How to Prevent Cheating in Online Hiring Assessments

Cheating in online hiring assessments used to mean a friend off-camera. Now it means AI tools answering in a second tab, deep-fake video, and increasingly sophisticated impersonation. The defence has to evolve at the same speed. Single-layer proctoring catches the obvious cases; layered defence with human review catches the rest. Auto-disqualification based on AI flags is dangerous in both directions — it rejects good candidates and creates legal risk.

The cheating threat model in 2026

The cheating patterns have shifted meaningfully over the last few years. The platform's threat model needs to keep up.

• AI tool usage. The candidate pastes the question into ChatGPT, Claude, or a coding-specific tool, then types the response back. Detection requires multiple signals — tab switches, typing patterns, abrupt high-quality answers.
• Second-screen consultation. A candidate with a second monitor or tablet reads the question, looks up references off-camera. Face tracking catches some cases; audio anomaly detection catches more.
• Off-camera helper. Someone else in the room or on a call providing answers. Audio anomaly detection is the primary signal.
• Identity substitution. A different person takes the test on the registered candidate's behalf. Facial recognition matched to ID at session start catches this.
• Pre-leaked questions. The candidate has seen the questions before, through a public bank or a friend who recently took the same test. Randomised question delivery and private custom banks reduce this.
• Deep-fake video. Emerging, rare so far. Detection requires liveness verification — randomised prompts the candidate must respond to in real time.

No single signal catches all of these. Layered defence with human review is the only architecture that handles the full threat model.

Layered defence works, single defences fail

Each layer catches some attempts and misses others. The layers together catch the great majority.

Identity layer. Government-issued ID matched against the candidate's face at session start. Catches identity substitution before the session begins.

Face and gaze layer. Continuous monitoring of where the candidate is looking. Catches off-camera glances and brief disappearances.

Audio anomaly layer. Background audio analysis — multiple voices, whispers, structured spoken responses. Catches off-camera helpers.

Browser layer. Tab switches, full-screen monitoring, keyboard shortcuts that would open external tools. Catches AI tool consultation.

Typing pattern layer. Distinctive patterns when a candidate reads vs types from memory vs pastes from elsewhere. Catches AI-generated answers.

Question delivery layer. Server-side randomisation, no bulk download, time-controlled reveal. Prevents pre-leakage and group cheating.

Each layer alone has false positives and false negatives. Together, they are robust.

The AI tool problem specifically

The fastest-growing cheating pattern is AI tool usage. Detection is hard because the tools are designed to mimic human output.

Realistic defences:

• Browser lock or kiosk mode for high-stakes assessments. Prevents tab switching entirely. Heavy on candidate experience but eliminates the vector.
• Distinctive typing pattern detection. A candidate who pastes large blocks of text mid-task generates different signals from one who types from memory.
• Task design that AI struggles with. Role-specific tasks with context unique to your codebase or workflow. Generic AI tools cannot replicate them.
• Practical phase scoring that catches AI output. AI-generated code often has tell-tale patterns — over-commented, unnecessarily abstracted, missing the specific style of the codebase.

If the candidate can be prevented from accessing AI tools (browser lock), do it. If not, design tasks that resist AI assistance and detect the attempts.

What humans catch that AI misses

The AI proctoring system generates flags. A human proctor interprets them. The interpretation step is where most false positives and false negatives are caught.

Examples of human review catches:

• A candidate glanced away three times in 90 minutes. AI flags this. Human reviews the session, sees the candidate was looking at a notepad on the desk, dismisses the flag.
• A candidate's audio picked up a child's voice briefly. AI flags as second-person assistance. Human reviews, notes the child was clearly not coaching answers, dismisses.
• A candidate's answers came in unusually fast and matched ChatGPT's voice exactly. AI does not flag this because each individual signal is below threshold. Human reviews the pattern, escalates for verification.
• A candidate's typing pattern suggested copy-paste mid-answer. AI flags. Human reviews and sees the candidate was reformatting code with auto-completion, dismisses.

The human review step is what makes online proctoring trustworthy. Without it, every assessment is a black box of AI judgement that nobody — candidate, hiring manager, legal team — fully trusts.

Why no platform should auto-disqualify

Some platforms market "instant decision" or "auto-disqualify on integrity flags." Avoid them. The reasoning:

False positives wreck candidate experience. A good candidate rejected for a flag they were innocent of is a brand-damaging incident. Multiply by months of operation and the brand cost is real.

False positives create legal risk. Auto-disqualification on AI judgement, without human review, increasingly runs into employment law issues. Several jurisdictions are formalising requirements for human review of consequential automated decisions.

The integrity signal is information, not verdict. A flag is one input into a hiring decision. The decision belongs to the human reviewer who can interpret the flag in context.

AssessExpert never auto-publishes a report or auto-disqualifies on flags. A certified proctor reviews every session, writes the integrity note in plain language, and signs off before the report reaches the hiring team.

The honest cheating rate

Studies of unproctored online assessments suggest cheating rates of 15-30%. With layered proctoring plus human review, the rate drops below 3%. With browser-lock plus layered proctoring plus human review, the rate drops below 1%.

Some level of cheating is unavoidable — no defence is perfect. The question is whether the residual rate is low enough to make the assessment data trustworthy. Below 3% is generally considered acceptable for hiring decisions; below 1% is the standard for high-stakes regulated assessments.

What to tell candidates

Be transparent. The invitation email should explain that the session is proctored, what is monitored, that a human reviews the data, and what happens if flags are raised. Most candidates accept this when explained; the resentment comes from surprise.

Avoid surveillance language that overstates the monitoring. "Your session is recorded for integrity purposes and reviewed by a certified proctor" is honest and reasonable. "We constantly monitor your face, eye movement, audio, and screen for any sign of cheating" is true but reads as hostile.

How AssessExpert handles integrity

Layered proctoring covers identity, face, gaze, audio, browser, and typing patterns. Question delivery is randomised and server-side. Practical scoring catches AI-generated submissions. Every flagged session is reviewed by a certified human proctor before the report publishes. The integrity note in plain language goes into every report. No auto-disqualification. For the platform overview, see Online Assessment Platform. For the report structure that includes the integrity note, see Candidate Reports and Scoring.

FAQ

What is the realistic cheating rate without proctoring?

Studies suggest 15-30% in unproctored sessions. With layered proctoring plus human review, it drops below 3%.

Should we block AI tools entirely?

If the role does not permit AI tools, yes — browser lock or kiosk mode. If the role does, design tasks that test AI-augmented work skill, not AI absence.

What happens to a candidate whose session is flagged?

The proctor reviews. Most flags are dismissed as innocent. Severe and clear cases result in decline. Borderline cases trigger a clarifying interview before final decision.

Can candidates dispute flags?

Yes. The session recording is retained for the legal minimum, and candidates can request review of disputed decisions.

How do we know the platform's proctors are calibrated?

Proctor calibration sessions and inter-rater reliability checks. Ask the vendor to show their calibration data.

What about deep-fake video?

Rare so far but increasing. Liveness verification — randomised prompts the candidate responds to in real time — is the emerging defence. Most platforms are adding it.

Next steps

If you want to see the proctoring layer end-to-end and review a sample integrity note, book a demo. The first call covers your integrity standards and the threat model relevant to your hiring.