Is This Photo AI-Generated? A 6-Step Check Anyone Can Do

Generation models spend their quality on the centre of the frame. Zoom to 300 percent and inspect the perimeter of the subject: ear structure, earring pairs, the hairline where it meets a shoulder or background, the arms of glasses, the seam of a collar, the join between wrist and watch. Fingers are still worth checking, but hands have improved, edges haven't. Anything that smears, duplicates, or terminates in nothing is a strong signal.

Real rooms obey geometry. Follow every straight line, door frames, tiles, window mullions, table edges, and check that it stays straight and consistent where it passes behind the subject. Look for a doorway that changes height across the frame, tiles that shift size mid-floor, a chair with three legs, a reflection that doesn't contain the subject. Then read any text in the shot: signage, book spines, logos, street names. Pseudo-text that looks like letters until you try to read it is close to conclusive.

Pick the brightest highlight on the face and decide where the light source must be. Now check the shadow under the nose, the shadow the body casts on the wall, and the catchlights in both eyes. In a real photo they all point back to the same source or sources. Generated images often have flawless skin lighting with a shadow falling the wrong way, or eyes carrying reflections of a window that isn't in the scene.

Save the original file, not a screenshot, and inspect its metadata with any free EXIF viewer. A camera photo usually carries a make and model, a lens, exposure settings, and a capture date. Generated images typically carry no camera data at all, sometimes an editing-software tag, and increasingly a provenance marker (C2PA content credentials) that names the generator. Metadata is weak on its own, because social platforms strip it when they re-encode uploads, but a file sent directly to you with no camera data and a same-day creation timestamp is worth noting.

Run the photo through more than one reverse image engine, and crop tightly to the face for a second pass, cropping defeats mirroring and border tricks. Then read the result properly. Matches on stock sites, an unrelated person's real account, or a scam-warning forum means a stolen photo. Zero matches does not mean real: a generated face has no siblings online. Zero matches only moves you from "recycled photo" to "no history", which is what step 6 is for.

A real person leaves a trail across time and platforms: an account older than a few months, photos with other people in them, tags from friends, comments from people with their own history, images where they look younger or heavier or differently styled. A fabricated identity has none of this depth. The final and most decisive step costs nothing: ask for a photo taken right now, in a pose you choose, holding an ordinary object in an odd way. Generation cannot follow an arbitrary instruction at speed, and a person who refuses every version of that request has told you what you needed to know.