A Candidate's Eyes Keep Moving Off-Camera: Cheating or Nerves?

Is looking off-camera evidence of cheating?

No. Eye movement doesn't separate people who are lying from people who aren't. Three studies (coded lies told to camera, observers trained in the pattern, and 52 recordings of people publicly appealing for a missing relative) found no eye-movement difference that tracked deception, and none of the analyses reached significance ¹. A candidate whose eyes keep leaving the frame has told you they are looking at something. Nothing more.

The base rate is the problem. Everyone's eyes leave the frame: reading the question you just typed into the chat, checking the recording light, looking at the second monitor where their notes sit, glancing at a child in the doorway. You are looking away too, taking notes, and on their screen it looks exactly like the thing bothering you.

What you actually noticed is a rhythm: a pause, a shift of the eyes, then an answer that arrives too complete. The last part is the signal. An answer can be checked, and checking it is a question you already know how to ask. Where the eyes went during the pause is not checkable and never will be.

The suspicion also runs one way. Nobody flags the candidate who looks away and then gives a mediocre answer, which means the reading gets applied to answers that were good, and you are reacting to quality you didn't expect. If the real worry is that a model is answering for them, that is a different problem with a different fix.

Why does gaze feel like proof when it isn't?

Because it is one of the most widely held beliefs on earth. Asked how to spot a liar, 63.66% of respondents volunteered gaze aversion before any other cue, and the belief turned up in 51 of the 58 countries surveyed; asked directly, 71.5% agreed ². A belief that universal arrives in the room before the candidate does. It feels like observation, and it is recall.

The specific version (that eye direction reveals the lie) has been tested and did not survive. Participants told lies and truths on camera while their eye movements were coded against the pattern the claim predicts, and nothing reached significance. Observers trained in the pattern did no better than untrained ones. Fifty-two videos of people appealing for the return of a missing relative, roughly half of whom turned out to be responsible for the disappearance, showed no difference either ¹.

Once the suspicion is in the room, it recruits everything else. The pause before the answer becomes retrieval rather than thought. The polished sentence becomes a script rather than preparation. The follow-up gets sharper, so the candidate you have already decided about faces a harder interview than the one before them, and the transcript will show that they did worse.

That asymmetry is the part that leaves a record: different questions, asked differently, of the candidates who were treated as suspects.

What does off-screen work look like in this role?

Ask what the job looks like at a desk. An analyst re-running a number, a consultant pulling up a framework, an engineer reading the docs, a claims examiner opening the policy: all of that is the work, and all of it moves the eyes off the camera. In a role where nobody works from memory, watching a candidate work from memory tests nothing you are hiring for.

So write two lists before the round rather than after it. What may be open on the candidate's screen: their own notes, the job description, the documentation, a calculator, the model they are about to walk you through. What may not: another person, a live chat with someone else, an answer key if you have been unwise enough to have one. Send both lists in the invitation. A stated rule is enforceable; an unstated one is a feeling you get later.

Then ask what the role genuinely requires unaided. Some jobs have real moments of it: a live incident call, a client asking across a table, a shift handoff. If that is the job, build a round that puts the candidate in that moment and score what happens there. Don't approximate it by watching a webcam and hoping the eyes report back.

The general form of that question (what an assistant actually does in the job, and therefore what an interview should let it do) is worth answering for your own roles first.

What is the risk of acting on gaze?

Real, and it lands in two places. Eye behavior varies with disability. Autistic adults describe eye contact as sensory overload, in their own words: "I can't concentrate while making eye contact, particularly if I need to listen" ³. A candidate with low vision may be reading you through the screen magnification the EEOC names as a reasonable accommodation ⁴. Screening on that behavior is screening on the disability.

The EEOC's 2023 guidance on visual disabilities lists screen readers, magnification software and larger high-contrast monitors among the accommodations an employer must provide so an applicant has an equal opportunity to participate in the application process, and states that individuals with disabilities should not be excluded from that process because an employer speculates about accommodation ⁴. A candidate running any of those is not looking at you. That is the accommodation working.

The second exposure is the informal rule. Apply "kept looking away" consistently and it becomes a selection procedure, and under the EEOC's guidance on tests and selection procedures, a procedure that screens out a protected group has to be shown job-related and consistent with business necessity ⁵. Gaze predicts a webcam angle. It is job-related to nothing, and there is no validity study to hand your counsel.

There is a practical version of the same point that needs no statute. You almost never learn why the eyes moved. The candidate with a tremor, the candidate whose partner is asleep in the next room, the candidate reading your question because their hearing aid is not handling your audio, and the candidate with an answer open in a doc are indistinguishable in a video window. Acting on the observation is a coin flip with a rationale attached.

Settle the accommodation question before the round rather than mid-call. Here is what that looks like in an assessment. None of this is legal advice, and the rules vary by jurisdiction; your counsel is the one who reads them.

Test the answer, not the eyes

Replace the inference with a question only this candidate's own answer can survive. Take what they just said and make them operate on it: which number here would change the recommendation, and by how much? A prepared answer has a depth of one. A candidate reading from a script fails the second question whether or not you ever noticed where they were looking.

Five changes do most of the work, and none of them need a tool:

• Spell out the rule in the invitation. One sentence on what may be open and whether a model is allowed. A rule a candidate can read is a rule you can enforce.
• Ask the second question. Not a new topic: one level down on the answer just given. Follow-up questions expose understanding faster than any reading of behavior.
• Write the rubric before the first candidate. Decide what a good answer contains, apply it to everyone, and record why each answer landed where it did. That record survives a challenge; a note reading "seemed to be reading something" does not.
• Give them the assistant on purpose. If the fear is that a model is doing the work, an open-AI round removes the fear and hands you something to read: how the problem got framed, what got delegated, what got refused. The proctoring-versus-open trade is worth making deliberately.
• Keep the observation out of your notes. If it is not in the rubric, it is not a finding. Write what the answer did and did not contain.

None of that requires deciding what the candidate was looking at, which is the point. You were never going to find out, and the answer in front of you was always the better evidence.

See how it works >