Only 0.1% of People Can Spot a Deepfake — Here's the 3-Step Method That Actually Works

Out of two thousand people tested in the U.K. and the U.S., only zero point one percent could correctly tell every real video from a deepfake. That's two people out of two thousand. The other ninety-nine point nine percent got fooled at least once.

That number should sit with you for a second,

That number should sit with you for a second, because it means almost nobody — not you, not me, not the sharpest person you know — can reliably eyeball the difference anymore. And this isn't just about being tricked by a funny video online. In Hong Kong, a finance worker at a multinational company sat in a video call with someone who looked and sounded exactly like the company's C.F.O. That deepfake was convincing enough to authorize thirty-nine million Australian dollars in payments — to criminals. If you've ever FaceTimed a coworker, joined a Zoom meeting, or even just watched a video someone texted you, this matters to you. If that feels unsettling, it should. But the reason most people get fooled isn't that they're careless. It's that they're using a detection method that stopped working years ago. So what went wrong with the old method, and what actually works now?

For years, people were taught to spot deepfakes by looking for visual glitches. Weird lip sync. Eyes that don't blink. Skin that shimmers or blurs around the jawline. And honestly, that advice used to be solid. Between roughly 2018 and 2021, the technology was young enough that those artifacts showed up consistently. Media outlets ran "spot the fake" challenges. Awareness campaigns told everyone to hunt for glitches. So people trained their eyes on exactly those cues — and felt confident doing it.

The problem is that deepfake generators have been specifically engineered to eliminate those exact tells. According to researchers cited by Which?, deepfakes get more convincing every year. The technology still struggles with hands, and lip sync can still slip in some cases. But replicating a face in a video call? That's gotten dramatically easier. Which means the checklist most people carry in their heads — look for the glitch, zoom in on the eyes, check the mouth — is tuned to catch last generation's fakes. Today's fakes are built to sail right past those checks.

That creates a dangerous irony

And that creates a dangerous irony. A video that looks perfectly smooth, with flawless skin and clean audio sync, might actually be more suspicious than a noisy, grainy clip shot on someone's phone. Real footage is messy. Real lighting is uneven. Real people fidget and blur. Ultra-clean perfection can itself be a red flag — because professional synthesis tools now produce output that's smoother than reality.

So if visual inspection alone won't cut it, what does? Researchers and forensic practitioners point to a three-checkpoint approach. Not a gut check. A structured method.

The first checkpoint is source verification. Before you even analyze a single frame, trace where the video came from. Who posted it first? Was it an established, verified account, or an anonymous one? Can you find the original upload, or only copies? According to the Which? investigation, the strongest signal for spotting a deepfake isn't what's in the video — it's the publishing chain around it. Source verification gives you documentary evidence. Visual judgment gives you an opinion. For an investigator building a case, that distinction is everything. For the rest of us, it means the next time a shocking video lands in your group chat, the first question isn't "does it look real?" It's "where did this actually come from?"

The second checkpoint is frame-by-frame consistency

The second checkpoint is frame-by-frame consistency. This is different from staring at a single frozen image. You're watching how things change over time. Does a facial feature flicker or vanish for a single frame? Does the person's skin tone shift slightly midway through? Does their hair color or eye color drift even a tiny amount across the clip? Synthesis algorithms still struggle with continuity across long sequences. A single frame might look perfect, but the transitions between frames can betray the forgery. Forensic examiners in the A.C.E.-V workflow — that stands for Analysis, Comparison, Evaluation, and Verification — treat this temporal comparison as a core discipline, not an afterthought.

The third checkpoint is contextual cross-checking. Does what this person is saying match what you'd expect from them? Is the situation plausible? Is there independent confirmation from a second source? That Hong Kong case is a perfect example of what happens when this step gets skipped. The finance worker saw a face that looked right and heard a voice that sounded right. But if someone had verified the request through a separate channel — a phone call, an email, a text to a known number — thirty-nine million dollars would have stayed in the company's account.

Now, you might assume that at least the algorithms can catch what our eyes miss. And facial recognition systems are powerful. But even they have hard limits. According to research analyzed by C.S.I.S., when a facial matching algorithm is set to a ninety-nine percent confidence threshold, its miss rate jumps to thirty-five percent. That means the system found the right person but wasn't confident enough to flag the match. A high match score doesn't eliminate false positives, especially in large-scale searches. For a professional running a database query, that means the algorithm is a starting point, never a conclusion. For everyday people, it means the technology protecting your accounts and verifying your identity has real blind spots — and those blind spots are measurable.