Real-time deepfakes have gotten good enough that the old advice — "it'll look blurry" or "the lag will give it away" — doesn't hold up the way it used to. The technology has improved fast, and the gap between a convincing fake and an obvious one has narrowed considerably. That doesn't mean there's nothing you can do. There are a handful of specific tests that still reliably catch a real-time deepfake, and knowing the actual scam patterns that come with them matters just as much as the technical detection itself.
A real-time deepfake overlays a synthetic face onto a live video feed, usually run through a virtual camera so it looks like an ordinary webcam to whatever app is receiving the call. The person on the other end might be using a completely different face than their own, in real time, while talking to you. This isn't science fiction anymore — it's accessible enough that someone with a decent computer and an evening's worth of setup can run a passable version of it.
It's worth being clear about why this happens. Sometimes it's genuinely playful — someone using a fun filter just to mess around. But increasingly, it's the opening move in a scam: build a quick connection using an attractive or trustworthy synthetic face, then ask for money, crypto, or access to something once trust is established.
This is the single most reliable test available right now, and multiple independent security sources confirm it still works as of 2026. Most real-time face-swapping models are trained heavily on front-facing footage, since that's what most webcam video looks like. Ask the person to slowly turn their head to the side, toward a 90-degree profile. Watch the jawline and the area where the ear meets the face. A live deepfake will frequently glitch, blur, or flicker right at that transition — the model simply doesn't have enough training data to render a clean profile in real time.
Ask the person to move their hand slowly in front of their face, as if waving or covering their mouth briefly. Real-time face-tracking models still struggle when something passes between the camera and the tracked face — the mask can warp, lag behind the hand's movement, or briefly disappear and reappear in a way a real face never would. This is one of the simplest tests to ask for, and it doesn't come across as suspicious or confrontational — it just looks like normal conversation.
Look closely at the eyes specifically. Real human eyes have small, consistent reflections of light sources — called specular highlights — that move naturally as the head turns. AI-generated faces often get this wrong: the reflections can look static, mismatched between the two eyes, or simply absent. It's also worth comparing the lighting on the person's face to the lighting visible in their background. A face lit like it's in a studio, sitting in front of a window with completely different natural light, is a real mismatch worth noticing.
Glitching or blurring near the ear and jawline when the head turns to the side.
The face warping, lagging, or briefly disappearing when a hand passes in front of it.
Static, mismatched, or missing light reflections in the eyes, especially when the head moves.
The face lit completely differently from the visible background behind the person.
Smiles or expressions that don't quite track with what's being said, or look slightly too smooth.
A faint mismatch between mouth shapes and sound, especially on sounds like "b," "p," and "m" which require the lips to fully close.
One pattern that's become more common and is worth knowing specifically: a video call that drops abruptly right when something starts to feel slightly off, followed almost immediately by a text message asking for money — often framed as an emergency, bad signal, or a dying phone battery. This isn't a coincidence. Scammers running a live deepfake know the longer the call runs, the more likely it is to glitch and give itself away, so cutting the call short before that happens, then moving the actual money request to text, is a deliberate way to avoid exposing the fake on camera. If a video call drops unexpectedly and is quickly followed by an urgent request for money, that combination on its own is worth treating as a serious red flag — independent of whether you spotted any visual glitch at all.
The underlying scam goals tend to be familiar even when the technology is new: a sudden financial emergency that needs cash or crypto right now, an "exclusive" investment tip from someone posing as an expert, or a request to share your screen, which can be used to capture login details or access private files. None of these require a sophisticated deepfake to work — the deepfake is just there to buy initial trust faster than a typical scam conversation would.
It's worth being honest about the current state of automated detection tools, since some marketing around them oversells what they can actually do. Detection software exists and continues to improve, but it's genuinely in an ongoing back-and-forth with deepfake generation technology, and newer generative models are sometimes specifically trained to defeat the detectors built for older ones. A tool telling you something is "90% likely real" is a useful signal, not a guarantee. The behavioral tests above — the profile turn, the hand test, watching the eyes — remain useful precisely because they test something fundamental about how these models are built, rather than relying on a detector that has to be constantly updated to keep up.
Chatzyo doesn't run any kind of automated content analysis on video streams — that's true by design, not by choice, since video on Chatzyo travels directly between two devices and never reaches our servers at all. There's no system here scanning for virtual cameras or flagging suspicious streams automatically. The honest answer is that the tests above are genuinely your best tool, the same as they would be on any other video platform. If you do run into someone using an obvious fake to scam you, the right move is to skip and report them through the in-chat Report button — our guide to reporting a user covers exactly how that works and what happens after.
Beyond the in-the-moment tests, a few general habits make this kind of scam less likely to work on you in the first place. Be cautious about how much footage of yourself talking directly to a camera you post publicly — that kind of footage is exactly what's used to train a convincing fake of your own face or voice, and reducing your public footprint reduces what's available to misuse. Never act on a financial request made purely over video or voice, no matter how convincing — verify through a completely separate channel, like a different app or a phone call to a known number, before sending anything. And if a screen-share request comes up in a random chat with someone you don't know, treat it as an immediate red flag rather than something to consider.
Deepfakes are a real and growing part of how scams work now, but the core defense hasn't actually changed that much — skepticism toward urgent financial requests, a few specific physical tests you can ask for naturally in conversation, and a willingness to disengage the moment something feels off. None of it requires special software or technical expertise. It just requires knowing what to actually look for, which is the whole point of this page.
For more on staying safe in random video chat generally, see our guide on talking to strangers online safely, or read our honest look at whether anonymous chat is actually safe.