendors skip: cheating has a cause

One more thing belongs in this picture. Candidates aren't cheating in a vacuum. They're responding to a process they believe is already automated against them. Most assume an AI reads their application before a human does, and three-quarters don't trust it to judge them fairly. Using AI back starts to look less like fraud and more like arms parity.

Which suggests the durable fix isn't only better detection: it's a process worth being honest in. Shorter, role-specific, transparently read, and honest in both directions: the employer discloses the real salary band and the real day-to-day, the candidate answers real questions about fit. That exchange is harder to fake and, more importantly, less worth faking.

Frequently asked