Reverse search, not fingers: how to tell if an image is AI

Counting fingers is dead advice. Current image models get hands right most of the time, write legible text on signs, and handle reflections well enough to survive a glance. If six fingers is still your test, you are going to miss nearly everything.

The useful move is to stop looking at the image and start checking where it came from. Four steps, about a minute, and you run them when a picture makes you feel something strong, because that is the feeling the picture was built for.

Who posted it first. Not who you saw it from. A real photo of a real event almost always has a traceable source: an outlet, a photographer, a named person who was standing there. Fakes tend to surface on accounts that post a great deal of viral material and nothing else. If nobody claims to have taken it, that on its own is most of your answer.

Reverse image search. Google Lens, Bing Visual Search, TinEye. You are looking for three things: older copies, which turn yesterday's storm photo into a 2023 photo; a fact-check, because somebody has often already done this; and the original, because plenty of viral images are real photos that have been edited.

Content Credentials. More cameras, phones and editing tools now attach provenance data under a standard called C2PA, and some platforms surface it as a small label. You can check an image yourself at contentcredentials.org/verify. The honest caveat is that absence proves nothing at all, because screenshots and re-uploads strip it constantly. When it is present it is one of the strongest signals available. When it is missing you have learned nothing.

Does the story hold. Would there be other photographs of this if it happened? A public event, a disaster, a protest, all of those produce dozens of angles, and one lone perfect frame is suspicious on arithmetic alone. Is the image too good: perfect light, perfect composition, perfect emotional beat, in the middle of chaos. And is it confirming precisely what some group already wants to be true, because fakes are built to be shared and outrage travels furthest.

Pixel clues still have a small role, as a second opinion rather than a verdict. Models spend their effort on the subject, so the background is where things still go wrong: faces in a crowd, text on a distant sign, railings that stop making sense. Physics is the other one, shadows pointing different directions, reflections that do not match. And there is a waxy over-polished finish to a lot of generated skin, which is a feeling rather than evidence.

Detector sites are worth less than people want them to be. Some are decent, none are reliable, and they fail in both directions: real photographs flagged, obvious fakes passed. Treat a score as one opinion. The source checks beat it.

The rule that covers everything: if you cannot work out where an image came from, do not pass it on as real. You can still pass it on with "no idea if this is genuine", which is one of the more useful things anybody adds to the internet at the moment.