Have you ever been on a video call and felt something was slightly off about the person on screen? That instinct might be worth paying attention to. AI face-swapping technology, commonly known as deepfake, can now run in real time during live video calls, and the tools that make it possible are easier to access than most people realize.

TrendLife researchers tested 13 of these detection methods against real deepfake tools to see which ones actually hold up. An iProov study (a biometric security firm) of 2,000 participants found that only 0.1% could correctly identify all AI-generated deepfake content they were shown, a finding that lines up with what the testing showed. Human detection ability is far more limited than most people assume.

The answer is complicated, but there’s genuine good news in it too.

Pre-recorded and real-time: two completely different problems

Before getting into what works, deepfakes come in two forms with very different detection difficulty.

• Pre-recorded deepfakes are generated ahead of time. Because there’s no real-time constraint, creators can adjust and re-render until the result looks nearly flawless.
  
  A 10-second video at 720p takes roughly 2 to 3 minutes to generate; at 1080p, 6 to 8 minutes. Even when artifacts appear during generation, they can be cut out or the video re-generated entirely. The final product often has no visible flaws.
  
  
• Real-time deepfakes work differently. During a live video call, every single frame has to be detected, tracked, and blended together within milliseconds.
  
  TrendLife researchers ran tests on a consumer-grade laptop with an RTX 4060 GPU (the kind of hardware a typical consumer might own), and the result was roughly 10 frames per second, about one-third of a normal video call’s frame rate, with visible stuttering.

One practical factor: most people take video calls on phones or tablets. Small screens already make it harder to see fine detail, which means a mid-quality real-time deepfake can pass more easily in that setting than it would on a large monitor. Every method described below needs to be read with that context in mind.

The two methods that held up in testing

Two of the 13 methods came through consistently against real-time deepfakes.

• Ask the person to wave a hand quickly across their face
  As the hand moves quickly across the face, the face-swap model can’t keep up with tracking what’s covered. Facial features (eyes, nose, and mouth) can appear to stick to the hand, as if they have ghosted through and onto it. This showed up consistently in testing. You don’t need a reason. Just ask. You can try waving your own hand in front of your face first.
  
  Most people will instinctively mirror the gesture, giving you a natural opportunity to observe. If they don’t follow along, you can always ask them directly to wave a hand across their face.
  
  
• Ask them to slowly turn their head
  Face-swapping models are trained mostly on front-facing images. When a face rotates past a certain angle, the model loses stable tracking points. The real face underneath can bleed through at the edges, producing brief double outlines or flickering along the jawline and hairline.
  
  A natural prompt works well here. Something like “I think there’s something behind you” gives the person a reason to turn without making the request feel deliberate.

Other methods (inability to blink, unnatural eye movement, lip sync) landed in “partially effective” territory in testing. Whether they reveal anything depends heavily on the software and hardware the other person is using. Treat them as supplementary signals rather than stand-alone tests.

Pre-recorded deepfakes: one unexpected weak point

For pre-recorded content, almost none of the 13 methods produce reliable results. The time available to fix problems before sending a video is the key difference, and most creators use it.

There is one exception: teeth.

Most face-swapping software uses a single source photo, which typically captures one angle and one expression. When the subject opens their mouth, the AI has to infer what their teeth look like. If the person being impersonated has distinctive features (a gap, a crooked tooth, a missing one) and the source photo doesn’t show them smiling, the inference often falls short. The result can look like a single white block without natural gaps or depth.

Across all 13 methods TrendLife tested, this was the only visual indicator that reached “partially effective” against pre-recorded deepfakes. That said, it only applies in specific conditions and isn’t something to count on every time.

If you suspect a video, try taking a screenshot and zooming in on the teeth. Details that are easy to miss in motion often become more visible in a still image.

How to stay safe

1. During a video call, run two checks. Ask the person to wave a hand quickly across their face, then ask them to slowly turn their head. These are the two methods TrendLife researchers confirmed work against real-time deepfakes, and neither requires any tools. If they find an excuse to avoid either test or refuse to cooperate, that’s a signal worth taking seriously.
   
   
2. Watch the video for smoothness. Real-time deepfakes demand significant processing power. On typical consumer hardware, the result is a noticeably choppy frame rate compared to a normal call.
   
   
3. Be cautious when strangers initiate video calls. Especially when the conversation moves quickly toward money or personal information. There’s no need to decide anything in the moment. Give yourself time to verify.
   
   
4. (For friends and family) Hang up and verify through another channel. Don’t make decisions involving money or personal information during the call itself. Follow up by text or phone using a number you already know is theirs.
   
   
5. (For friends and family) Set up a verification word in advance. Scammers sometimes impersonate people you know. A shared question or phrase that only the two of you would know is a quick way to confirm someone’s identity when it matters.
   
   
6. Use security tools to spot fake video calls. Trend Micro ScamCheck‘s AI Video Scan detects AI face-swapping scams during video calls in real time, alerting you to potential impersonations.

You now know more than most people

Deepfake technology is genuinely impressive, but it isn’t without limits. The processing constraints of real-time face-swapping create specific, testable weak points, and the two methods above are grounded in what those constraints actually are, not what feels intuitive.

For pre-recorded video, staying appropriately skeptical and using detection tools is the most practical approach right now. The underlying principles are unlikely to change quickly: it’s difficult to make real-time face-swapping perfect without sacrificing speed, and speed tends to leave traces.

The technology will keep improving, and detection methods will keep up with it. In a live video call, though, there’s always room to ask a simple question. Knowing which questions to ask is already a real advantage.