Guide

How to spot an AI generated photo

Six fingers and crooked smiles are history: today's generators fail elsewhere. People who shoot for a living do not look for curious details, they look at the physical consistency of the image. These five checks are the same ones Uthentic applies systematically, and you can run them by eye with no tools at all.

Leggi questa guida in italiano

Light tells the truth

A real photograph tells a single story about light: one dominant source, shadows consistent in direction and hardness, highlights falling where they should. Generated images often mix several stories: a face lit from the right with a ground shadow going left, eyes with catchlights in different positions, soft shadows on one surface and hard ones on the next. Cover half the image with your hand and ask where the light comes from: if the answer changes from one half to the other, that is a signal.

Optics leave a signature

Every lens compresses space in its own way. A long lens portrait has a compressed background and a gradual fall-off; a wide angle stretches the edges. In generated images the blur is often applied like a mask: it jumps from sharp to blurred, cuts hair in half, forgets an object sitting on the same plane as the subject. Look at where the blur starts and ends, not at how pretty it is.

Materials and repeated detail

Skin, fabric, metal and hair react to light in different, non repeatable ways. Generators tend to produce pores that are too regular, fabric weaves repeating identically, metal highlights without scratches, teeth and nails that are too uniform. Lettering is another good clue: small text, signs, logos and numbers are often plausible from afar and meaningless up close.

File data, when it is there

A captured photo carries the camera's data: body, lens, shutter speed, aperture, ISO, date. Its absence proves nothing — many social platforms strip it — but its inconsistency does: an aperture incompatible with the depth of field you see, a focal length that does not match the perspective. Some tools instead write a provenance trail (C2PA) declaring how the file was created and edited.

Real does not mean untouched

This is the distinction almost everyone skips. Between the captured photograph and the fully generated image sits the authentic photo that has been edited: a replaced sky, a removed person, an extended background. The signs are local — edges that do not line up, noise changing within a single area, textures repeating in a corner — and should be read as editing, not as generation.

When the eye is not enough

Easy cases are solved by eye. The ones that matter — a well retouched photo, a recompressed screenshot, an image generated and then photographed off a screen — need measurements. Uthentic queries independent engines, fuses their readings after calibration and points out the areas and lighting conditions where deception is most likely. Every change to the fusion goes through a repeatable benchmark of 10,179 tests, including a dedicated category of real photographs worked in post production.

No tool, ours included, replaces judgement: a report exists so you know which evidence the decision rests on. When the evidence is not enough, the report says so.

Analyse a photo How the detector works