Candidate fraud and deepfake interviews: what works without biometrics

Candidate fraud once meant a padded résumé. It now includes a stranger taking the interview in someone else’s name, a face-swapped video feed, and stolen identities applying for remote jobs on behalf of a sanctioned state. The FBI, the Justice Department and the employers on the receiving end have documented every one of these. The tempting response is to turn the first interview into a checkpoint: scan the face, analyze the voice, score whether the candidate is “live.” This post argues for a first line of defense built on better questions, and it spells out where that defense stops.

What candidate fraud looks like in 2026

Candidate fraud is any attempt to get hired as someone you are not, or on the strength of experience you do not have. Gartner’s working definition is narrower and useful: a would-be hire “that pretends to be someone else and/or has someone else complete an interview pretending to be them” (Computerworld). In practice it takes four shapes:

• Proxy interviewing: someone else takes the interview.
• Deepfake candidates: real-time face or voice manipulation on a video call.
• Stolen or synthetic identities: a real person’s name, someone else behind it.
• Organized schemes that combine all three. The most visible are North Korean IT workers drawing salaries from Western companies.

The public record, in order:

• June 28, 2022. The FBI warned that deepfakes and stolen personal data were being used to apply for remote tech jobs, in interviews where “the actions and lip movement of the person seen interviewed on-camera do not completely coordinate with the audio” (FBI, I-062822-PSA).
• January 23, 2025. The FBI said North Korean IT workers “have been observed using artificial intelligence and face-swapping technology during video job interviews” (FBI, I-012325-PSA).
• April 21, 2025. Palo Alto Networks’ Unit 42 reported that a researcher with no prior experience built a real-time deepfake in just over an hour, on a five-year-old computer (Unit 42).
• June 30, 2025. The Justice Department announced searches of 29 known or suspected “laptop farms” across 16 states. In one indictment, prosecutors said the scheme compromised the identities of more than 80 Americans to land remote jobs at more than 100 U.S. companies (DOJ).
• July 31, 2025. In a Gartner survey of 3,000 job candidates, 6% admitted to interview fraud. Gartner also forecasts that by 2028, one in four candidate profiles worldwide will be fake. Treat that number as a projection (HR Dive).
• November 2025. In a Greenhouse survey of 665 U.S. recruiters and hiring managers, 18% said they had dealt with deepfakes (HR Dive).
• December 2025. Amazon’s chief security officer said the company had blocked more than 1,800 suspected North Korean applicants since April 2024 (The Register).
• October 13, 2026. The FBI warned that these operations increasingly pay Americans to front for them: to apply under their own names, and to “appear on camera, participate in job interviews or meetings” (FBI, I-101326-PSA).

For hiring teams, that last entry matters more than the deepfakes. The next section explains why.

Biometric checks are the wrong first line of defense

When the threat is a fake face, the instinct is to inspect faces: liveness checks, face matching against an ID, voice analysis, an “integrity” score. For a first-round interview, the one stage that touches every applicant, that instinct fails in four ways.

It is an arms race you start behind. The tells that expose today’s deepfakes (a hand passed over the face, lips that lag the audio) belong to today’s tools. If a newcomer can build a working one in just over an hour, detection is chasing a target that keeps improving.

It misses the fraud that is growing. A paid stand-in has a real face and a real voice. Liveness detection confirms that a live human is on the call; it says nothing about whether that human will do the job. And when the person on camera is the identity holder renting out their own name, as the FBI’s October 2026 notice describes, a face match passes because the face matches.

It imports bias. Face analysis does not perform equally across people. In NIST’s December 2019 evaluation of 189 algorithms from 99 developers, “the majority of face recognition algorithms exhibit demographic differentials,” with false-positive rates for Asian and African American faces often 10 to 100 times higher than for Caucasian faces in one-to-one matching (NIST). One-to-one matching is exactly what an ID-to-video check does. Every false flag is a real candidate treated as a suspect. Even the FBI cautions that blunt filters can be unreliable, block qualified applicants and harm the very people whose identities were stolen, and it asks employers to calibrate their controls against false positives (FBI, I-101326-PSA).

It imports legal risk. In the EU, biometric data used to uniquely identify a person is a special category under GDPR Article 9: processing it is prohibited unless a narrow exception, such as explicit consent, applies (GDPR Art. 9). Tools that go further and claim to read stress or nervousness from a voice or face are inferring emotions. The EU AI Act’s prohibition on emotion recognition in the workplace has applied since February 2, 2025 (European Commission), and the Commission’s guidelines make clear that “hiring processes also fall within the workplace context” (FPF). We mapped what those rules require in AI interviews and the law.

A first round that treats every candidate as a suspect spends its error budget on the honest ones.

Content is harder to fake than a face

Real-time AI can change what a candidate looks and sounds like. Real-time help from a model, a coach or a handler in the next room can hand them a polished first answer. Neither is good at supplying a specific, coherent history that holds up two follow-ups deep and stays consistent across a conversation, and that asymmetry is the one to build on. The same logic answers the far more common case of real candidates leaning on AI mid-interview; we cover that in candidates are using AI in interviews.

The security guidance points the same way. The FBI’s January 2025 advice includes using “soft” interview questions that ask applicants for specific details about their location or education (FBI, I-012325-PSA). When Amazon described how it screens out fraudulent applicants, it listed structured interviews alongside background checks and credential verification (The Register).

What to listen for:

• Specificity. One real instance, with its constraint, its people, its numbers.
• Depth under follow-up. A script covers the headline. The second follow-up (“what did you drop, and what did that cost you?”) goes where nobody prepared.
• Ownership and failure. What they did, distinct from the team, and what went wrong.
• Consistency. Come back to something they said ten minutes earlier: the team size, the stack, the timeline. Lived memory tends to hold at the level of detail that matters. A borrowed story tends to drift, or to stay vague exactly where it should get sharper.

One caution: an inconsistency proves nothing on its own. Nervous people misspeak, and everyone simplifies. Treat a contradiction as a question for a person to raise later.

This also means you rarely need to catch anyone. When every rating must cite a specific moment against a versioned rubric, a candidate who cannot produce specifics does not score well, whether they are a fraud, a stand-in or just not ready. Thin evidence shows up as thin evidence; see evidence-based scorecards.

A round-one playbook against candidate fraud

1. Ask for one instance, then follow up twice, using techniques fixed per competency in advance (a specific example, ownership, the failure case), so every candidate faces the same pressure.
2. Return to earlier threads. Record contradictions as quotes, in the candidate’s own words, for a human to raise later.
3. Score the evidence. Every rating cites a transcript moment. Polish, accent and confidence get no vote.
4. Flag anomalies for a person. Someone who can read the evidence and ask decides what they mean. Nothing is rejected automatically.
5. Verify identity once, at the human stage. Before an offer or before system access, the FBI points to E-Verify, a passport or multiple forms of identification, or live verification (FBI, 2026), and to doing as much of hiring and onboarding in person as possible (FBI, 2025). Do it once, for a stated purpose, owned by a person, and only for finalists, so it never turns into ambient surveillance of every applicant.
6. Check continuity between stages. In the live round, pick up a thread from round one: “you mentioned the migration that slipped. Walk me through the rollback.” The person who took the first interview knows what they said.
7. Keep basic application hygiene. The FBI recommends cross-checking HR systems for applicants who share résumé content or contact details (FBI, I-012325-PSA).

What content checks cannot do

The approach has three real limits.

• A capable stand-in can pass. If the proxy genuinely has the experience, their answers will be specific and consistent. Content shows that the answers came from someone who did the work. It cannot show that the someone is the applicant.
• The switch can happen after the interview. The FBI notes that “sometimes an individual is employed to pass the initial interview, but the on-the-job work is completed by a different individual” (FBI, I-072325-4-PSA). Only onboarding and access controls can catch that.
• No single layer is enough. The FBI’s October 2026 guidance says these workers “cannot be identified reliably with a single indicator,” and recommends layered detection and continuous identity verification throughout employment (FBI, I-101326-PSA). A well-designed interview is one of those layers.

Where Hure lands

Hure is built for the content layer and only that layer. Candidates take an 8–10 minute voice-to-voice interview in their own language, from a signed, single-use link tied to one interview. The link identifies the interview; it does not identify who is speaking. The interviewer follows up using only the techniques the hiring team approved for each competency: a specific example, the candidate’s own role, a failure case. It returns to threads the candidate opened earlier and notices when two things they said do not line up, quoting them verbatim. Every score is anchored to a sealed, versioned rubric and cites the moment in the transcript that earned it.

Hure’s limits are just as deliberate. It does not detect deepfakes, verify identity, or analyze faces or voices. It uses no voiceprints, no liveness scoring and no emotion inference. Those signals would bring bias and legal risk into the stage that touches every applicant, and they would not catch a real person sitting in for someone else anyway. Identity belongs at the human stage, and so does the decision: a person makes the final call, with the transcript evidence in hand. See how the interviewer works.