AI images have gotten scarily good, but “good” isn’t “perfect” — generators still make characteristic mistakes because of how they build images. Here’s what still gives them away.
The 7 tells
- Text in the image. Still the strongest tell. Signs, labels, jerseys — zoom in. AI text degrades into almost-letters, inconsistent fonts, or gibberish past the first word or two.
- Hands and fingers (in the background). Main subjects got fixed; background people didn’t. Count fingers on anyone standing behind the subject.
- Physics of light. Shadows pointing different directions, reflections that don’t match the scene, jewellery shining with no light source.
- Pattern continuity. Follow a repeating pattern — brick wall, fence, fabric print, hair strands — and watch it melt, merge, or restart mid-way.
- Too-perfect skin with hyper-detailed everything else. That waxy, poreless, evenly-lit face against sharp detailed clothing is a signature AI look.
- Background logic. Doors to nowhere, furniture fused into walls, crowds where faces repeat — generators care about the subject, and backgrounds expose it.
- The vibe test on symmetry. Glasses with mismatched arms, earrings that differ, collar points at different angles — humans and cameras produce symmetry; AI approximates it.
The 10-second settle
Reverse image search (Google Lens or TinEye) settles most cases: a real photo has a trail — original source, other angles, earlier posts. An AI image typically exists nowhere except where you found it. For news images, this habit matters more than any visual tell.
Why this matters beyond curiosity
Fake product photos, fake reviews with generated “customer” pics, romance-scam profiles, fake news events — AI images are now standard scam infrastructure. The same generators from our free AI tools guide are available to everyone, including people building fake shops. When a website’s imagery smells wrong, that’s a red flag we weigh in website legitimacy checks too.
Do AI-detector tools work?
Inconsistently — they produce both false alarms and misses, so treat them as one signal, not a verdict. The visual tells plus reverse search remain more reliable.
