Signs a Video Selfie Is Fake: 8 Red Flags to Watch

A convincing fake face no longer requires a film studio. Consumer apps and open-source models now generate a moving, talking likeness from a handful of photos, and the output is good enough to pass a casual glance on a video call or an onboarding screen. For anyone trying to confirm that the face in front of them belongs to a living person, the practical question is which cues still hold up. The clearest signs of a deepfake video tend to cluster around physics the model never learned: how light wraps a face, how skin behaves at the edges, how a real body keeps an involuntary rhythm. This report walks through eight red flags worth watching, then explains why the most reliable tell is one the human eye cannot see at all.

Researchers at Idiap and partner institutions reported that high-quality deepfakes can now carry a realistic, measurable heartbeat inherited from the source video, meaning the simple presence of a pulse signal is no longer sufficient proof that a face is real (Frontiers in Imaging, 2025).

The visual signs of a deepfake video, ranked by reliability

Manual inspection still works against weaker fakes, and knowing what to look for raises the bar for an attacker. The catch is that every visual artifact below is a moving target. As generation models improve, the cues that exposed last year's fakes quietly disappear. Treat these as warning signs, not guarantees.

The eight red flags fall into two groups: things you can see in a single frame, and things that only show up across motion and time.

• Edge and boundary glitches. Watch the hairline, jaw, and the line where the face meets the neck or glasses. Warping, flicker, or a faint seam where a swapped face is blended onto a head is a classic synthetic media detection cue.
• Lighting that does not match. Shadows falling the wrong way, specular highlights that do not track the room, or a face lit differently from its background. Work led by researchers studying lighting inconsistency has shown that forged frames often carry less stable illumination than authentic ones.
• Teeth, ears, and jewelry. Generative models struggle with fine repeating structures. Smeared individual teeth, asymmetric or melting earrings, and shifting ear shapes are common.
• Eyes and gaze. Reflections in both eyes that do not agree, pupils that drift, or a gaze that never quite locks onto the camera.
• Unnatural blinking. Blinks that are too rare, too frequent, or mechanically regular. Real-time detection systems built around blink number, duration, and cycle timing exploit exactly this, because involuntary blinking is hard to reproduce.
• Lip-sync drift. Audio and mouth shapes that fall out of step, especially on plosive sounds, point to a face animated separately from the voice track.
• Skin texture that is too smooth or too busy. Waxy, poreless skin or, conversely, a shimmering high-frequency texture that crawls between frames.
• Stiff head and neck motion. A face that moves while the neck and shoulders stay oddly locked, or motion that stutters when the person turns.

One frame versus the whole clip

Detection signal	What it catches	Visible to the human eye	Defeated by better models	Works on a still frame
Edge and blending artifacts	Face-swap seams	Sometimes	Increasingly	Yes
Lighting inconsistency	Composited or relit faces	Rarely	Partially	Yes
Teeth, ears, jewelry	Generation errors	Sometimes	Increasingly	Yes
Blink and gaze patterns	Animated faces	Rarely	Yes	No
Lip-sync alignment	Voice-driven puppets	Sometimes	Yes	No
Frame-level classifiers	Known generator fingerprints	No	Yes	Yes
Blood-flow (rPPG) analysis	Absence of living physiology	No	Resistant	No

The pattern in that table matters more than any single row. The cues a person can spot are the same cues generation models are trained to erase. The signals that survive are the ones tied to physiology and motion over time, which is where automated and contactless methods pull ahead of manual review.

Industry applications: why eyeballing fakes does not scale

For consumers, spotting one suspicious video selfie is a useful skill. For the teams that verify millions of faces, manual inspection is not a control. Fraud rings operate at volume, and a reviewer who has to study a clip for thirty seconds is already losing.

Identity verification and KYC

Remote onboarding is the primary target. An attacker who can inject a synthetic face into the camera feed bypasses document checks entirely if the liveness layer only looks for visual artifacts. KYC providers increasingly treat AI face warning signs as a starting point and layer automated liveness on top, because a face swap that fools a reviewer can be produced in minutes.

Banking and fintech fraud teams

Account-opening fraud and account-takeover both lean on synthetic faces to defeat selfie checks. Fraud teams need a signal that does not degrade as generators improve, which is why biometric liveness verification has shifted toward physiological evidence rather than pixel forensics alone.

Video call and support verification

Real-time impersonation on a video call, including the now-documented executive deepfake scams, has pushed the question from "does this look real" to "is this a live body." Behavioral cues like blinking help, but they can be scripted into a fake, so they cannot stand alone.

Current research and evidence

The academic picture has shifted quickly. Early detectors leaned on the idea that deepfakes lacked a heartbeat. Remote photoplethysmography, or rPPG, measures the tiny periodic color changes in skin caused by blood pumping through facial capillaries, and synthetic faces historically had no such signal.

That assumption is now under pressure. A 2025 study reported in Frontiers in Imaging found that modern high-quality deepfakes can inherit a realistic pulse from the genuine driving video used to animate them, so a naive "is there a heartbeat" test can be fooled. Separately, a cautionary analysis from researchers at the University of Modena and Reggio Emilia warned that rPPG heart-rate estimates are noisy enough under real-world conditions that simple global-pulse checks should not be trusted on their own.

The constructive response in the literature is to go deeper than a single global heart rate. A comprehensive review of rPPG-based detection from Torrens University Australia points toward analyzing how blood flow varies across different regions of the face. A real face shows a coherent, physiologically plausible pattern of pulse propagation; a stitched or generated one tends to show spatial inconsistencies that are extremely hard to fake on purpose. Work such as the local-attention rPPG approaches published in peer-reviewed venues pursues exactly this spatial-temporal signal, and the open-source DeepFakesON-Phys line of research from biometric labs demonstrated that physiological features can drive strong classifiers.

The takeaway for practitioners is nuanced. The presence of a pulse is no longer proof of life, but the absence of a coherent, spatially consistent blood-flow pattern remains a powerful red flag, and one that does not show up to a reviewer scrubbing through frames.

The future of deepfake detection

Three shifts are likely over the next few years. First, visual artifact detection will keep losing ground as generators close the gap on edges, teeth, and lighting, so single-frame forensics will move from primary control to supporting evidence. Second, provenance and watermarking, including approaches like noise-coded illumination that embed an invisible flicker into a scene's light source, will add a second track of defense at capture time. Third, physiological liveness will move from global pulse detection toward fine-grained spatial blood-flow analysis that is far harder to synthesize than a single averaged heartbeat.

The strategic point for verification vendors is that no single signal wins. The durable architecture is layered: visual checks to catch cheap fakes, provenance to verify capture, and contactless physiological analysis to confirm a living body is present. The signals that survive model improvement are the ones rooted in biology rather than pixels.

Frequently asked questions

Can I reliably spot a deepfake just by looking? Against low-effort fakes, yes, by checking edges, lighting, teeth, blinking, and lip-sync. Against high-end synthetic media, no. The visual cues humans rely on are precisely what newer models are trained to remove, so manual inspection should be treated as a first filter, not a verdict.

Does a heartbeat prove a face is real? Not anymore. Research in 2025 showed that high-quality deepfakes can inherit a realistic pulse from their source footage. A simple global heart-rate check can be fooled, which is why current work focuses on spatial consistency of blood flow across the face rather than the mere presence of a pulse.

What is rPPG and why does it help? Remote photoplethysmography reads the subtle color changes in skin caused by blood circulating through facial capillaries, using only a standard camera. Because it measures a physiological property rather than image appearance, it targets something a generated face does not naturally possess in a coherent, region-by-region pattern.

Are blinking and gaze checks enough on their own? No. Blink rate and gaze are useful behavioral signals, but they can be animated into a convincing fake. They work best combined with physiological liveness and visual artifact analysis in a multi-layer stack.

Circadify is building toward this layered future, with contactless blood-flow analysis designed to separate a living face from synthetic media regardless of how polished the visuals become. Verification vendors and fraud teams evaluating liveness detection can see how rPPG-based signals fit into an existing stack through an enterprise security demo.