FAXTR
FAXTR Guide · AI Detection

How to Spot a Deepfake Video Call

A face and a voice you recognize no longer prove who's on the line. Here's what actually catches a real-time impersonation — and the one rule that works even when the video looks perfect.

For most of the internet era, video was treated as proof. If you could see someone's face moving and hear their voice, identity felt settled. Real-time deepfakes have quietly ended that assumption. Tools now exist that map one person's face and voice onto a live call — expressions, lip movement, and speech pattern included — convincingly enough to survive a short conversation. The FBI has warned that scammers use AI video during live calls to impersonate executives and even officials, and a widely reported finance case in Hong Kong ended with a worker approving transfers after a video meeting in which every other participant was synthetic.

That sounds alarming, and it is — but the practical defense is more reassuring than the technology. Deepfake calls almost never win on the strength of the visuals alone. They win when a convincing face makes you skip the checks you'd normally run on a strange request. So the most useful skill here isn't pixel-peeping. It's noticing what you're being asked to do, and refusing to let a familiar face rush it.

Start with the request, not the face

Before you study the video, listen to the ask. These operational red flags matter more than any visual glitch, because they're present in nearly every deepfake scam and they don't depend on how good the rendering is.

01

Urgency plus money or access

A wire that has to go out before end of day, an MFA code they need read aloud, a gift-card errand, a password reset they'll 'walk you through'. Real colleagues and family almost never combine a hard deadline with a request that moves money or credentials.

02

The call came to you, unprompted

You didn't schedule it. A 'CEO' or 'relative' appears on a call you didn't set up, on a platform they don't normally use. The convenience of catching you off guard is the point.

03

Pressure to keep it off the usual channel

They discourage you from confirming by text, email, or a callback — 'no time', 'don't tell the others yet', 'just do it now'. Secrecy is the tell, not the face.

04

A request to bypass a normal control

Skip the approval step, ignore the vendor-change policy, install a remote-access tool 'so I can help'. If the ask is to route around a safeguard, the safeguard is what's being attacked.

Visual and audio tells worth a glance

The rendering catches up every year, so treat these as hints that raise suspicion, not proof. A clean-looking face doesn't clear anyone; a glitchy one on a bad connection isn't necessarily fake either. Look for a cluster of these, not a single frame.

Lip-sync drift

Watch the mouth against the audio for a few seconds. Real-time models often lag or smear on fast speech, plosives, and laughter.

Edges that blur on motion

The boundary where the face meets hair, glasses, or a hand tends to shimmer or warp when the person moves quickly.

Blinking and gaze that feel off

Too little blinking, or blinks at odd intervals; eyes that don't quite track where they're looking.

Lighting that doesn't match the room

Face light and background light disagree — a warmly lit face in a cool room, shadows falling the wrong way.

Teeth, tongue, and ears glitch

Small, high-detail areas are hard to render live. Teeth that merge, an earring that flickers, a tongue that smears on wide vowels.

Liveness tests you can run mid-call

If a call feels wrong, you don't have to guess from a distance — you can push the model into territory it handles badly. These take seconds and are hard to refuse without looking evasive.

Ask them to turn fully to profile

Many live face-swap models are trained mostly on frontal views. A full side profile — nose, ear, jawline in one plane — is where they distort or briefly collapse.

Have them wave a hand slowly across the face

Passing a hand over the eyes and nose forces the model to redraw occluded areas in real time. Watch for the hand melting into the face or the face rebuilding behind it.

Ask something only the real person knows

Not a fact that's searchable online — a shared detail: what we argued about last week, the name of the café by the office. Impersonators run on public information.

Request an unscripted physical action

Pick up the mug on your left, stand up and step back, tap your shoulder twice. Sudden, specific movements are hard to fake convincingly on the fly.

The rule that beats a perfect fake

Here's the part to remember even if you forget everything above. When a video call asks you to move money, hand over a code, or grant access, hang up and reconnect on a channel you already trust — the person's known mobile number, an internal directory line, a walk to their desk. Don't redial the number that called you, and don't use a contact detail they gave you on the call. A real request survives a two-minute callback without complaint. A deepfake usually can't follow you onto a channel it doesn't control.

For families and small teams, a pre-agreed safe word does the same job faster. Pick a word now, share it in person, and use it whenever a call turns urgent about money or credentials. It costs nothing and it neutralizes a voice or face clone instantly, because the scammer simply doesn't have it.

A word on detection tools

Real-time deepfake detectors and liveness-verification systems are improving, and some enterprise video platforms are starting to build them in. For an individual on an ordinary call, though, there's no reliable one-click "is this person real" button yet — and any tool that claims certainty is overselling. Lean on behavior and verification, not a score. If you later obtain a recording of the call, you can run it through the same AI-video and deepfake analysis covered in our other guides, but that's an after-the-fact step, not live protection.

Not sure a claim or clip is real?

FAXTR searches 100+ fact-checking organizations in one query — free, no login required. Report a suspected deepfake so others can be warned.