Home › Voice Detection › For Recruiters

Deepfake job interviews and the AI voice on the call.

To spot a deepfake or AI voice in a job interview, lean on live, unscripted follow-up questions, a camera policy applied to everyone and formal identity checks. A voice detector can review a recorded clip afterwards as one signal, but interview audio is heavily processed and a score must never decide a hire.

The FBI has warned about it twice: voice spoofing in remote-job interviews (June 2022) and face-swapping in video interviews (January 2025). The controls that actually work are procedural rather than technical. Here is an honest account of what to do.

Check an audio clip free What actually works ↓
First check free, no signup Audio not kept on our servers
Known limitation — please read before acting on a result. Genuine recordings can be wrongly flagged. Real human speech, especially compressed audio such as phone calls and messaging-app voice notes, can come back as “AI-generated”, sometimes with a high score. We have not yet published an evaluation of how often this happens. Until we do, treat a synthetic verdict on a recording you believe is real as unproven, and verify the speaker directly — calling back on a number you look up yourself settles it in a way no detector can.
Reality check

Why the detector is the weakest part of your defence.

Four structural problems, none of which is about detection quality.

  1. Conferencing audio is heavily processed. Noise suppression, automatic gain control and aggressive codecs are doing a great deal to the signal before you ever see a file. That processing removes the same irregularities detection depends on, which compresses scores toward the middle for everyone — genuine candidates included.
  2. The error that matters lands on a real person. A missed impostor costs you a bad hire you will discover. A wrongly flagged candidate is a person who loses a job over something they cannot disprove, and who may have a discrimination claim if the flagging correlates with accent, equipment quality or a speech difference. That asymmetry should govern your whole process.
  3. Detection is not a live stream. It scores a recorded file, or a 10, 20 or 30 second clip of a browser tab's audio captured with the TextSight Voice Chrome extension — a web call included. It does not watch the whole call, and by the time you have a result the conversation has usually moved on. Recording a candidate also needs their notice first.
  4. It cannot catch a proxy. If a different real person sits the interview, their voice is a genuine human voice. A voice detector has nothing synthetic to find; only identity checks do.
Never reject a candidate on a detector score. It is not forensic-grade and must never be the sole basis for an employment decision. If a result concerns you, the correct next step is a further live conversation — not a rejection, and not an accusation.
Controls

What actually works instead.

All of these are cheap, none of them depends on a tool, and they apply equally to every candidate.

Live, unscripted conversation

The strongest control there is. Ask follow-up questions that depend on the previous answer, interrupt gently, change direction mid-topic. Real-time voice systems and coached impostors both can struggle with genuine conversational turn-taking, and a listener tends to notice the lag.

Ask about specifics only the real candidate would know

Not credentials — the detail of the work. Why that architecture rather than the obvious alternative, what went wrong on the project, who disagreed with them. Someone relaying answers from elsewhere cannot keep up with a genuine technical follow-up.

Camera on, and consistent

Apply it as a stated policy for all candidates at a defined stage, not selectively when someone raises your suspicion. Selective application is where discrimination risk enters.

Identity verification at the right point

Formal identity checks belong in your onboarding process where they can be applied uniformly and lawfully, not improvised mid-interview because a voice sounded odd. The FBI's January 2025 notice advises identity verification during interviewing and onboarding, and completing as much of the hiring and onboarding process as possible in person.

Structure the process, not the tooling

A live technical exercise, a scheduled call at a fixed time, and consistent interviewers across a panel do more against impersonation than any audio analysis. They also improve hiring quality generally, which is a better reason to adopt them.

Each control, and what it stops

ControlWhat it is meant to stopWhat it does not stop
Live, unscripted follow-upsScripted answers, someone relaying answers from off-screen, voice tools that lag on turn-takingA capable stand-in who genuinely knows the subject
Probing the candidate's own claimed workCoached impostors and borrowed CVsA proxy who did similar work themselves
Camera on, same rule for everyoneAudio-only impersonation; makes sound-and-picture mismatches visible (the 2022 IC3 notice describes coughs and sneezes that do not match the video)Face-swapping, which the 2025 IC3 notice says has been used in video interviews
Same interviewers across stagesA different person turning up for a later roundThe same impostor attending every round
Identity verification at onboarding, in person where possibleApplicants using stolen identities and personal dataOn its own, proving the verified person is the one who interviewed
A voice check on a recorded clip, with noticeNothing on its own; it gives one signal on whether a voice may be syntheticA real human proxy; and processed call audio can be wrongly flagged

FBI sources checked 2 November 2026. The table describes what each control is designed for; we have not measured how often any of them succeeds.

Where a check does help: reviewing a recorded screening call after other signals have already raised a question, as one input into whether to run another live round. That is a legitimate, proportionate use. Making it a routine screen on every candidate is neither.
If you do check

Doing it fairly, if you do it at all.

If your process includes checking recorded interview audio, a few things keep it defensible.

  • Tell candidates. Recording and analysing interview audio has data-protection implications in most jurisdictions, and doing it covertly is the version that creates liability. Notice and a lawful basis are for your own counsel to confirm, not for a guide to assert.
  • Apply it uniformly. Every candidate at a defined stage, or none. Ad-hoc checks triggered by an interviewer's hunch are exactly where bias enters.
  • Use a control. Another recording of the same candidate from the same platform gives you a baseline. The comparison method matters more here than almost anywhere, because platform processing dominates isolated scores.
  • Write down the limits with the result. If the note in your ATS says only “flagged as AI”, it will be read as fact by whoever sees it next. Record the audio conditions and the caveat alongside it.
  • Give the candidate a route to respond. A process with no way to contest a flag is not a process. A short live call resolves nearly every genuine case.
Not forensic-grade. A TextSight voice result is a triage signal, not proof, and it must never be the sole basis for a disciplinary, employment, financial or legal decision. Our full position on what a result is worth is here.
FAQ

Hiring team questions.

How can you tell if a job candidate is using a deepfake or AI voice?
Mostly by how the interview is run, not by a tool. Ask follow-up questions that depend on the previous answer, probe the detail of work the candidate says they did, keep a camera policy that applies to everyone at a stage, and verify identity formally at onboarding. The FBI has reported audio that does not line up with the video, such as a cough or sneeze that does not match what is on screen. A voice detector can review a recorded clip afterwards as one signal, never as proof.
Can I reject a candidate because a voice detector flagged their interview?
No. It is not forensic-grade analysis and must never be the sole basis for an employment decision. Beyond the fairness problem, a rejection resting on an unexplainable score is difficult to defend if challenged. If a result concerns you, run another live conversation.
Why do genuine candidates get flagged?
Conferencing platforms apply noise suppression and automatic gain control that strip out the irregularities detection relies on, which pushes everyone's scores toward the middle. Short clips, heavy accents captured through aggressive processing, low-quality microphones and quiet rooms all add noise to the result. The error lands on a real person, which is why it should never decide anything alone.
What actually stops candidate impersonation?
Live unscripted conversation with genuine follow-up questions, specific technical probing on the candidate's own claimed work, a consistent camera policy applied to everyone at a defined stage, and formal identity verification in onboarding where it can be applied uniformly. Procedural controls beat audio analysis substantially here.
Do I have to tell candidates I am analysing their audio?
Recording and analysing interview audio carries data-protection obligations in most jurisdictions and covert analysis is the version that creates liability. Notice and a lawful basis are questions for your own legal counsel — but as a matter of practice, telling candidates costs nothing and removes most of the risk.
Can I check a live interview in real time?
Not as a live stream. You can check a recording of the call afterwards, or, with the TextSight Voice Chrome extension (free account required), capture a 10, 20 or 30 second clip of the audio playing in a browser tab, a web call included, and check that clip. Tell the candidate before you record any of it: covert recording is the version that creates legal risk. During the interview itself, unscripted conversation is still what works.
Is this useful for high-volume screening?
We would advise against it as a routine screen. The false-positive burden falls on genuine applicants, platform processing makes scores unreliable at exactly the audio quality screening produces, and a systematic flagging process that correlates with accent or equipment quality carries real discrimination risk. Reviewing a specific call after other signals raised a question is the proportionate use.
Has the FBI warned about deepfake job interviews?
Yes. On 28 June 2022 the FBI's Internet Crime Complaint Center (IC3) reported complaints of applicants using voice spoofing, or potentially voice deepfakes, and stolen personal data in online interviews for remote roles. On 23 January 2025 it reported North Korean IT workers using AI and face-swapping technology during video job interviews, and advised identity verification during interviewing and onboarding and as much of the process in person as possible.
Can a voice detector catch a proxy candidate?
No. If a different real person sits the interview, their voice is a genuine human voice and a voice detector has nothing synthetic to find. Proxy interviews are caught by identity checks and by keeping the same person in view across stages, not by audio analysis.

Last verified 2 November 2026.

Related

More voice detection guides.

How accurate is it, really?

Why we publish no single accuracy percentage, what false positives look like, and how to weigh a result.

See our position →

Compare two clips

The control-recording method: run a known-genuine clip alongside the suspect one and read the gap.

Learn the method →

Choosing a detector

The questions to ask any vendor, the claims to distrust, and how to run your own test.

Read the buyer's guide →

Further reading

Have a recorded screening call? Treat it as one signal.

Your first check is free with no signup — 3 a day with a free account. Your audio file is not kept on our servers.

Check an audio clip All voice guides
Never a sole basis for a hiring decision · false positives land on real candidates