How to Tell a Real Photo From an AI-Generated One
AI-generated photos are good enough now that your eyes cannot be trusted alone. No single check proves an image is fake, but a 60-second routine of five checks catches most fakes before you share them.

AI-generated photos are good enough now that your eyes cannot be trusted alone. A fake image can have perfect lighting, believable skin and a convincing background, and it can still be entirely invented. The good news is that most fakes still leave traces. Here is how to spot AI-generated images before you believe one or share one.
No single check proves an image is fake. The generators improve every month, and the tells that worked last year are already fading. What works is a routine: several quick checks, done together, that catch most fakes in under a minute. Think of it like checking a used car. One clean panel means nothing. Five clean checks start to mean something.
Start with the hands, the text and the details
Hands are still the classic tell. Count the fingers, look at how they bend, and check whether a hand holding a cup actually grips it. Teeth are another weak spot: too many, too even, or merging into each other. Earrings that do not match, glasses with one arm missing, a watch that melts into the wrist.
Text is the fastest check of all. AI image generators still struggle with letters. Read every sign, shirt, book spine and label in the frame. Gibberish text, duplicated words, or letters that morph halfway through a word are strong signs of generation. Real photos have real text, even when it is blurry.
Then look at the background. AI images often get the subject right and the background wrong: a window that opens onto nonsense, a crowd where faces blur into each other, a building with windows that do not line up. Zoom in. The background is where the generator stops paying attention.
Check the light, the shadows and the reflections
Light is physics, and generators still fake it badly. Look for shadows pointing in different directions, or a face lit from the left while the shadow falls to the left too. Check reflections in mirrors, windows and sunglasses: they often show a different scene from the one in front of the camera, or a person who is not there.
Skin is another giveaway. AI faces often have skin that is too smooth, too even, almost plastic, especially around the eyes and mouth. Real skin has pores, blemishes and uneven tone. Hair is similar: individual strands that end nowhere, or a hairline that dissolves into the forehead.
None of these alone is proof. A real photo can have odd lighting, and a good AI image can have perfect hands. The point is to stack the checks. One odd detail is a curiosity. Three odd details in one image is a verdict.
Use the metadata and the tools
The most reliable checks are not visual at all. Many cameras and phones now attach C2PA content credentials to photos, a signed record of where the image came from. If a photo carries credentials from a known camera or app, that is strong evidence it is real. If a photo claims to be from a news event and carries no credentials at all, be skeptical.
Reverse image search is the next step. Upload the image or paste its URL into a search engine and see where else it appears. A real news photo usually has a history: the same image on a wire service, a photographer's portfolio, an earlier article. An AI fake usually appears out of nowhere, attached to a fresh account or a single viral post.
Detection tools can help, but treat them as a second opinion, not a judge. Models like the ones in Pangram's detection tools are trained to spot generation artifacts, and they are right more often than the human eye. They are also wrong sometimes, in both directions. Use them to confirm a suspicion, not to replace your own checks. The same discipline applies to video, which has its own tells, covered in our guide on how to spot an AI-generated video.
How to spot AI-generated images in practice
Here is the 60-second routine. First, zoom in and read the text. Second, count the fingers and check the hands. Third, look at the background and the light. Fourth, reverse search the image. Fifth, check the source: who posted it, when was the account created, and does the image appear anywhere credible.
If the image passes all five, it is probably real, or at least real enough to treat as plausible. If it fails two or more, do not share it. The cost of sharing a fake is higher than the cost of waiting: fake images of disasters, protests and public figures have already caused real harm, and the share button is the distribution network.
The generators will keep improving, and the visual tells will keep fading. The routine will not save you forever. But the source check and the reverse search will age well, because they do not depend on the image at all. They depend on whether anyone can vouch for it. In a world where any image can be faked, provenance is the only check that cannot be faked along with the pixels.


