Does Mettl Detect Cheating? Flags, Proctors, and the Index

Mercer Mettl does not decide that you cheated. It collects signals from up to four layers, rolls them into a credibility index rated High, Medium, or Low, and hands that to the employer. As of September 9, 2026, the documented layers are browser rules that run on any test, an AI model with eight named flags, an optional live proctor or authorizer, and a code similarity check for coding questions. You are not shown your flags or your index.

Layer one: rules on every test

Three controls apply regardless of proctoring. Mettl's test settings documentation says "By default, the Mettl platform does not allow copy/paste on or from the test window." The anti-cheating measures page lists "Blocking 'copy-paste', 'print-screen' controls." And browsing tolerance, which "counts every time a test-taker navigates away from the test window to give warnings," can terminate the test at a limit the employer sets. The mechanics are in whether Mettl detects tab switching.

Two more are quieter. Employers can restrict the test to "a defined IP range," and a watermark setting "will show a unique ID in the background of each candidate's test screen," which ties any leaked screenshot to the person who took it.

Layer two: the AI flags

When webcam proctoring is on, Mettl's remote proctoring overview says its system "looks for & highlights any suspicious activity using advanced video and audio analytics" and raises eight flags: FACE_NOT_VISIBLE, ANOTHER_PERSON, AWAY_FROM_WINDOW, CANDIDATE_LOOKING_AWAY, CANDIDATE_FACE_PARTIALLY_VISIBLE, CANDIDATE_NOT_PRESENT, CANDIDATE_IMPERSONATION_SUSPECTED, and MOBILE_PHONE_DETECTED.

Mettl's marketing goes further, claiming the model "auto-flags up to 18 digressions with upward of 98% accuracy" (Mettl proctoring FAQ); other Mettl pages say 95%. No audited figure is published. Take the eight documented flags as the real list and the accuracy claim as a vendor statement. Every one of those flags is a webcam observation, and each has an innocent cause: a bad angle, a lamp behind you, a family member, a glance at your keyboard. What the camera side captures is detailed in whether Mettl uses your camera.

Layer three: people

Mettl can put a human in the loop before and during the test. An authorizer reviews your "Live clicked photo and ID card photo" and can reject them. A live proctor, per the Proctoring Guide, watches the webcam and test screen, can "Chat with the candidate," and can "Pause and end the candidate test if needed." In record-and-review, the review happens after the fact against the "entire video recording." Manual proctoring is itself one of the seven inputs to the credibility index, so a proctor's judgment is scored alongside the model's.

Layer four: code similarity

For coding questions, Mettl's test settings documentation describes a plagiarism check that will "take a test-taker's code and compare it with all other test-takers' codes to identify areas which might have a potential overlap." It compares candidates to each other, not to the internet, and it does not analyze typing patterns; Mettl documents no keystroke-level model.

How it becomes one number

Mettl's Candidate Credibility Index documentation says the index "takes inputs from Mercer | Mettl's proctoring flags and test finish status," drawing on seven inputs: five visual flags (Candidate Not Present, Additional Person Present, Away from test window, Mobile phone detected, Candidate's Face is Partially Visible), Manual Proctoring, and Browsing Tolerance. "A client can customize the weightage of the different flags," so the same behavior can score differently at two employers. The result is High, Medium, or Low, shown on the report and in PDF and Excel exports. The page marks the feature as beta.

The product guide's own framing is probabilistic: candidates "with a low credibility index have a higher likelihood of having cheated and vice-versa." It is a likelihood the employer interprets, not a finding.

What you see, and do not see

You see warnings. Browsing tolerance shows one each time you leave the window and can show the count remaining. A live proctor can message you. A terminated test ends on screen. What you never see is the flag list, the index, or the recording. A later request to re-take, or silence, is what a low index looks like from your side. How to handle that conversation is in what happens when a coding assessment flags you.

What this means for candidates

Assume the camera, microphone, and screen are recorded for the whole test and that a model scores every frame with no context for why you looked away. Set up a quiet room, a clean desk, and a single display; keep the test window in front for the entire session; write code in the editor rather than pasting. If the rules prohibit references or AI, prepare with them beforehand and take the test without them. The full inventory of what each tier records is in what Mettl proctoring records and flags.