Guide

Can You Tell If an Image Is AI-Generated Just by Looking at It?

No. A strange hand, a broken word, or a reflection that does not match can make you look twice. Those clues are not a test. A convincing picture is not proof that a camera took it, and an odd picture is not proof that a generator made it.

Looking is a suspicion, not a verdict

People still try to settle this by eye. Cooke, Edwards, Barkoff, and Kelly asked people to sort real and synthetic media. For images, average accuracy was 49.4%, close to a coin toss. Nightingale and Farid tested real faces against faces from a generative model. Without training, people were right 48.2% of the time. A separate group, given training and feedback, reached 59.0%. That is better than guessing, and it is still not reliable.

Those studies used particular pictures, not every image you will meet. The percentages describe those studies. They are not Real Image Check’s accuracy. They still mark the limit. A careful look can raise a question. It cannot reliably answer whether the image is AI-generated.

Clues people notice, and why each one fails

None of these is proof in either direction. Each one is a reason to slow down.

• Hands. Extra or melted fingers are a familiar warning. A generator can also draw five ordinary fingers. A real photo can hide a finger, blur it, or crop it at the frame. Counting fingers is a reason to look closer, not a verdict.
• Text. Nonsense on a sign is suspicious. A real sign can be out of focus, in another language, or cut in half. A generator can also render a short, correct word. If the file is soft or tiny, the letters may not mean what they seem to mean.
• Reflections. Eyes, glasses, and windows sometimes show a scene that does not fit. A real mirror, a second light, or a window onto another room can do that too. A reflection that matches is not proof of a camera either.
• Shadows. Shadows that fall in two directions are worth a second look. A room can have a window and a lamp. A flash can light a face while the background keeps its own light. One shadow direction does not prove the picture is real, and two do not prove it is generated.
• Fine detail. Jewelry that melts into skin, or hair that turns into a smear, can be a clue. JPEG compression, a beauty filter, and a small preview do the same thing to a real photo. A clean close-up can still be generated.
• Repeated patterns. A wall, a fence, or a crowd can look copied. Real wallpaper and real fabric repeat on purpose. A generator can also vary a pattern enough that the repeat is hard to see. Compare one repeat with the next, then stop.
• Inconsistent objects. A strap that vanishes, a chair leg that joins the floor, or an ear that does not match the other ear is a reason to pause. Editing, a collage, and a bad crop create the same breaks in pictures a camera started. Objects that agree with each other are not proof either.

A clean picture can be generated, and a strange one can be real

A generator does not have to leave an obvious mistake. The face study above used pictures that already looked photographic, and people still could not sort them reliably. The other way around is just as common. Motion blur, a dark phone photo, a wide lens, and heavy editing make real photos look wrong.

If the method is only “it looks off,” both kinds of picture will fool you. A lack of mistakes is not evidence that a camera took the picture.

What the file does to what you can see

The file in front of you may already be a damaged view of some earlier picture.

• Resizing throws away detail.
• Compression, especially a low-quality JPEG, replaces fine edges with blocks and smear.
• A crop removes the context that would explain a cut-off hand or a half sign.
• An edit can add the oddness, or paint over it.
• A screenshot is a new picture of a screen, often smaller, and with the display’s own scaling.

Any of those changes makes visual inspection harder. None of them tells you which way the original leaned. For what a detector can estimate from a screenshot, a resized preview, or a compressed copy, see Can an AI image detector check a screenshot?

A checklist that is not a test

Use this to slow down. Do not add up the items, and do not treat a finished list as a yes or a no.

• Look at the whole picture, then zoom in on one area, then zoom back out.
• Check the same kind of clue twice: both hands, two bits of text, or the shadow and the light.
• Ask what a camera would do in that place: focus, motion, a window, a mirror, or a flash.
• Notice the file. Is it tiny, heavily compressed, cropped, edited, or a screenshot?
• Write down what you saw in one sentence. Stop. That sentence is a suspicion, not a result.

For a wider check that also asks where the image was published, read How to tell if an image is AI-generated.

Three different questions

Looking asks whether anything seems wrong. The answer is a suspicion.

Content Credentials ask whether the file carries a signed note about its history. The C2PA explainer describes those notes as signed statements bound to the file. Checking one tells you whether the note is intact and whether the signer is on a trust list. It does not decide that the statement is true. A crop or another edit in a tool that does not keep the note can leave that history incomplete. A missing credential does not mean the image is a real photo. Many files never had one.

A detector asks whether the file resembles AI-generated imagery. That is a third question. It is not a sharper pair of eyes, and it is not a provenance check.

What Real Image Check estimates

The checker scores one JPEG, PNG, or WebP image, up to 20 MB. The score is calculated in your browser. The image is not uploaded for analysis. This browser keeps the preview, file name, and score until you clear that check or leave the page. Those details are not sent to the server. The server stores an operational event for the check, such as whether it finished, the plan, a short error code, the model version, how long it took, and a shortened visitor identifier. That event does not include the image, the preview, the file name, or the score.

The score runs from 0.000 to 1.000. At 0.65 or above, the label is Likely AI-generated. Below 0.65, the label is Likely a real photo. Scores on the same side of 0.65 share that label. A result of 0.66 and a result of 0.99 receive the same label. The score is an estimate, not a probability or proof.

The checker does not establish who made the image, prove the image is authentic, name a generator, or read a watermark or Content Credentials. It returns one score for the whole image. It does not point at the hand or the letter that looked wrong. A picture that looks strange can still be labeled likely a real photo. A picture that looks clean can still be labeled likely AI-generated. For what it means when that label is the wrong one, see Can AI image detectors get it wrong?

Use the score as a separate estimate. If the result would change what you publish, pay, or believe, go back to where the image came from.

Questions