Hiring integrity

The candidate who is not who they say they are

Two different problems get filed under one word. One is a person being someone else. The other is a machine doing the work. They need different evidence and different answers, and a product that treats them as one thing solves neither.

Two problems, one word

Impersonation — the person is not the person

A proxy sits the test. A different face appears on day one. In the extreme, the whole identity is manufactured. This is not a hypothetical: the US Department of Justice has documented co-conspirators compromising the identities of more than 80 US persons to obtain remote jobs at more than 100 US companies between 2021 and October 2024, with the FBI searching 29 laptop farms across 16 states. One facilitator was sentenced to 102 months for placing overseas workers into remote IT roles at more than 300 US companies.

Sources: US DOJ, coordinated nationwide actions · US DOJ, sentencing in the $17m IT-worker scheme

Gartner projects that by 2028 one in four candidate profiles worldwide will be fake, and in a 2Q25 survey of 3,000 candidates, 6% admitted to interview fraud — impersonating someone else or having someone pose as them.

Source: HR Dive, reporting Gartner (31 July 2025)

Substitution — the work is not their work

The person is real. The output is not theirs. This is now cheap and commoditised: overlay assistants sell for $20–50 a month and render their output in a layer that screen capture does not see.

The numbers circulating for how often this happens should be read with the label attached. Fabric reports 38.5% of candidates flagged for cheating behaviour across 19,368 interviews, rising to 48% for technical roles. CodeSignal reports technical assessment cheating going from 16% to 35% in a year. Karat estimates 80% of candidates use language models during code tests even when explicitly banned. Every one of those figures comes from a company selling the detection. We found no independent, non-vendor measurement. Treat the direction as well evidenced and the magnitudes as unverified — including when we would benefit from you believing them.

Why detection is the losing side of this

The most useful statement on this subject was published by a competitor. In a post titled “Why it's impossible to prevent cheating”, TestGorilla writes that no assessment “has ever been 100% cheat-proof”, and that its detection “doesn't and cannot tell you who cheated; it can point to behaviours which you can contextualise”. Its recommendation is to stop building walls and move to live, real-time formats where outsourcing the answer stops being practical.

Source: TestGorilla — Why it's impossible to prevent cheating

Two further reasons the forensic route fails. Detectors for machine-written code carry high false-positive rates, so the tool that flags a cheat also accuses honest candidates — and an accusation you cannot substantiate is a worse outcome than no signal at all. And the countermeasure market moves faster than the detection market, because it has more customers.

What our recording catches, and what it does not

The structural answer

If detection is unreliable, the only remaining lever is asking for something a proxy or a model cannot produce on the candidate's behalf. Three things do that, and none of them is a monitoring feature.

1. Make them explain their own work, live

This is the cheapest fix available to any employer, including one that never buys anything from us. A candidate who did not write the solution cannot walk you through the decision they did not make, and the failure is obvious within two minutes. It is also the moment that produces the incident most engineering leaders describe: a 95% take-home score, and a follow-up call where the candidate cannot read their own code.

2. Assess in a format where the answer is spoken in real time

For sales, collections, support and any customer-facing role, Savvanta's assessment is a live simulated call with an AI customer, scored on what was said and when it was said. There is no window in which to consult anything. This is also, precisely, what TestGorilla's own post recommends and does not sell.

3. Bind the person to the attempt for its full duration

Webcam, microphone and a full-monitor screen share are collected in a single gate before the attempt exists. Deny any of them and the assessment does not start and the clock does not run. There is no reduced-proctoring path and no employer toggle that switches it off, so two candidates for the same job are never assessed under different conditions.

The part that gets skipped

Anti-fraud measures hurt honest candidates when they are built carelessly, and the damage is invisible to the buyer because it looks like a low score. Two things we do specifically to stop that, both of which make our numbers look worse rather than better:

• Dropped recording chunks are counted and reported separately from performance. A partial recording and a weak candidate look identical downstream. Upstream bandwidth correlates with income and region, so silently grading whatever survived is both invalid and a fairness problem.
• Screen capture is dropped where it proves nothing. For spoken assessments the screen is not the work, and removing it takes the sustained upload from roughly 2.56 Mbps to 1.06 Mbps — which materially widens who can complete an assessment on a domestic or mobile connection.

Candidates can appeal on seven grounds, including that a proctoring check flagged them unfairly and that a recording failed. A reviewer has to write reasons, and the candidate reads them.

What to do this quarter, whoever you buy from

1. Retire the unsupervised take-home as a decision instrument. Keep it as a conversation starter if you like it, but stop scoring it.
2. Add a short live walkthrough to every technical process. Fifteen minutes, the candidate's own submission on screen, questions about the decisions. This catches more than any detector and costs you nothing.
3. Make one person accountable for identity at offer stage, separately from whoever runs the assessment. Continuity during a test and identity at hire are different checks.
4. Write down what you will do when a flag fires before one does. A process that has no answer to “the tool flagged them, now what?” will either ignore every flag or wrongly accuse someone.
5. Stop buying detection accuracy claims from the companies selling detection — including the numbers quoted higher up this page.

Frequently asked questions