Interview identity fraud, and the candidate you must not wrong

Two failures are possible here and only one of them is usually discussed. You can hire someone who is not who they appeared to be. You can also flag an honest applicant on weak evidence and quietly cost them a job. A hiring process has to be built against both.

How it presents to a hiring team

Almost never as a dramatic moment. It shows up as small frictions that each have a reasonable explanation.

A screening call where the audio is a little processed and the camera is off, for a good reason. A candidate whose answers are excellent but slightly delayed, as if arriving through something. A voice at the second stage that does not sit quite where the first one did. A different interviewer noticing something the first did not, and neither raising it because neither is sure. Then, after an offer, an early request to change payroll details, or equipment sent to an address that does not match anything on file.

The two shapes to know are simple. In the first, the person interviewing is not the person who intends to do the job — a proxy sits the technical stages for someone less able, or a single operator fronts multiple identities. In the second, the identity itself is fabricated or borrowed, and the objective may be a salary, or access to your systems, or both.

What has changed is that the interview stage no longer filters either one. Real-time voice conversion takes a live speaker and renders their speech in a different voice with a delay small enough to hold a conversation, so “we spoke to them, they were fluent and quick” is not the evidence it used to be.

Process that works without accusing anyone

The best controls here are ones every honest candidate passes without ever knowing they were controls.

• One unscheduled callback, placed by you to the number on the application, at a time not agreed in advance. Brief and ordinary. It is a channel check, not a test of the person.
• Continuity of interviewer. At least one person who is present at more than one stage. Cross-stage inconsistency is caught by people, not by tools.
• Identity through the existing route. Your right-to-work or background check is the place identity is established, done properly, against documents, by the team that does it for everyone. Do not invent a parallel identity check inside the interview.
• One live, unrehearsed stage. A conversation about the candidate’s own claimed work, following their answers wherever they go, is a strong control and a good interview anyway.
• An in-person or supervised final stage where the role justifies it. The most reliable check available, and worth the cost for privileged access.
• Separate the payroll change from the person. Bank detail changes verified by callback on the directory number, with a hold period, never on the strength of an email or a call that came in.

Consent and retention: if calls are recorded or assessed, say so before the call, retain only what you need, and apply the same rule to every candidate for a role. Uneven application is where both the legal and the fairness problems begin. [VERIFY: have your employment counsel confirm the recording, notice and retention rules for each market you hire in.]

The false-positive duty

This part is not optional, and it is the reason to be careful about buying detection for hiring at all.

A detector returns a probability. Applied to interview audio, it is working in poor conditions — compressed conferencing codecs, noise suppression, packet loss, headset processing, short answers — and every one of those pushes real human speech towards looking odd. The people most likely to be affected are not randomly distributed either: someone on a weak connection, someone in a shared or noisy room, someone using assistive technology, a speaker whose accent or speech pattern is less represented in the data. A flag that lands more often on those candidates is a discrimination problem wearing a technical costume, whatever the average accuracy figure says.

And the cost is asymmetric. If you wrongly hire, you find out and you can act. If you wrongly reject, the candidate is told nothing, learns nothing, and carries a loss they cannot see or contest. So write the rule down before you ever look at a score:

• No automatic rejection, ever. A flag is not an outcome. It routes to a named human who reviews it.
• No silent rejection. If a signal is contributing to a decision, the candidate is offered a verification step they can actually pass — a live call, a supervised session, standard identity documents — rather than being dropped without explanation.
• Never the only evidence. A verdict may support a concern that already exists on other grounds. It may not create one on its own.
• Same rule for everyone in the role. No selective checking of particular candidates, which is where bias enters even with good intentions.
• Log the decision. What was flagged, who reviewed it, what the candidate was offered, what was decided. If you cannot explain it later, do not do it now.

We would rather say this plainly than sell into it: on conferencing audio a “likely human” result is weak evidence and a flag is a reason to look, not a reason to decide. See phone and call audio and where we fail before designing anything around a score.

If you believe it has happened

• Treat it as an access problem first. If the person was onboarded, credentials, devices and data access matter more urgently than the employment question.
• Preserve the recordings in their original files, along with the application, correspondence and payroll details. Do not forward audio through chat apps before it is assessed.
• Involve HR and counsel before confronting anyone. Getting this wrong about a real employee is its own serious harm.
• Report it. [VERIFY: link the national fraud reporting bodies and, where relevant, sanctions or employment-verification authorities for your markets — verify every URL before publishing]

Questions