AI image generators got frighteningly good. The old tells — melted hands, gibberish text — are mostly fixed. Here is what still works in 2026, from quick eyeballing to tools that actually measure.
The eyeball checklist
Light and shadows. AI struggles with consistent light. Check whether shadows fall in the same direction and whether reflections in eyes, glasses, water and mirrors match the scene.
Texture and detail. Generators love over-smooth skin, wax-like surfaces and improbably perfect hair strands. Zoom into edges: jewellery, teeth, fences and bricks often dissolve into mush.
Text and logos. Signs, book spines, clothing prints and brand logos still come out subtly wrong — wrong letterforms, warped kerning, half-imagined logos.
Physics and context. Six fingers is fixed, but cutlery that intersects plates, cups floating above saucers, and crowd scenes where faces blur into each other are still common.
Provenance. Where did the image come from? A random Telegram forward is not a news agency photo. No original source, no EXIF data, first appearance on a fresh account — be suspicious.
What machines catch that you can't
Your eyes read pixels; detectors read statistics:
- Frequency and noise fingerprints. Diffusion models leave characteristic spectral patterns. Error-level analysis (ELA) exposes regions that were generated or spliced rather than captured by one sensor.
- Classifier ensembles. Models trained on outputs from Midjourney, DALL·E, Stable Diffusion, Flux and video-frame extractors score how likely the whole image is machine-made.
- C2PA content credentials. More cameras and apps cryptographically sign images at creation. A valid signature is strong evidence of authenticity — and its absence on a "professional" shot is a flag.
The honest caveat
No method is perfect. Heavy compression, cropping and screenshots strip the signals detectors rely on, and every generator generation gets harder to catch. That is why guessing a single verdict from one weak signal is a mistake — you want multiple independent checks weighted together, and an honest "inconclusive" when confidence is low.
Check it in seconds
Upload the image to RealCap and get a REAL or CAP verdict from multiple detection engines plus classic forensics — free, three checks a day. Pro adds full forensic reports with ELA heatmaps.