Why AI video looks fake, and how to fix it
When someone says AI video "looks fake", they are almost never reacting to resolution or frame rate. They are reacting to a short list of specific, fixable things. This is that list, with what to do about each.
1. It is too clean
The most common tell is that generated footage looks like it was shot by a competent crew with good lighting. Real phone video is not like that. It has blown highlights near windows, mixed color temperature, focus that hunts for a moment, and a horizon that is not quite level.
Fix: ask for the camera in the brief. "Shot on a phone held at chest height", "available light only", "slight handheld drift". Naming the camera and the light source does more than any amount of adjectives about realism.
2. Nobody is doing anything with their hands
Generated people have a tendency to stand and talk. Real people fidget, hold a coffee, gesture past the frame, adjust a strap, shift their weight from one foot to the other.
Fix: give every person in the brief a physical task. Not "she talks about the product" but "she talks while unlocking a bike, glancing down at the lock twice". The task also anchors the hands, which is where generation most often goes wrong when hands have nothing to do.
3. The speech is too well written
Written dialogue is smoother than spoken dialogue, and the smoothness is audible. People restart sentences, use the wrong word and correct it, trail off, and talk over each other.
Fix: write the disfluencies in. "I mean — honestly? Probably like an hour. No, more." Overlap two speakers deliberately, so one person's laugh starts before the other has finished. If you read your script aloud and it sounds like a script, it will sound like one in the ad.
4. The eyes are wrong
This is the subtlest and the most damaging. Generated faces look at the camera steadily. Real people in conversation look away constantly — up when recalling, sideways when uncomfortable, down before delivering something honest. Sustained unbroken eye contact reads as uncanny even when the face is otherwise perfect.
Fix: direct the gaze explicitly. "She looks up and to the left before answering, then back to the interviewer." A half-second of looking away is worth more than any improvement in facial detail.
5. The background is empty or static
A street with nobody on it. A café with no one moving. Background extras who walk a fixed path and never look at anything. Real environments are full of small, purposeless motion.
Fix: specify background activity that is incidental rather than choreographed. "People passing behind, none of them looking at the camera; a bus pulls away mid-shot." Incidental is the operative word — background action that is too relevant reads as staged.
6. The audio does not match the room
A voice recorded in a booth over footage of a busy pavement. This is one of the fastest ways to lose a viewer, and almost nobody notices they are doing it because they are listening for whether the voice sounds good rather than whether it sounds present.
Fix: the voice should carry the room. Traffic under a street interview, a little reverb in a kitchen, wind on a beach. If your pipeline produces a clean voice, put the ambience back in deliberately at the mix rather than leaving it dry.
The related error is audio bitrate. Video compression is forgiving; audio compression is not. Thin, hissy audio makes footage feel cheap in a way viewers attribute to the video. Encode audio well even when you compress the picture hard.
7. Continuity breaks between shots
The jacket changes shade. The mic swaps hands. A car in the background is in two places. These happen when an ad is assembled from separately generated shots rather than cut from continuous takes.
Fix: generate long takes and cut from them, the way an editor cuts real footage. Where you must join separately generated material, join on a change of angle or a moment of motion rather than on a static match, and keep the same reference images across every shot so the person is genuinely the same person.
8. The performance has no subtext
Everyone in generated video tends to mean exactly what they say. Real people are frequently saying one thing and feeling another — polite but bored, amused but pretending not to be, confident about something they are not sure of.
Fix: write the subtext into the direction. "He agrees, but he is not convinced, and it shows for a beat before he covers it." This single change does more for believability than most technical adjustments.
9. Everything is at the same pace
Generated performance tends toward an even tempo. Real speech has bursts and stalls — a fast run of words, then a long gap, then three words.
Fix: specify the rhythm. Mark where the speaker rushes and where they stop. In the cut, resist trimming every gap; the gaps are what make the rest feel fast.
10. The product is too prominent
The label faces the camera. The bottle is centered and lit. This is the visual grammar of a commercial, and it tells the viewer that everything preceding it was setup.
Fix: let the product be handled badly. Partly out of frame, held at an angle, put down mid-sentence. Real people do not present objects to cameras.
How to work through the list
Do not try to fix all ten in one brief. Take a take that is not working, watch it with the sound off, and identify which single item on this list is loudest. Fix that one, re-render, and watch again. In most cases two or three of these are doing all the damage, and the rest are fine.
The useful discipline is to grade every take against these criteria before spending on the next one. A take that fails on eye contact and background motion will not be rescued by a better color grade, and knowing that before you render the next variation is worth more than any single improvement in model quality.