Deepfake interview detection comes down to making the face do something the software can't track: ask the candidate to pass a hand slowly in front of their face, turn to a full profile, and change the light. Real-time face swaps break on all three. That advice comes from Palo Alto Networks' Unit 42, whose researcher built a working interview deepfake in "just over an hour" on a consumer graphics card. If it takes an hour to build, assume some of your applicants have one.
Key Takeaways
- The FBI has warned since 2022 that deepfakes are used to apply for remote software jobs, and named lip-sync lag as a tell.
- Unit 42 calls a hand passed over the face the most effective live test; fast head turns and lighting changes also produce visible artifacts.
- Gartner predicts that by 2028, one in four candidate profiles worldwide will be fake. Treat that as a forecast, not a measurement.
- Detection in the interview is one layer. Identity checks before the offer and at onboarding catch what the call misses.
How common are deepfake interviews?
Common enough that the numbers keep coming from different directions. The FBI's June 2022 PSA reported complaints about deepfakes and stolen personal data used to apply for "information technology and computer programming, database, and software related job functions." Its January 2025 PSA said North Korean IT workers "have been observed using artificial intelligence and face-swapping technology during video job interviews."
Surveys point the same way. Greenhouse's 2025 AI in Hiring Report found that 65% of US hiring managers have caught applicants using AI deceptively, including "showing up as deepfakes (18%)." In Gartner's own 2025 candidate survey, 6% of 3,000 job candidates admitted to interview fraud (posing as someone else or having someone pose as them). That same release carries the prediction everyone quotes: "by 2028, one in four candidate profiles worldwide will be fake."
The money is moving too. In August 2026 Deel announced it is acquiring Clarity, a company focused on identity verification and deepfake detection. Global hiring platforms don't buy deepfake detection for fun.
The deepfake interview detection checklist
Run this on every final-round video call. It takes about two minutes and honest candidates rarely notice.
| Check | How to do it | What a deepfake does | Source |
|---|---|---|---|
| Hand over face | "Can you wave your hand slowly in front of your face?" | Face tracking breaks; the mask flickers or slides | Unit 42 |
| Profile turn | "One more quick camera check: turn to the side for a second." | Artifacts at the jawline and ears on fast turns | Unit 42 |
| Lighting change | Ask them to move closer to a window or switch a lamp | Face lighting doesn't match the room | Unit 42 |
| Lip sync | Watch the mouth on plosives (p, b, m) | Speech and lips drift apart | FBI 2022; Unit 42 |
| Cough or laugh | Make them laugh. Notice sneezes and coughs | Sound doesn't match the visual | FBI 2022 |
| Unscripted follow-up | "Why that design? What broke first?" | Latency while someone else answers | Sep 2026 IC3 alert |
| Eyes and freezes | Note glances off-screen, repeated freezes | Reading from another monitor; masking glitches | Sep 2026 IC3 alert |
Unit 42's HR guidance also says cameras on from the first call, and record interviews with consent. A recording lets a second person review a call that felt off, which beats arguing from memory. Recording rules differ by state and country, and some require every party's consent; this is general information, not legal advice, so check with counsel before you record.
Questions beat filters
The tests above catch the face. The technical conversation catches the person behind it. A deepfake setup often has a second human feeding answers, or a model doing it, and both struggle with depth on the candidate's own history.
Pick one project from their resume and go four levels down. What did you build. Why that approach. What broke in production. What would you change now. Real engineers get more specific with each question; a feed gets vaguer, slower, or switches topic. This is the same structured approach we recommend in structured vs unstructured interviews, and it doubles as a fraud check for free.
Small talk works too. The September 2026 alert suggests asking about "hometown, weather, or hobbies." It sounds silly. It's hard to fake in real time if you're not where you say you are.
A Concrete Version
A 15-person fintech is hiring a senior React Native engineer. Final round, 60 minutes, two interviewers.
- Minutes 0-5: camera check. One interviewer says, "Before we start, can you move your hand across your face once? We do this with everyone now." Takes ten seconds. Nothing happens. Good.
- Minutes 5-40: a walkthrough of the candidate's last app. Four-deep follow-ups on one offline-sync bug.
- Minutes 40-45: "Last camera check, same as everyone: can you turn toward your window for a second?" The candidate does it, laughs, complains about the Bogota rain.
- Minutes 45-60: questions from the candidate.
Cost: about five minutes of a one-hour call. In a different version of this call, the hand passes and the jawline shimmers, the follow-up answers come after a three-second pause every time, and the candidate's camera "breaks" when asked to turn. Three flags. You end the process politely and log what you saw.
Saying "we do this with everyone" matters. It turns the check into procedure instead of an accusation.
One note on our side of this. Sol, our hiring agent, shows blind profiles and a vetted shortlist; it does not check faces, so the camera checks above stay your job.
The Honest Counterpoint
Two problems with everything above.
First, detection is an arms race and the tests age. Hand-over-face works against today's consumer tools because face tracking loses landmarks. Better models will fix that. Any checklist you write in 2026 needs a review in 2027.
Second, false positives are real. Bad bandwidth causes freezes and lip-sync drift on honest calls, and a lot of strong engineers work on connections you'd hate. A single glitch proves nothing. Look for clusters of flags plus weak answers on depth.
The bigger point: interview detection is the wrong place to rely on alone. Gartner's own research note said only 7% of recruiting leaders used identity document validation software as of its 2024 survey. Matching an ID document and a liveness check to the person on your call catches most of what a webcam test misses. See candidate identity verification for how that works.
For transparency: Ruzora's vetting is an AI-led technical interview and a graded coding assessment, plus references and a founder screen. We don't run deepfake detection or document checks today; proof-of-human interviews are on our roadmap. Run the checklist above on anyone you interview, including engineers we send.
Frequently Asked Questions
Can you detect a deepfake in a live job interview?
Often, yes, with simple tests. Unit 42 found that passing a hand over the face disrupts facial landmark tracking, and that fast head movements and lighting changes produce visible artifacts. None of these tests is perfect, so combine them with depth questions and an identity check.
What does the FBI say about deepfake job candidates?
The FBI's June 2022 PSA described deepfakes used to apply for remote tech jobs and pointed to lip movements that "do not completely coordinate with the audio." Its January 2025 PSA said North Korean IT workers used AI and face-swapping in video interviews.
Should I tell candidates I'm checking for deepfakes?
Tell them you run the same short camera checks with everyone. Honest candidates don't mind, and it makes the test fair and consistent. If you record interviews, get consent first.
The Bottom Line
Deepfake interview detection is a two-minute habit: hand over the face, a profile turn, a light change, and four-deep questions on real past work. Make it routine, count flags rather than reacting to one glitch, and back it with an identity check before the offer.
Want candidates who already passed a coding assessment and an AI-led technical interview before you meet them? Request a shortlist.
Roberto Espinoza is CEO of Ruzora, which helps US startups hire pre-vetted senior LATAM engineers, with a vetted shortlist in 72 hours. See available engineers.
