Start with your eyes, then let several engines check what your eyes cannot see.
Most people ask the question in front of a single image: a profile picture, a news photo, a listing, a screenshot sent by someone else. Below are the things worth checking yourself in under a minute — and what each of them does not prove. Then you can run the same image through Uthentic, which reads it with several independent engines and tells you where the doubt sits instead of handing you one number.
Follow every shadow back to its source. In a real scene they all agree on where the light comes from and how hard it is.
Mixed lighting is normal indoors, and a reflector or a flash can create a second, softer direction.
A real object touches its shadow. Look for the dark contact point under feet, wheels and furniture, and for a shadow that softens as it travels.
Overcast light makes shadows very soft, so a faint shadow is not evidence on its own.
Windows, eyes, chrome, water and glasses should reflect what is actually around, in the right position and the right size.
Coated lenses and polarising filters legitimately kill reflections in real photographs.
Real texture keeps its irregularity at 100% zoom: pores, weave, stray hairs, paper grain. Generated texture often turns smooth, waxy or repetitive.
Beauty retouching, heavy noise reduction and small social-media files also destroy texture.
Text on signs, logos, jewellery clasps, teeth, keyboards and hands are where generation still breaks down.
Motion blur and shallow depth of field can mangle the same details honestly.
None of these signals is a verdict on its own. A single odd shadow means "look closer", not "generated" — which is exactly why a serious answer needs more than one reading.
These are two different questions, and most tools answer only the first one. A fully generated image has no camera behind it: no sensor pattern, no lens signature, no capture metadata, and a texture that is statistically too even across the whole frame.
A real photograph worked in post production keeps its camera fingerprint everywhere except in the area that was changed. Skin smoothing, a replaced sky, a removed passer-by or a generative fill leave a local discontinuity: compression and noise behave differently inside that patch than outside it. Uthentic reports that case as an authentic photograph with an edited area, instead of calling the whole image AI — the mistake that produces most false accusations.
The image goes through several independent engines rather than one model: open detectors trained on generated images, a vision model that reasons about the scene, file-level forensics on EXIF data and C2PA provenance, a local reading of skin micro-texture and how pigment responds to reflected light, plus a monochrome integrity map of the frame. Their outputs are calibrated and fused, so a single wrong reading cannot decide the verdict, and a separate arbitration step is used only when the signals genuinely conflict.
The report shows the suspicious areas on the photo itself, the signal behind each flag, and where light or context makes deception most likely. When the evidence is not strong enough in either direction, the report says exactly that instead of forcing a percentage. Your image is needed only while the analysis runs, is deleted afterwards, and is never used to train models.