How to Tell If You're Interviewing a Deepfake

Written by Liza Carey | Sep 22, 2026, 9:08:51 PM


Here is the number that should change how you run video interviews.

A 2025 meta-analysis of 56 separate studies found that humans identify deepfakes correctly about 55% of the time. That is barely better than guessing. And it holds across the board, including for people who work with video every day and believe they have a good eye for it.

Now put that next to what recruiters are actually reporting. In a Greenhouse survey of more than 4,000 people, 31% said they had personally interviewed someone they suspected or later confirmed was using deepfake technology. Nine in ten had encountered or suspected AI-generated answers during an online meeting.

So the problem is not that fake candidates are rare. It is that the primary control most firms rely on, a trained person watching a screen and trusting their gut, performs at roughly the level of a coin flip.

That is worth sitting with, because nearly every article published on this subject tells you to watch for red flags. Unnatural blinking. Odd lighting around the hairline. Audio that lags the mouth. All real, all worth knowing, and none of it a control you can build a hiring process on when the underlying detection rate is 55%.

The better approach is not looking harder. It is asking differently.

Why real-time questions work when observation doesn't

Both deepfake systems and coached impostors share a weakness: they are optimized for the expected. A prepared answer to a predictable question is where they perform best. Real-time synthesis, physical interaction, and unscripted redirection are where they degrade.

You do not need to detect the fake. You need to create a moment the fake cannot handle.

Seven questions that hold up

1. Interrupt mid-answer with a follow-up that depends on what they just said. Scripted responses do not survive redirection. If someone is reading or being fed answers, the seam shows when they have to abandon a prepared block and respond to something specific from three seconds ago.

2. Ask about something visible behind them, then ask them to interact with it. Not "tell me about your setup." Ask them to pick up the object on the shelf and turn it around. Physical interaction with the environment is difficult for synthesis pipelines and awkward for someone using a virtual background to hide a room full of prompts.

3. Ask them to sketch something and hold it up. Then ask a question about a specific part of the drawing. This combines physical action, occlusion of the face, and content that has to be generated on the spot.

4. Request a specific movement. Turn fully in profile. Pass a hand slowly across the face. Stand and step back from the camera. Many face-swap systems fail on extreme angles and on occlusion, where a hand crosses in front of the face. This feels odd to ask, which is why it is worth scripting into the interview so it lands as routine rather than accusatory.

5. Ask about a failure. Fabricated career narratives are optimized for achievement. They tend to be thin and generic on the specifics of a project that went wrong, because failure stories require detail that is hard to invent convincingly under time pressure.

6. Return to a detail from three questions earlier. Impostors working from notes lose continuity. A real candidate discussing their own experience can pick a thread back up without effort. Someone reading a prepared profile often cannot.

7. Get a detail wrong on purpose. Reference a technology or a company fact incorrectly and see whether they correct you. Someone who actually holds the expertise will usually push back. Someone performing it will frequently agree.

The part most guidance leaves out

These seven questions help. They are not sufficient on their own, and the reason is structural.

Video is one signal. If your entire verification strategy lives inside the interview, you are betting everything on a single channel that is specifically the channel being attacked. The firms handling this well have moved verification off video entirely, and moved it earlier.

Two changes matter more than anything you can do on camera.

  • Verify identity before the assessment stage, not at offer. Most processes verify at the end. If the identity check comes after the technical evaluation, and it fails, you have just discovered that your entire assessment was of the wrong person. Everything you learned is void.

  • Include at least one live, synchronous touchpoint. Fully asynchronous processes, where a candidate records answers on their own schedule, carry meaningfully higher exposure. Recorded submissions remove every real-time pressure the seven questions above depend on.

There is also a counterintuitive finding worth knowing. Gartner research found that 62% of candidates said they were more likely to apply to a role when the organization required in-person interviews. The assumption that tightening verification costs you pipeline appears to be backwards, at least for some roles.

What to do when you suspect something

This is where most firms improvise, and improvising is expensive.

Do not confront in the moment. An accusation in a live interview creates legal and reputational exposure, and it eliminates any chance of confirming what actually happened. Complete the interview normally. Document immediately afterward with timestamps and specific observations. Preserve the recording before taking any other action. Escalate to a named internal owner rather than handling it at the recruiter level.

And verify through an independent channel before drawing a conclusion, because a bad connection and a synthetic feed can look remarkably similar. Being wrong about this in either direction is costly.

Where this fits

The seven questions above are one piece of a longer verification sequence we put together for Human Capital in the Age of AI, our guide to hiring and building teams in a market where the old signals have stopped working reliably.

The full guide covers the complete escalation procedure, what to do about client notification, how to structure verification across the funnel rather than at a single checkpoint, and how sourcing strategy itself changes your fraud exposure before a single interview happens.

Read the guide → 

 

References:
View full post