Interview Integrity

How our AI interview deepfake detection works

Remote interviews can be gamed: someone could try to sit in for the real candidate, or use a real-time face-swap or voice clone. Our AI interviewer runs a layered set of checks to confirm a genuine human is on camera, in real time, without storing a faceprint or recording the interview. Here is exactly what happens and what a candidate will experience.

Liveness challenges Facial signal analysis Voice naturalness Human review of every flag

Where the check fits into the interview

A liveness check never interrupts a question or an answer. It only appears in the natural gap after the candidate has finished answering and the interviewer has acknowledged the response, just before the next question.

Interviewer asks question 10 Candidate answers Interviewer acknowledges Liveness check (only sometimes) Interviewer asks question 11
The candidate is given a moment to complete the check. Most interviews include one or two of these, placed at unpredictable points so they cannot be anticipated.

Layer 1: Liveness challenges

A short, friendly prompt that a real person can do instantly but a pre-recorded video or a real-time face-swap struggles with. The interviewer reads the first one aloud and shows it on screen; later ones appear on screen.

Left Centre Right We track the rotation of your head, not its position

What you will be asked

The check is one of a small set, chosen at random:

Turn your head left then right Rest your hand on your chin Read 4 digits aloud

The system measures the natural movement (the rotation of your head, the way your hand softens the detected face, or the digits you say matching what was shown). You do not need to be precise; the checks are tuned to accept normal human movement and allow for poor lighting or a brief stumble.

Layer 2: Passive facial signal analysis

This runs quietly in the background, in the candidate's own browser. It looks at signals that a real human face naturally produces and that synthetic faces usually do not. Nothing here is used to identify anyone, and no faceprint is created or stored.

forehead sampled for pulse 1. Heartbeat in skin tone 2. Blink rate open closed 3. Lip-sync timing mouth movement vs audio

The signals

A real face has a faint, regular colour change as the heart beats (this is the same idea behind contactless heart-rate apps), blinks at a human rate, and moves its lips in time with the audio. The system measures these as plain numbers, a pulse estimate, a blink rate, a lip-sync delay, and uses them only to flag something that looks artificial.

Because this layer looks at facial signals from the webcam, it is only switched on when the candidate has given specific consent for it on the consent screen, and a recruiter has enabled it for the role.

Layer 3: Voice naturalness

Cloned or replayed audio often sounds subtly too smooth: little variation in pitch, fewer natural pauses, an unusually even tone. The system scores how natural the speech sounds on a 0 to 100 scale. This is a soft signal, it nudges the overall risk score but never ends an interview on its own.

Natural human speech Unusually even (suspicious)

What happens if a check is not satisfied

  1. The system records the result and the exact reason (for example, no clear head turn detected).
  2. It captures a few still images from the webcam at that moment as evidence.
  3. Depending on the recruiter's settings, the interview either ends politely, counts it as a warning, or simply notes it for review.
  4. The result and evidence are attached to the interview so a human can review them.
A failed check is never the final word. A person always reviews the evidence before any decision is made, and honest candidates who hit a false alarm (for example, from poor lighting) can be re-invited.

What the recruiter sees

On the interview results, recruiters get a clear summary: how many checks passed and failed, a deepfake risk score, a voice naturalness score, the reason for any failure, and the evidence screenshots. Administrators also have a dedicated monitoring view listing every flagged session.

Deepfake / liveness checks Passed: 2 Failed: 1 Risk score: 78 / 100 Voice naturalness: 41 Evidence

Common questions

Is my interview being recorded?

No continuous video is recorded. Only a few still images are captured, and only if a liveness check is not satisfied, as evidence for a human reviewer.

Are you storing my face or building a faceprint?

No. The facial signal analysis produces plain numbers (such as a blink rate) and is never used to identify you. No faceprint or facial template is created or stored.

What if I have poor lighting or a slow connection?

The checks are deliberately forgiving and treat unclear situations as inconclusive rather than failing you. If a genuine candidate is wrongly flagged, a human reviewer can re-invite them.

Do I have to consent to this?

You are shown a consent screen before the interview. The facial signal analysis layer only runs if you give specific consent for it. You can also request a human-conducted interview instead.