Clues are not equal
Rank clues by how hard they are to produce accidentally. A verified content credential, or motion that violates physics across several frames, is hard to explain away and should shift your confidence a lot. A single odd hand, a smeared hair edge, or a missing metadata block is easy to produce accidentally and should shift it only a little. When a strong clue and several weak clues disagree, the strong clue usually wins — but say so explicitly rather than averaging silently.
Common false alarms and why they happen
- Heavy compression and low light smear edges and blur fine detail, mimicking texture artifacts in genuine photos.
- Aggressive beauty or noise-reduction filters smooth skin and hair into the waxy look often read as AI texture.
- Wide-angle and fisheye lenses bend hands and faces, producing proportions that look anatomically wrong.
- Stylized art, illustration, and heavy color grading break lighting and shadow logic on purpose.
- Screenshots and re-saved files lose metadata, so a missing EXIF block is common in authentic media.
- Motion blur and rolling shutter in real video create frame-to-frame differences that resemble temporal flicker.
Why these traps matter
Each false alarm is a real artifact with an innocent cause. That is exactly why no single clue is conclusive: the same visible feature can come from generation or from ordinary capture and processing. The defense is not to find a better single clue but to require several independent clues pointing the same way before raising confidence.