He blinked at the right moments. He smiled when I made a joke. He nodded along, thoughtfully, while explaining his approach to a difficult stakeholder conversation. But something was off. Not anything you could point to. The kind of thing you feel on your skin before you can name it. It was just a little too good to be true.
He wasn't real. The face was. The person underneath it was someone else entirely, feeding answers through voice modulation, or simply leaning on AI in real time, while a stranger's likeness moved across the screen, mapped to match his expressions. Researchers who study this now say a fake candidate capable of passing a full video interview can be built in about seventy minutes, with tools that require no image editing experience at all.
This isn't a hypothetical from a security conference. It's something I've run into myself, in real interviews. It's not common… yet. In my experience it's closer to five percent of candidates. But it happens often enough that it changes what you learn to notice. An experienced recruiter who has sat across from hundreds of real candidates develops a feel for the ones who aren't quite real. Not proof. A feeling worth trusting enough to slow down and look closer. Deepfake fraud attempts in hiring jumped by roughly 1,300 percent in under two years. By early 2026, close to four in ten candidates tracked across live interviews were being flagged for AI assisted cheating, a rate that tripled in three months. Half the businesses surveyed on the topic say they've already encountered some version of it. Google and McKinsey have both reintroduced mandatory in person interviews, not as a perk, as a defence.
The interview used to guarantee one thing. Two real people, paying attention to each other. That guarantee is gone. A candidate can now show up as just a face and a voice, with nobody real behind either. Even if it only happens occasionally, that changes what the interview is testing for. Not just whether someone can do the job. Whether the person in front of you is the one who'll actually show up to do it.
This isn't an abstract trend if you're a founder running interviews between customer calls. It's the exact process you're relying on to make a decision that shapes everything downstream. A bad hire costs differently at twelve people than at twelve thousand. If that process can be faked, even sometimes, it isn't testing what you think it's testing.
The instinct, when something breaks, is to patch it with more process. A liveness check before round one. An identity scan bolted onto the calendar invite. These help at the margins. But the real failure was never technical. Somewhere along the way, the interview stopped being a moment two people show up for. It became a box in a pipeline, one that something, or someone pretending to be something, just needs to pass through.
None of this needs new software. It needs an experienced interviewer back in the room, someone who has done enough of these to trust the feeling when something is off. A few things worth watching for, if you want somewhere to start.
Ask for something unscripted. A live walk through of a project, a screen share, a question nobody prepared for. Deepfake avatars are trained on rehearsed material and struggle the moment a conversation goes somewhere nobody scripted.
Watch the lag. A slight delay between what's said and how the face reacts, a smile that lands half a beat late, a blink that doesn't quite match the sentence. Voice and face are usually generated separately, and they don't always stay in sync.
Change the frame. Ask someone to turn their head to the side, hold a hand up in front of their face, or move to a different spot in the room. Face mapping software still struggles with angles and lighting it wasn't trained on.
Notice the emotional flatness. Real surprise, real hesitation, real irritation look different from a performance of them. If every answer arrives a little too smooth, that's worth a second look, not a pass.
None of it replaces judgment. It just gives an experienced interviewer somewhere to point it. The webcam was never the weak link. The person watching through it always was.