I was watching a political speech from 2020 and something about the mouth movements felt off, so I ran it through two different detection tools.

Elliot Mason · 2026-03-19T16:40:14.021222+00:00

One was a free online checker that just gave a 'likely fake' percentage. The other was a forensic analysis plugin for DaVinci Resolve that my buddy let me use. The free tool said 87% fake, but the plugin showed me exactly why: inconsistent lighting reflections in the eyes and a weird blur pattern on the teeth that didn't match the rest of the face. Seeing the specific artifacts made it click for me. Has anyone else found that just getting a 'score' is way less useful than actually seeing the breakdown of flaws?

elliot_roberts5mo ago

Remember seeing a fake where the mic didn't cast a shadow.

8

miasanchez4mo ago

The thing nobody's talking about is how these deepfakes mess with subpixel rendering. I noticed that on one video, when I zoomed in 400 percent on the shirt collar, the checkerboard pattern of pixels was actually shifting in time with the audio track. Like the generator was trying to sync a static image to talking, but the compression artifacts were breathing in and out with the voice. That's the kind of weird quirk that a simple percentage will never tell you.

3

emma_rodriguez305mo ago

The breakdown beats a vague score every time.

1

nathan_palmer5mo ago

Yeah, my buddy had a similar thing happen with a clip from a podcast. The basic checker just flagged it, but when he looked at the audio waveform, there was this weird, perfect gap in the background hum every time the host spoke. Seeing that little digital fingerprint made it real.

1