That Video of Your Boss Asking for Money Won't Glitch Anymore

For years, the advice was simple. Watch the blinking. Watch the hands. If a video of someone asking you for money looked jumpy or warped, you knew it wasn't real. That tell is disappearing. Researchers from The Chinese University of Hong Kong, working with Alibaba's Qwen team and Liblib A.I., built a system that generates full-body video of a person, continuously, at about twenty frames per second. Twenty frames per second is the point where your eye stops noticing the seams.

Why should that matter to you, sitting in your car

So why should that matter to you, sitting in your car or your kitchen right now? Because the next video someone sends you — your boss, your daughter, your bank — might look completely fluid and completely fake. And I want to be honest with you. That's an unsettling thing to sit with. If it makes you uneasy, that's a reasonable reaction, not a paranoid one. But feeling powerless is the part we can fix today. I'm going to walk you through what actually changed inside these systems, why the old glitches vanished, and the one habit that still protects you. So how does a machine make a fake body move smoothly for minutes on end?

Start with why old fakes fell apart. Older systems generated video in short bursts, then stitched them together. Every stitch was a chance for the face to freeze or the hand to melt. The article uses an analogy I keep coming back to. Old deepfakes were like a photocopy of a photocopy. Each pass lost detail and added noise, and investigators learned to read those seams. The new approach is more like a digital file. Nothing degrades, because nothing's being copied.

The trick is memory. This system, called LiveAnimate, keeps one permanent reference pose of the person. Then it stores about five representative body and hand positions from earlier in the video. When it draws a new frame, it looks back at how that body already moved and stays consistent with it. That's why it doesn't drift. And because it only keeps five snapshots, the cost doesn't balloon as the video gets longer. For an investigator, that means frame-by-frame comparison stops being the smoking gun. For you, it means the thing your gut used to catch — that flicker of wrongness — may simply not be there.

You'd think detection software would just handle this

Now, you'd think detection software would just handle this. On paper, it looks like it does. Detection tools often record accuracy above ninety-five percent on test datasets. In the real world, they perform meaningfully worse. Why the gap? Because a lab video is pristine. A real video gets compressed by a messaging app, re-uploaded, screen-recorded, and shrunk. All that squeezing erases the tiny fingerprints the detector was trained to find. And the generators keep changing, so a detector trained on last year's fakes gets fooled by this year's.

The volume is the part that stopped me cold. Deepfake content went from roughly five hundred thousand pieces in 2023 to about eight million in 2025. That's sixteen times more in two years. And projections put A.I.-driven fraud losses in the United States at around forty billion dollars by 2027. That's not a curve humans can keep up with by squinting harder at videos.

Which brings us to the belief most of us still carry. We assume that if a video looks smooth and real, it probably is — and if it's fake, an expert would spot it instantly. That belief made sense for a long time. Making convincing synthetic video genuinely required a research lab and enormous processing time. So realism really did equal authenticity, most of the time. That barrier collapsed. A polished video no longer tells you someone's honest. It only tells you they had access to modern tools.