What Do You Do When the Interviewee Isn't Who Did the Take-Home?

Why won't the video feed settle it?

Because people are bad at both halves of that judgment. Asked to sort authentic videos from deepfakes, participants averaged 57.6% accuracy, barely above chance, with confidence running well ahead of accuracy, and neither an awareness warning nor a cash incentive moved the number ². Face matching does no better: passport officers comparing photo ID against the live person holding it wrongly accepted 14% of fraudulent photos ³.

Experience does not rescue it either. In that same study of passport staff there was no relationship between employment duration and face-matching accuracy, and the officers scored no better than the general population on a standardised matching test ³. You are doing the harder version of their task, on a compressed call, against a headshot and a memory of a phone screen.

The published tells are real, and they are still not diagnostic. In 2022 the FBI's Internet Crime Complaint Center reported complaints of deepfakes and stolen personal information used to apply for a variety of remote and work-at-home positions, some of them information technology, programming, database and software work, and named what interviewers had noticed: the lip movement of the person on camera not completely coordinating with the audio, and coughing or sneezing landing out of alignment with the picture ¹. Every one of those is also what a saturated home connection looks like at 4pm.

So finish the interview the way you would have finished it, and keep the theory to yourself until the panel has scored independently. Said out loud, it recruits three more people into looking for evidence for it, and none of them can unhear it. Work through the dull explanations first, because one of them is usually the answer: nerves, a coached first round, a script on a second monitor, one interviewer who ran the screen warmly and another who ran the final cold. The gap between a strong remote screen and a flat onsite has ordinary causes long before it has sinister ones, and a candidate's eyes going off-camera is the same trap in miniature.

Which questions can a stand-in not answer?

The ones nobody saw coming. A stand-in can be handed the submission and its conclusions. Handing over the discarded options, the dead ends and the shape of the afternoon is a different job. In a deception study where pairs were interviewed separately about a shared restaurant meal, up to 80% of liars and truth tellers were correctly classified, particularly on the drawings, and the effect ran on the questions they had not anticipated ⁴.

That research is about alibis rather than hiring, and the part that transfers is the mechanism rather than the number: rehearsal covers what a person can see coming. The opening questions everyone expects separated nobody in that study. The unpredictable spatial ones did.

So book thirty minutes and make the whole round a walkthrough of their own submission. Say so in the invitation, because someone who did the work reads that as normal and an ambush only makes a nervous candidate look guilty. Then ask for the things a finished artifact does not carry:

• Software engineering. Which test did you write first, and what did it catch? What broke when you got the repo running? Which function did you nearly extract, and why did you leave it alone?
• Financial analysis. Which figure did you rebuild by hand because the packet's number wouldn't tie? What would the growth rate have to be for the recommendation to flip?
• Marketing. Which statistic did you go looking for the original of, and what did it turn out to say? What came out of the brief after you opened the source?
• Data and analytics. Which column did you drop, and what did the result look like before you dropped it? What did the assistant assert that the data would not carry?
• Legal operations. Where did the playbook and the signed precedent disagree, and which one did you follow? What did you leave for someone senior, and why?

One more move fits in ten minutes on the call: hand them something small that their own submission already implies. Add a segment to the analysis. Add a case to the test. Whoever built the original moves around inside it; whoever was handed it searches it. That is the same mechanism behind working out who actually built a portfolio project, and the same three rungs that make follow-up questions expose real understanding do the work here: specify, invert, falsify. Score it on a written rubric, ask the same set of every finalist, and record the answers rather than your impression of them. See how Olive measures this.

Is asking for ID the safe move?

Vendor advice reaches for it first, and the answer is still no. Federal law puts identity and work-authorization checking at hiring: 8 U.S.C. §1324a(b) makes the person hiring, recruiting or referring attest that they examined documents the statute itself lists ⁵. A final-round interview is not that step, and a document demand aimed at one candidate is a process only that candidate got.

The neighbouring provision is the one worth reading before improvising. In the United States, 8 U.S.C. §1324b, added by the Immigration Reform and Control Act in November 1986 and amended through 1996, makes it an unfair immigration-related employment practice to discriminate in hiring because of national origin or citizenship status, and it treats a request for more or different documents than §1324a(b) requires, or a refusal to honor documents that on their face reasonably appear to be genuine, as that same violation where it is done for the purpose or with the intent of discriminating ⁶.

An off-book ID demand in a video call is not literally that provision, because it is not part of the verification form at all. What it is instead is a step you applied to one person and not the rest, decided after you looked at their face and listened to their voice, and intent is the element that gets argued from exactly that pattern. This is public law rather than legal advice, jurisdictions differ, and the policy goes to counsel before it goes into a process document.

If you want identity verified, the fix is structural and dull. Name the stage, which is after the offer is accepted. Name the method your onboarding already uses, apply it to every hire in the same order, and publish it where a candidate can read it before they apply. Write that rule while nothing is at stake. A step invented in the twenty minutes after a video call will not survive the first person who asks why they got it and the last hire did not.

Don't let a hunch about AI use become an identity accusation

They are two problems, and most of the guides on this sell them as one. A candidate who used an assistant on the take-home is the candidate, using a tool. Someone else doing the work is a fraud question. Reading the first as evidence of the second is how an honest applicant ends up suspected of something close to a crime because their prose read clean.

The instrument people reach for there does not work. Seven detectors run over 91 TOEFL essays written by non-native English speakers produced a 61.3% average false-positive rate: 97.8% of those human-written essays were flagged as AI-generated by at least one detector and 19.8% by all seven, while the same tools classified US eighth-graders' essays accurately ⁷. The error is not spread evenly. It falls on people writing in a second language, which is national origin wearing a percentage.

Notice where that leaves the two suspicions stacked. The candidate whose written English reads as slightly formal, whose home connection is poor, and whose take-home came back unusually clean is the one most likely to attract both doubts at once and the least likely to have earned either. Keeping hiring managers from rejecting people for sounding like AI is the same discipline as this, one step earlier in the funnel.

If the submission genuinely reads as assisted, that is its own question with its own answer, and the answer turns on whether the brief said no before they started rather than on how the finished work reads afterward. Keep the two files apart. A note that says "felt AI-written" sitting next to a note that says "possibly not the same person" is one observation being counted twice.

What if the walkthrough confirms your doubt?

Then you have a hiring decision rather than a fraud case, and you can make it without alleging anything. Decline on what the round showed: work submitted under their name that the person on the call could not account for. Keep the questions and the answers in the file, keep the theory out of it, and hold the finalist you liked to the same bar.

What you say to the candidate matters more than what you privately concluded. Give them the reason you can support, which is the walkthrough and what it did not establish, and nothing you would not want read back to you in six months. Explaining an AI-related rejection to a candidate runs on the same rule: name the observable, never the theory.

There is one case where this stops being a hiring matter. If the role reaches customer data or production systems and the pattern matches the FBI advisory, meaning stolen personal information behind a remote technical application, it goes to your security and legal teams, and the complaint goes to IC3, which is where those reports are asked for ¹. It does not go to the candidate in an email written at 6pm.

Then change the round so the next one is not decided by a hunch. Two edits carry most of it. Put a live component with the same rubric in front of every finalist, so an unwatched artifact is never the only observation you hold: the take-home and the live session answer different questions and the cheap version is both. Then ask the take-home for the decisions rather than the deliverable, which means what was checked, against what, and what changed as a result.

Say the limit out loud. None of this makes identity certain, and no interview does. It makes the round cheap for the candidate who did the work and expensive for anyone who did not, which on a Tuesday, holding one suspicion and no evidence, is the whole of what you can fairly do.

See how it works >