How to Tell If an AI Video Is a Deepfake

Telling if an AI video is a deepfake comes down to five places synthesis still struggles to fake convincingly: audio-to-lip timing, blink rate, shadows across consecutive frames, the edges where a generated face meets real footage, and the texture of a cloned voice. A doctored photo only has to fool you for one frame. A deepfake video has to hold up for thousands of frames in a row while staying locked to a soundtrack, and that's where the seams show. Below is what actually separates real footage from a synthetic clip, plus the tools built to catch what a human eye misses.

Why Video Deepfakes Fail Differently Than Fake Photos

Our guide to spotting an AI-generated photo covers static artifacts: warped hands, garbled text, backgrounds that don't connect. None of that transfers to video. A single still frame from a deepfake can look flawless. The forger's problem is that a face has to move, speak, blink, and cast a shadow the same way for thousands of consecutive frames while staying locked to an audio track. That's expensive to fake, and it's where video-only tells live. Video deepfakes also sit inside a wider set of AI risks that show up anywhere synthetic content can pass as real. Video is the highest-stakes format, because people trust a face and a voice more than they trust text.

The Video Tells, at a Glance

Run through this on any clip you're unsure about. No single tell is proof, but two or three together are a strong signal.

Audio and Lip-Sync: the Fastest Check

Check this first, it's the hardest thing for a forger to fix. Generating a face is one model's job; generating matching audio is another, and the two are often produced separately. Slow the clip down if your player allows it and watch the mouth on plosive sounds, the ones needing a hard shape like a closed lip for "b" or "p". In real footage, lip closure and sound arrive together within a frame or two. In a synthetic clip there's often a soft drift, sound landing slightly ahead of or behind the mouth shape that should have produced it. Researchers building deepfake detection tools lean on exactly this signal, running independent transcriptions of the audio and the lip movement and flagging clips where the two disagree.

Blinking, Eye Movement, and Micro-Expressions

Human blinking isn't random, but it isn't perfectly regular either. It clusters around speech pauses and averages 15 to 20 times a minute, with the interval between blinks varying. Early face-swap models notoriously under-blinked, because their training data was scraped from photos, where eyes default to open. Newer generators fixed the average rate but often produce blinks on a suspiciously even rhythm, or eyes fixed on the lens instead of making the small involuntary shifts a person makes while thinking. Watch a face for thirty seconds without the audio. A metronome blink pattern, or eyes that never break contact with the camera, is a flag.

Lighting, Shadows, and the Edges Around a Face

Frame-to-frame lighting consistency is one of the hardest things for a generator to hold. Watch the shadow under the chin or beside the nose as the person turns their head. In real footage that shadow moves predictably, tied to one light source in the room. In a synthetic clip the shadow can subtly reset or shift direction between frames, because the model generates each frame without a perfect anchor to the last one. The other place to look is the boundary where a generated or swapped face meets the frame: hairline, ears, the jaw against a collar. A faint shimmer or "swimming" texture there as the head moves is temporal warping, a reliable tell because it's genuinely hard to eliminate across an entire clip. Bitdefender's breakdown of deepfake tells covers several more variants of this same edge-artifact pattern.

Voice-Cloning Red Flags in the Audio Track

A cloned voice can nail a person's pitch and accent from a few seconds of source audio, but it usually struggles with breath sounds, pacing on emotional words, and background room tone. Real speech has small intakes of breath between phrases; cloned audio often omits them or places them mechanically at the start of every sentence. Genuine emotional speech speeds up, slows down, and cracks in places a synthetic voice smooths into an even cadence. Check whether the ambient room tone matches the room shown on screen. A flat, echo-free voice over video of someone supposedly in a car or a busy office is a mismatch worth noticing.

A Real 2026 Case: What an $83,000 Deepfake Video Looked Like

This isn't theoretical. In August 2026, ScamWatch HQ reported that an Ontario woman lost $83,000 after encountering deepfake videos of Canadian Prime Minister Mark Carney promoting cryptocurrency platforms. The clips followed a pattern now standard for this fraud: a short video using a cloned face and voice, edited to resemble a fragment of a legitimate broadcast interview, routing viewers to a platform run by a fake "account manager" who kept the balance climbing until a withdrawal required one more deposit. The target choice matters: deepfake investment fraud consistently clones people with financial credibility, central bankers, finance ministers, known investors, because a fabricated endorsement from someone like that supplies institutional legitimacy an ordinary scam ad can't fake.

The same reporting notes the Better Business Bureau warned that same month about deepfake celebrity endorsements pushing supplements, and that Meta and ByteDance removed more than 45,000 deceptive deepfake ad campaigns in 2026 alone, a count that by definition excludes everything that ran and worked before it was caught. By ScamWatch HQ's estimate, deepfakes now account for roughly 11% of all global fraud. If you run or advise a business, our guides on what to do if someone deepfakes your business and how to spot an AI scam cover the response side in more depth.

Detection Tools Built for This

You don't have to rely on your eyes alone:

• Intel FakeCatcher reads photoplethysmography signals, tiny changes in facial pixel color caused by blood flow under real skin, which a generated face doesn't reproduce correctly. Intel reports roughly 96% accuracy in controlled testing and around 91% on real-world video.

• Reality Defender runs an ensemble of models across video, audio, image, and text for a real-time risk score. It's aimed at financial institutions and enterprise security teams, but it shows where serious video authenticity verification is headed: models cross-checking each other instead of one person squinting at a screen.

Neither tool is infallible, and both are built for institutions rather than one person deciding whether to trust a call. The manual checklist above still does most of the work day to day.

A Five-Step Check Before You Trust a Video Call

A live conversation gives you options a pre-recorded clip doesn't when you need to spot a fake video call in the moment.

1. Ask the person to turn their head slowly in profile. Real-time face-swap tools struggle with extreme angles and glitch at the jawline.

2. Ask an off-script question with a specific, personal answer. Live overlays add latency, and a scripted attacker may stall.

3. Watch for lag between a sudden movement, a hand crossing the face, and the video catching up cleanly.

4. Cross-check on a second channel. Call the person back on a known number before acting on anything financial.

5. Treat urgency as a signal, not an instruction. Documented deepfake fraud leans on time pressure to stop verification.

Video interviews are a documented vector for impersonation, too. Our guide on how to spot an AI-generated resume when hiring pairs well with this checklist if you're vetting people you've only ever met on a screen.

Frequently Asked Questions

What's the fastest way to tell if a video is a deepfake?

Check audio-to-lip timing first. A slight drift between the sound and the mouth shape producing it is often visible within the first ten seconds, especially on hard consonants like b and p.

Can deepfake videos fool voice authentication systems?

Some can, particularly systems that only check voiceprint similarity. Voice-cloning tools pass casual authentication now, which is why banks increasingly layer in liveness checks and out-of-band verification rather than relying on voice alone.

Is there an app or tool that detects deepfake videos automatically?

Yes. Intel FakeCatcher and Reality Defender are two enterprise-grade options, and consumer scanners exist too, though accuracy varies. Treat any automated result as one input, not a verdict.

How common are deepfake video scams in 2026?

Common enough to be a mainstream fraud category. Reporting from August 2026 put deepfakes at roughly 11% of all global fraud, with platforms removing tens of thousands of deceptive ad campaigns over the year.

Do deepfake videos look worse after you screenshot or re-record them?

Often, yes. Re-recording or re-compressing a clip tends to amplify existing artifacts, edge flicker especially, because compression targets the same high-frequency detail where synthesis errors already live.