# How to make AI UGC not look like AI

**The exact post-production pass I run on client ads. Free.**

Most AI UGC gets clocked as fake in about a second. Not because of the model or the prompt, but because of five fingerprints AI leaves on every clip, and one finishing pass that removes them.

Here is the pass, in the order I run it.

## The 5 tells (and the fix for each)

**1. Color sits too "digital."**
AI renders clean, even, slightly waxy color. Real phone footage has compressed highlights and a warmer, less-perfect curve.
Fix: pull the highlights down (this kills the plastic sheen on skin), add a touch of contrast and saturation, nudge it warmer. Small moves. You are de-glossing, not stylizing.

**2. Skin is plastic, zero texture.**
The number-one giveaway. AI skin has no pores, no grain, no sensor noise.
Fix: add fine film grain (subtle, and temporal so it moves frame to frame), then a light sharpen on top. This breaks the smooth AI gradient and reads as recorded, not rendered.

**3. The motion floats.**
AI shots sit too still or glide too smoothly. No human is holding the phone.
Fix: add micro-handheld. A tiny, slow drift on the frame (a few pixels of movement, plus about 4% scale so you have room to move it). Barely perceptible, but the brain reads that a person filmed this.

**4. It is too clean for a phone.**
Real phone lenses have edge softness, a hint of chromatic aberration, and light vignetting.
Fix: a touch of chromatic aberration and a gentle vignette. Subtle. You are mimicking a $600 phone, not a broken lens.

**5. The audio was recorded in a vacuum.**
AI and TTS voice is too clean and sits in dead silence with no room. This is the tell almost nobody fixes.
Fix: normalize loudness to social spec (around -14 LUFS), warm the EQ, de-ess the harsh highs, and add a subtle room reflection plus an ambient noise floor so the voice sounds recorded in a real space. If the voice cadence itself is robotic, post will not fully fix it, re-voice it.

## The order (it matters)
Grade, then grain, then micro-motion, then lens (chromatic aberration and vignette), then audio, then a final loudness normalize.

## The actual secret
Restraint. Every one of these done too hard looks like a filter, which is its own tell. Done too light and the plastic stays. The dialed-in amount per clip is the whole craft, and it is different for a bright skincare shot than for a moody night ad.

That is the pass. Free to run locally with ffmpeg or any editor that has grade, grain, and audio tools. The 30 minutes of taste is the part that separates AI slop from a clip people assume a real creator filmed.

---

Built by Phil Grabowski. If you would rather not dial this in per clip, it is exactly what I do for brands. Send me one ad and I will show you the before and after.