The Hong Kong finance employee logged into the video conference call at 9:47 on a Tuesday morning in February 2024. The CFO was there. Two regional heads. The compliance lead. The faces in the grid matched the faces he saw on the company directory. The voices matched what he expected. He authorised fifteen transactions over the next several minutes. The total was HK$200 million — roughly $25 million.
Every person on the call except him was an AI-generated video overlay. The CFO had never scheduled the meeting. The compliance lead had been in a different country. The “people” on the call were deepfakes built from publicly available footage of the actual executives. Each was run through a real-time face-swap pipeline and composited into a video conferencing session.
By the time the company verified the transactions with the real CFO, the funds had moved.
This is no longer the most extreme example. In May 2024, the British engineering firm Arup confirmed the $25M case was theirs. By mid-2025, Mandiant and the FBI’s IC3 had logged dozens of similar incidents at smaller dollar values. Several hit US and European companies, and several involved solo employees rather than groups. The FBI’s 2024 Internet Crime Report flagged deepfake-enabled business email compromise as one of the fastest-growing categories (FBI IC3 2024 Internet Crime Report, ic3.gov).
The technology has matured. The question for anyone receiving an urgent video call is now simple: how do you verify the person you are looking at is real?
The short answer. You verify a suspect video out-of-band, not by staring at it. Deepfakes still leave tells such as lip-sync mismatches on B/P/M sounds, odd blink timing, mismatched eye reflections, and hairline or hand artifacts. But modern fakes defeat the human eye on a compressed video call, so the reliable move is to break the frame. Ask the person to turn fully to profile or reference a private shared memory, and always call them back on a known number before acting on any urgent request. For a recorded clip, an automated detector adds a confidence score.
How to spot a deepfake video
A handful of structural weaknesses in current generative video models produce artifacts that a careful observer can sometimes catch. None of them is conclusive on its own. All of them are getting harder to spot as models improve.
Lip-sync mismatch. Real human lip movement is governed by the physical mechanics of forming each phoneme — a B sound requires lips to fully close before the vowel; an F sound requires the upper teeth on the lower lip; a hard P creates a slight puff that you can sometimes see in cheek movement. Deepfake models trained primarily on visual data often produce lip movement that matches the timing of speech but not the precise shape of each phoneme. Watch for B, P, and M sounds where the lips do not fully close, or F and V sounds where the teeth-and-lip contact looks soft. The mismatch is usually subtle on a single frame and more obvious if you slow the video down.
Blink rate and pattern. Humans blink on average 15–20 times per minute, with a natural rhythm that varies based on focus and emotion. Early deepfakes blinked too rarely because face datasets did not include enough closed-eye reference frames. Newer models have addressed this — they generate blinks at human-typical rates. But blink mechanics still trip them up. A natural blink takes 100–400 milliseconds, with the eyelid moving down faster than it moves up. Many deepfakes generate symmetric blinks (same speed both directions) or blinks that look slightly delayed relative to whatever should trigger them.
Eye movement and pupil reflection. Real eyes show consistent specular highlights across both pupils — the reflection of light sources in the environment appears at matching positions in each eye. Deepfakes often have inconsistent or absent eye highlights, or highlights that do not match what should be visible in the room. The eye direction can also drift in ways that do not correspond to natural attention — looking slightly off-camera when the person should be looking directly at you.
Hand and finger artifacts. Generative models have improved on faces faster than on hands. If the person on the call uses their hands while speaking, watch for fingers that briefly merge, hands that hold objects in physically implausible ways, or fingertips that shimmer at the edges. The “six fingers” trope from early AI image generation is mostly past, but real-time video models still struggle with rapid hand motion.
Audio splicing and prosody breaks. A real-time voice clone running alongside the video overlay typically does not generate the same speech the deepfake would produce naturally. The voice comes from one model; the lip animation comes from another. Listen for pauses that fall in unusual places, breath sounds that do not match the chest movement, and prosody (the musical contour of speech) that breaks between sentences as the model resets. A normal speaker will rarely have a perfect, identical-energy pause between every sentence.
Hairline and ear artifacts. The boundary between the deepfake face and the underlying head is one of the hardest things for the model to composite cleanly. Watch the hairline as the person turns their head; a deepfake often shows a slight blur, lag, or color shift at the boundary. Ears are a related weakness — they are unique to each person and rarely present in the training data at the right angles, so deepfakes often produce slightly wrong ear shapes or ears that flicker as head angle changes.
Metadata stripping. This is not visible in the video itself but matters if you have the file. Real video recordings carry EXIF or container metadata: device model, recording timestamps, codec parameters, sometimes GPS. Deepfake outputs are re-encoded after generation and lose all original metadata; the only metadata you see is the encoder used to produce the final file. A video where the only metadata says “ffmpeg” or “OBS” is not necessarily a deepfake — those are common in legitimate processing pipelines — but the absence of any device-of-origin metadata in a video that purportedly came from a phone is a flag.
Why human inspection fails on modern models
The artifacts above are real and exploitable. The hard truth is that they are also unreliable.
Several university studies between 2022 and 2025 measured human ability to detect deepfakes under realistic conditions. The MIT Media Lab’s Detect Fakes project found that untrained participants identified high-quality deepfakes correctly only 57% of the time, barely above chance. Trained participants did better, around 71%. Both numbers drop further when the video is short (under 10 seconds) or low-resolution. They also drop over a video-call codec, whose own compression artifacts mask the deepfake signal.
A real-time video call is the hardest case. The codec running on Zoom, Teams, or Google Meet introduces compression. Bandwidth fluctuations introduce frame drops. Lighting and camera quality vary. The human eye has perhaps two or three seconds to assess each suspicious moment before the conversation moves on. Under those conditions, the structural artifacts that a careful frame-by-frame analyst would catch get lost in the noise of an ordinary video call.
The Hong Kong case illustrates the limit. The employee was a trained finance professional. The call was on his company’s normal conferencing platform. He had seen the CFO speak in person and on dozens of prior calls. He still complied with the transfer requests because the call looked normal enough and the social pressure to act quickly was high.
Detection is not just a perception problem. It is a context problem.
What an automated check actually measures
Automated deepfake detection works by analysing signals the human eye is not optimised to see.
Face artifact detection. Models look at pixel-level statistics that change when a face is generated rather than recorded. Real cameras produce sensor noise patterns that correlate across frames in specific ways; synthetic faces produce different correlation patterns. Detection models, trained on large datasets of known-real and known-synthetic video, learn to flag the synthetic distribution. Intel’s FakeCatcher takes a related approach by analysing pulse signals — micro-color-changes in skin caused by blood flow — which deepfakes typically do not replicate convincingly because they were not in the training objective.
Voice splice detection. The audio track is analysed separately. Real speech has continuous prosodic contours and consistent background noise floors. Spliced or synthesised audio shows discontinuities at splice boundaries, voice-clone signatures (the particular spectral artifacts each TTS model produces), and mismatches between speaker identity and acoustic environment. Pindrop, Reality Defender, and several open-source projects publish detectors trained on these signals.
Lip-sync correlation. A pre-trained model can measure, frame by frame, how well the visible lip movement matches the audio waveform. Real speech shows high correlation; deepfakes typically show lower or inconsistent correlation. This signal is harder for generative models to defeat than face-artifact signals because lip-sync is a joint property of two separately generated streams.
Metadata and provenance. Automated checks read the file’s metadata and compare against known patterns. C2PA (Coalition for Content Provenance and Authenticity) provenance signatures, when present, cryptographically attest to the recording chain. Their absence does not prove a deepfake, but their presence on a video from a C2PA-supporting camera or platform raises confidence.
Temporal consistency. Frame-to-frame analysis catches inconsistencies that any single frame would not reveal — a freckle that disappears for two frames, a strand of hair that shifts in a way uncorrelated with head movement, a background that shows compression artifacts but the face does not. These are hard to spot by eye and obvious to a comparison model.
Combined, these signals produce a confidence score. No single automated detector is perfect. The most reliable commercial systems publish accuracy in the 90–96% range on test datasets, and lower in the wild against the newest generative models. Still, combining signals raises confidence enough to be useful. A video with low face-artifact, low voice-splice, and missing C2PA provenance reads as high-confidence synthetic. A video with strong face artifacts, weak voice splice, and signed provenance reads as high-confidence real.
The arms race continues. Each generation of detectors prompts a generation of generators that work around the latest signals. Detection is a moving target, not a settled answer.
What to do during a suspicious video call
If you are on a call and something feels off, the in-the-moment options are limited but real.
Ask the person to perform a specific physical action. Turn their head fully to the side. Cover and uncover an eye. Hold up a specific number of fingers in front of their face. Real-time deepfake overlays often degrade visibly when the face is partially occluded or rotated past the angles in the training data. A request to “show me your full profile, looking at the corner of the room” is unusual enough that a scripted impostor will struggle.
Reference something only the real person would know. This is the voice-clone defense and it applies to video too. A specific shared memory, a recent private conversation, an inside joke. A deepfake operator running the call from a script does not have that context. The pause while they improvise is often visible.
Move the conversation to a separate channel. “Let me call you back on your mobile in two minutes.” A scammer running a deepfake cannot intercept your outbound call to the real person’s known number. If the real person answers, the impostor on the original call was not them. If the impostor objects to switching channels — “this is urgent, I cannot wait, please act now” — that objection is itself the warning sign.
Slow down. Every documented deepfake-fraud case relies on time pressure. The CFO who needs the wire today. The board member who needs the credentials right now. The family member who needs the money before the next phone call. Slowing the decision by even 30 minutes — to verify out-of-band — defeats almost every documented attack.
What you can put in place ahead of time
For organisations, the structural defenses are concrete. Use dual-channel verification for high-value transactions. Require callback protocols on known numbers for any wire-transfer authorisation. Use hardware-key MFA for any system the call might be trying to reach. And run employee training that walks through deepfake scenarios directly, rather than abstract “watch out for phishing” guidance.
For individuals, the same logic applies at smaller scale. A family code word, agreed in advance, that any caller in distress is expected to be able to provide. A protocol of always calling back on a known number before acting on any urgent request. A general posture that authentic urgency is rare and manufactured urgency is the norm in fraud.
For media verification, when you see a video of a public figure saying something inflammatory, the fastest check is corroboration. Ask whether reputable news outlets are reporting the same content with footage from independent angles. A single video from a single source, on a single platform, with no corroboration, is the highest-risk profile for synthetic content.
When automated verification helps
A consumer-grade deepfake check is useful in two scenarios.
First, asynchronous video: a recording sent to you in an email, posted to social media, or shared in a group chat. You have time to run the file through a detector before you act on it or share it further.
Second, suspicious live calls where you can record a short segment. Most enterprise conferencing platforms allow recording. A 10-second clip is enough for most detectors to produce a confidence score.
KnytX Deepfake Check isn’t live yet. Once it launches, it’s planned to offer 10 checks per month on the Pro tier and unlimited checks on the Executive tier. Each check is built to produce a face-artifact score, a voice-splice score, a lip-sync correlation, and a metadata report. The output is a confidence rating, not a binary verdict. It is meant to raise or lower your trust in a video, not to replace the dual-channel verification that should be your first defense. The planned feature breakdown is on the deepfake check product page.
The Hong Kong employee did not have time to run a detector mid-call. Most victims do not. The detector matters most for the videos that arrive before the decision. Think of the executive video that “needs review,” the family video that “needs sharing,” or the news clip that “needs sharing now.” Those are the moments where 30 seconds of automated verification is the difference between forwarding and pausing.
What you can do right now
See what a scan finds → walks through an example report of the photos and footage public sites and data brokers can expose. The training data for a deepfake of you comes from somewhere. Reducing the available source material does not prevent every attack, but it raises the cost of a high-quality impersonation. KnytX Shield isn’t live yet; join the waitlist to run your own scan at launch.