Practical guide · Hoox

The visual checklist — 12 AI UGC tells to avoid

With visual signs and correction prompts for each tell. The screenshot test 90% of creators skip.

Why 90% of AI UGC gets tagged in 1 second

The TikTok scroll lasts 0.8 seconds per video. If your AI avatar triggers the slightest uncanny valley, the user scrolls. Worse: they comment "AI slop" and kill your reach. The screenshot test is the brutal filter we apply to every Hoox generation before publishing. You grab 5 random frames, zoom to 200%, and hunt for the 12 tells below. If a single tell shows up, you rerun.

Core rule: a real human is never perfect. If your avatar is, you have a problem. The whole trick to AI UGC that passes is to reinject controlled imperfection.

The 12 visual tells

Tell 1 — Uniform studio lighting

Visual sign: face evenly lit 360°, zero hard shadows, no contrast on the side of the nose or under the chin. Looks like perfectly balanced key + fill + rim lights.

Fix prompt: natural window light from left, soft shadow on right cheek, slight underexposure under chin, single source lighting, iPhone front camera quality

Tell 2 — Skin too smooth / no pores

Visual sign: permanent Instagram filter mode. No visible pores on nose, no redness on cheeks, plastic texture under zoom.

Fix prompt: visible skin pores on nose and cheeks, slight redness around nostrils, micro-imperfections, natural skin texture, no beauty filter, raw iPhone camera

Tell 3 — Background too clean / staged

Visual sign: spotless white wall, perfectly arranged shelf, coffee placed dead center. Too "Pinterest mood board".

Fix prompt: messy real bedroom background, unmade bed corner visible, random objects scattered, slight clutter, lived-in space, depth of field blur

Tell 4 — Weird hand-eye coordination

Visual sign: hand moves 0.3s late after gaze, fingers brush the object without really holding it. The gesture floats.

Fix prompt: natural hand grip on product, fingers wrap around object, micro-adjustment movement, eyes track hand movement, hand enters frame from below

Tell 5 — Lip-sync micro-misalignment

Visual sign: lips close 1 frame too early on "p" and "b" sounds. At 0.5x speed, it's blatant.

Fix prompt: Regenerate with slower voice (-15% speed), pick an avatar trained on your target language. On Hoox, select avatars tagged FR-native or EN-native.

Tell 6 — Fixed/glassy eye gaze

Visual sign: eyes lock on camera with no micro-saccades. No natural blink, no gaze drifting 0.2s to the side.

Fix prompt: natural eye micro-movements, occasional glance to side, blink every 4 seconds, eyes slightly squint when smiling, subtle eye moisture reflection

Tell 7 — Hair edges too sharp (no frizz)

Visual sign: hair contour cut with a scalpel against the background. No flyaway hairs, no frizz. Too "after-effects rotoscope".

Fix prompt: natural flyaway hair strands, slight frizz around hairline, individual hairs catching light, soft hair edges, baby hairs visible

Tell 8 — Facial symmetry too perfect

Visual sign: smile strictly identical on left and right. Right eye exactly the same size as left, pixel-perfect.

Fix prompt: asymmetric smile, slightly raised left eyebrow, one eye slightly smaller, natural facial asymmetry, real human face proportions

Tell 9 — Clothing fabric too draped/perfect

Visual sign: wrinkle-free t-shirt, hoodie hanging like a shop mannequin. No creases, no wear marks.

Fix prompt: worn cotton t-shirt with natural creases, slight wrinkles around shoulders, fabric pulled at chest, lived-in clothing, casual fit not styled

Tell 10 — Inconsistent reflections

Visual sign: eyes don't reflect the window even though light comes from there. Glasses show no reflection. Phone held in hand shows no face reflection.

Fix prompt: catchlight in eyes from window source, glasses showing window reflection, phone screen reflecting face partially, consistent reflection physics

Tell 11 — Inconsistent shadow direction

Visual sign: nose shadow on left but neck shadow on right. Multiple contradictory light sources.

Fix prompt: single light source from upper left, consistent shadow direction across face neck and shoulders, no conflicting shadows, physically accurate lighting

Tell 12 — Missing audio room tone

Visual sign (audio): pure -∞ dB silence between words. No ambient noise, no reverb. To the ear, it's a studio voice-over, not an iPhone video.

Fix prompt: In post-prod, add a -42 dB room tone in continuous background (HVAC, distant traffic, or ambient bedroom). Tools: Audacity or CapCut > Audio FX > Room reverb 8%.

QA protocol — 90 seconds per video

Screenshot test 5 frames: pause at 1s, 3s, 5s, 7s, 9s. Zoom 200%. Hunt for tells 1, 2, 6, 7, 8, 10, 11.
Audio check 3 sec random: pick a silent zone between two words. If you hear pure -∞ dB, fail. Tell 12.
Lip-sync check 0.5x: play at 50% speed. Focus on "p", "b", "m". Tell 5.
Hand-eye scan: watch only the hands during one playback. Tell 4.
Background scan: watch only the decor. Tell 3.

Decision tree — tell detected → action

Tells 1, 6, 8, 10, 11 (lighting / face / physics) → rerun with modifier natural imperfect + new seed.
Tells 2, 7, 9 (skin / hair / fabric texture) → rerun with prompt enriched with imperfections (see fix prompts above).
Tell 3 (background) → change the environment in the prompt, or crop tighter on the face.
Tell 4 (hand-eye) → rewrite the scene so the hand enters from the bottom of the frame, more natural.
Tell 5 (lip-sync) → switch avatar (Hoox tag native), slow voice by 15%.
Tell 12 (audio) → fix in post-prod, no need to rerun.

On Hoox, QA is built in: our avatars are trained on real iPhone footage, natural lighting, preserved skin texture. You can push realism by adding imperfect natural in your custom prompts.

What now?

Print this checklist. Stick it next to your screen. Before every publish, run the 5 checks in 90 seconds. You'll cut your "AI slop" detection rate by 3x, and multiply your organic CTR. The screenshot test is what separates AI UGC creators who scale from those who get shadowbanned.