How to Tell if an Image Is AI Generated

The most reliable question is not what the picture looks like but who published it first. Run a reverse image search and look for an earlier upload in a different context. A picture that appears in a news article, a company page and a stock library has a traceable origin; one that only exists as a screenshot in a group chat does not.

This step settles more cases than any visual analysis, and it takes less than a minute. It is also the only step that works when the image is genuinely ambiguous.

Signs, screens, book spines, jerseys and packaging all carry words. Generators have improved at short labels, but longer text at an angle still tends to dissolve into plausible-looking shapes. Ask whether the word you see would exist in that place at all.

Be careful with resolution. A small sign in a 600-pixel-wide photo is unreadable whether it was drawn by a camera or a model, so unreadability only counts if the sign should have been legible.

Where a hand holds an object, two different things have to line up: anatomy and physics. That is where generated images still slip, with fingers that merge into a cup handle, a thumb on the wrong side, or a strap that passes through a shoulder.

Real hands have their own problems — motion blur, awkward crops, bad angles — so look for structural impossibility rather than for one strange finger.

Pick a shadow in the foreground and a shadow in the background. Both should be consistent with one light source or a small number of plausible ones. Generators handle a single portrait light well and multi-source scenes badly, which is why interiors and night streets are more revealing than sunsets.

Reflections are the same test in a different medium: windows, water, sunglasses and car paint must show what is actually in front of them.

Count the windows on a facade, the slats on a fence, the chairs around a table. Equal spacing that changes angle, identical faces in a crowd, or a railing that continues through a wall are structural errors.

Sorting the scene by depth helps: what is in front, what is behind, and does anything belong to both?

Metadata, camera information and content credentials prove something when they are present and intact: an image with a verifiable signature from a camera or a generator has a documented origin. Their absence proves nothing at all, because every social platform strips metadata on upload.

That asymmetry is the reason detector scores should never be the last word. A detector returns a probability based on the models it was trained on, and it loses accuracy as soon as the image is resized or recompressed.