Anti-cheat and fraud detection: how Hiearlly keeps AI assessment results trustworthy

Automating assessments is easy. Automating trustworthy assessments is the hard part. The moment a test is unproctored and online, someone will try to game it — and at hiring volume, some of them succeed.

The three ways an assessment gets gamed

• Impersonation. A stronger friend, or a paid proxy, takes the test on the candidate's behalf.
• Copy-paste. The answer comes from GitHub, a past attempt, or another tab — never from the candidate's own reasoning.
• AI assist. An LLM generates a plausible-looking answer in seconds.

A traditional multiple-choice or take-home test detects none of these. It trusts the submission and moves on. That is why so many teams quietly distrust their own screening data.

Detection has to be built into the pipeline, not bolted on

Hiearlly's anti-cheat and fraud engine runs across the whole evaluation — the assessment and the AI video rounds together — rather than as a one-off proctoring add-on. It looks for:

• Identity consistency across the assessment and the video interview, to surface impersonation.
• Edit and interaction telemetry that flags paste and AI-assist patterns.
• Behavioural signals that don't match a genuine, first-hand attempt.

Because the same candidate is assessed and interviewed on video, the two stages cross-check each other. A submission that looks perfect on paper but doesn't match the person who showed up to the interview is exactly the case a keyword filter can never catch.

Why “advisory, not automatic” matters

Automated rejection on a fraud signal is tempting and wrong. Signals are probabilistic; people are not. A nervous candidate on a slow connection can look, briefly, like a suspicious one. Treating the integrity band as advisory keeps a human in the loop for the decisions that actually affect a person's livelihood, while still catching the fraud that manual screening misses entirely.

Trust is what makes automation usable

The whole point of automating screening is to act on the results without re-checking them by hand. That only works if you trust the score. Anti-cheat and fraud detection aren't a feature bolted onto the assessment — they are what makes an automated assessment worth automating in the first place.