How to Spot a Deepfake Video in 2026: A Practical Checklist

In early 2024, a finance employee at the engineering firm Arup joined a video call with the company's CFO and several colleagues, then wired out HK$200 million (about US$25.6 million) across 15 transfers. Everyone else on that call was a deepfake. The employee had even suspected the initial email was phishing — seeing familiar faces on video is what dissolved the doubt.

To spot a deepfake, slow the footage down and interrogate the face: boundary seams at the hairline and jaw, unnatural blinks, lip-sync drift, and lighting that doesn't match the room. Then check the audio, the source, and — before you act on anything — verify through a second channel. No single tell is definitive anymore; the checklist is what works.

Why your gut needs a checklist

People wildly overestimate themselves on this task. When iProov showed 2,000 UK and US consumers a mix of real and AI-generated images and video in early 2025, just 0.1% correctly classified every sample — two people in two thousand — even though participants knew they were being tested. Meanwhile 57% said they were confident they could spot a deepfake. That gap between confidence and performance is exactly what fraudsters monetize, and the volume is industrial now: Entrust's 2025 Identity Fraud Report logged a deepfake attempt every five minutes during 2024.

A checklist fixes the two failure modes of gut instinct: it slows you down, and it points your attention at the specific places where synthesis still breaks. Those places have moved over the years — this is the 2026 map. (For the broader numbers on fraud and detection rates, see our deepfake statistics roundup.)

How to spot a deepfake: the face-level checks

Face-swap and reenactment deepfakes share one architectural weakness: a generated face has to be composited into real footage, frame after frame. The seams are where to look.

The boundary

Watch the perimeter of the face — hairline, jawline, ears, neckline — especially during fast head turns or when a hand crosses the face. Look for a soft halo or double edge, skin tone that steps at the jaw, earrings that flicker, or wisps of hair that smear like wet paint for a frame or two. Occlusions are the hardest case for swap models, so a hand brushing hair away is a gift: pause and step through it.

The eyes

Blink failure, the famous 2018 tell, is mostly fixed. What remains is subtler: eyes that look glassy or slightly "locked" while the head moves, corneal reflections that don't match each other or the room's actual light sources, and eyelids that lose their crease mid-blink. Step through a blink at quarter speed; real lids deform and cast tiny shadows, synthetic ones often just interpolate.

The mouth

Lip-sync models put their error budget inside the mouth. Common artifacts: teeth rendered as a continuous white strip without individual edges, a tongue that stays suspiciously still through speech, and audio that runs a beat ahead of or behind the lips — most visible on plosives like p and b, where lips must fully close. If the speaker's mouth is doing generic open-close motion while the audio carries crisp consonants, be suspicious.

The light on the face

A swapped face inherits lighting baked in from its source data. Compare it to the scene: window light from the left should warm the left cheek and shadow the right; a face that stays evenly lit while the person walks past a lamp is composited. This is the same physics check that exposes still images — our guide on telling whether an image is AI-generated covers it for photos, and it transfers directly to paused video frames.

Listen before you look again

Audio deepfakes ride along with most video fakes, and they have independent tells. Do one pass with your eyes closed. Cloned voices tend to be too clean: no audible breaths between phrases, no mouth clicks, no room reverb that matches the visible space. Prosody is the other giveaway — emotional flatness, evenly spaced words, stress landing on odd syllables, and a total absence of the false starts and mid-sentence corrections that riddle real speech.

None of this is conclusive alone. A podcast studio produces clean audio legitimately. But a "CEO" calling from a car with zero road noise and studio-grade diction? That mismatch between claimed context and acoustic reality is the tell.

Sync is the intersection check. Fake audio and fake video are usually generated separately and married afterward, so watch for drift that comes and goes — perfect alignment in one sentence, a two-frame lag in the next. Genuine recordings don't oscillate like that; an encoding problem delays audio consistently, not intermittently.

Context beats pixels

Arup's employee lost the pixel game but could have won the context game with one phone call. Build these habits for any video that asks something of you — money, credentials, outrage, a share:

• Trace the source. Where did this clip first appear? An official account, or a three-day-old account with eleven followers? Reverse-search a few key frames to find earlier versions; recycled or re-captioned footage outnumbers true deepfakes.
• Check for provenance signals. Some AI video now ships with C2PA Content Credentials and visible watermarks, and platforms increasingly label synthetic media. AI Detector 360 reports surface these provenance signals alongside the pixel analysis, but the caveat stands everywhere: labels help when present, and they're routinely stripped by re-encoding, so absence means nothing.
• Verify out-of-band. For anything involving money or access: hang up, call back on a number you already have, or confirm over a separate channel. This defeats a perfect deepfake, which no visual check can claim.

Live calls deserve one extra move. Real-time face swaps have improved enough to sit through job interviews and board calls, but they still handle geometry changes badly. If a call feels wrong, ask for something a live rig hates: a slow head turn to full profile, a hand waved in front of the face, standing up and stepping back from the camera. Watch what happens at the moment of occlusion — and treat a sudden "bad connection" right after the request as an answer in itself.

The five-minute verification workflow

When a video matters enough to verify properly:

1. Watch normally, noting your gut reaction (1 min).
2. Re-watch at 0.25x, running the face checklist: boundary, eyes, mouth, lighting (2 min).
3. Audio-only pass for breaths, prosody, room tone (30 sec).
4. Source check: earliest upload, account history, key-frame reverse search (1 min).
5. Detector scan. Upload to the AI Detector 360 video detector, which samples frames across the whole timeline and scores each one — so a fake face that only appears in the middle third still gets caught. You'll get a per-frame timeline, an overall probability and an honest confidence level, not just a red or green light (30 sec).

Fully AI-generated footage — the Sora and Veo variety, with no real camera involved — fails in different ways than face swaps, and we've written a separate guide on detecting AI-generated video for that case. For suspicious thumbnails and stills pulled from videos, run them through an AI image detector first; it's faster and often decisive.

If the video is evidence in something serious — fraud, harassment, a news story — preserve the original before it vanishes. Download the file rather than screen-recording it, note the URL, uploader and timestamp, and keep the first copy untouched. Every re-upload and screen capture adds compression that destroys exactly the frame-level detail a detector needs, so the earliest file you can secure is also the most testable one.

Two honest closing notes. First, high-end deepfakes can pass every visual check on this page; that's why source verification and out-of-band confirmation carry the real weight. Second, detection tools — ours included — lose accuracy on heavily compressed reposts. Stack the checks, trust convergence, and never let one channel move your money.