Ask an AI video tool for a landscape and it will impress you. Ask it for a person and it will eventually embarrass you — usually at the hands, then at the face, then by forgetting who the person was. Here is the failure list, and how working creators shoot around it.
People are the hardest thing in generated video, for a simple reason: viewers are experts on people. We forgive a strange tree instantly; we spot a wrong thumb in half a second. This guide names the four ways generated people fail, explains plainly why, and gives the working rules — distance, duration and subject choice — that keep your videos on the safe side of the line. It ends with the most important rule of all for a creator channel: which faces should never be generated at all.
Watch any long generated clip of a person and you will meet them in order. Fingers come first — extra digits appear, merge, or vanish during any action involving the hands, because hands are small, fast and constantly changing shape. Faces soften next: eyes drift, teeth blur, skin turns to wax, most visibly in close-up and during speech. Objects in hands misbehave — cups change shape, tools bend. And identity quietly walks away: the person at the end of the clip is a cousin of the person at the start, and in the next generated clip, a stranger. None of this is one brand's shame; it is the current state of the whole technology, improving year by year and not yet solved.
A landscape is forgiving because it is texture without rules — a hillside has no correct number of trees. A hand has strict anatomy, a face has strict symmetry, and both move in ways the model must keep consistent across dozens of frames it generates in sequence. Each frame is a fresh prediction, and small prediction errors compound: a slightly wrong finger in frame ten becomes a very wrong hand by frame forty. This is also why shorter clips of people survive better — there is less time for errors to grow. Understanding this changes how you direct: you stop asking for the impossible shot and start asking for the shot the technology can actually hold.
Working creators treat these limits like a director treats a small budget — as framing rules:
A good editor hides a surprising amount: cutting a take before the artefact grows, covering a weak moment with a caption card or a cutaway, and keeping generated people on screen briefly all protect the viewer from the uncanny. But be honest about the boundary — editing cannot repair a melted face that sits on screen for five seconds, and no transition makes six fingers acceptable. The selection step is where quality is decided: reject flawed takes at review, not in the timeline. If the only take of a needed shot has broken hands, the correct fix is a different shot — wider, shorter, or from behind — not a hopeful edit.
There is one category where workarounds are the wrong answer entirely: the presenter. On a channel like this one, the viewer's trust is built on a real person who really did the work — real hands on a real keyboard, a real face explaining a real fix. Generated people are for scenes, crowds and illustrations; they are not a substitute for the host, and audiences can feel the difference even when they cannot name it. Use AI for the world around your video. Keep yourself, and anyone presented as a real person, genuinely real — the disclosure rules in this series exist for exactly this boundary.
See the whole process live on screen — every step, exactly as it is done in the video.
▶ Watch Full VideoBrief, still, well-lit close-ups sometimes survive — a face that barely moves for two seconds can hold together. The risk climbs with every added demand: speech, emotion, head turns, hand gestures. Treat close-ups as a gamble you verify frame by frame, never as a default shot.
Not practically, for a normal creator workflow. Repairing anatomy means painting or regenerating individual frames, which costs more time than the clip is worth. Cropping the hands out of frame, shortening the shot, or regenerating wider are all faster and look better.
Each generation starts fresh from your words alone — the model has no memory of the face it invented last time, only the description. Similar prompts give similar people, not the same person. Workflows that need one consistent character across a story are exactly where generated video is weakest today.