Guide

How to Tell If an Image Is AI-Generated

Wondering whether an image was made by AI? Look for visual clues, check where the image came from, and use an AI detector for another perspective. No single clue or detector score can tell you for certain.

Look for visual mistakes

Zoom in before you decide. Generators still leave mistakes, and a real photo can look odd for ordinary reasons.

• Hands. Count fingers, wrists, and where a hand meets a sleeve or another person. Extra, missing, or melted digits are a common tell.
• Text. Read signs, labels, shirts, and book spines. Letters that almost form words, or a logo that changes shape, are worth a second look.
• Reflections. Compare eyes, glasses, windows, and water. Reflections that show a different scene, or catchlights that do not match, are clues.
• Lighting. See whether shadows fall in one direction and whether the face is lit like the room behind it.
• Merged objects. Look for buildings, furniture, jewelry, or limbs that blend into each other or into the background.

These clues are not definitive. Newer generators make fewer of these mistakes, and heavy editing, beauty filters, or a small preview can create the same kinds of oddness in a real photo. A clean-looking picture is also not proof that a camera took it. For why a visual check still cannot settle the question, see Can you tell if an image is AI-generated just by looking at it?

Find the original source

Ask where the image was first published. Reverse image search, in Google or another search engine, can show earlier copies, a photographer’s post, a news article, or a generator page.

Open the earliest credible copy you can find and compare it with the version you have. A cropped, captioned, or re-uploaded file can drop the context that explained it. Finding no earlier copy does not prove the image is synthetic. Finding an earlier camera original is stronger evidence than a visual hunch.

Watermarks and Content Credentials

Some generators add a visible mark. Others embed an invisible watermark, or attach Content Credentials, a signed note about the tool that created or edited the file. When that signature is still intact, it is evidence about the file’s history.

Those signals are easy to lose. A screenshot, a crop, a re-save, or a social platform can strip a watermark or the credential. A missing mark does not mean the image is real. A mark from one company also says nothing about images made by a different tool.

Real Image Check does not read watermarks or Content Credentials. If you need to know whether OpenAI’s own provenance signal is present, use OpenAI’s verifier. That check reports OpenAI signals only. It does not decide whether some other generator made the picture.

What an AI image detector can tell you

A detector looks for patterns in the pixels. It does not know who posted the image, and it does not understand the event the way a person does. The number it returns is a model’s estimate.

Use that estimate after the checks above, not instead of them. Screenshots, compression, resizing, edits, and generators a model has not seen can all change the result. That is a limit of this kind of check. Real Image Check has not published an accuracy study, and this page does not borrow one from anywhere else.

How Real Image Check scores an image

The checker scores one JPEG, PNG, or WebP image, up to 20 MB. The score is calculated in your browser. Real Image Check does not upload or store the image, a preview, its file name, or the detection score.

The first analysis loads the detection model in the browser. Later checks in that visit reuse it. The image still stays on your device. Recent checks remain in this browser until you clear them or leave the page.

The checker does not name a generator, highlight regions, or read a watermark. It checks one image at a time.

The label uses 0.65 as its cutoff:

• 0.65 or above: Likely AI-generated.
• Below 0.65: Likely a real photo.

The score is an estimate, not a probability or proof. A result of 0.66 and a result of 0.99 receive the same label. A result of 0.64 is labeled likely a real photo, and it can still be wrong.

When a score is weaker evidence

Treat the number as weaker when the file is no longer close to the original.

• Screenshots. You can save a screenshot as a JPEG, PNG, or WebP and check it. The screenshot is a new picture of the picture, so the score describes that copy.
• Compression and resizing. Small previews, repeated saves, and messaging apps can wipe out the patterns a detector uses.
• Edits. Filters, crops, and mixed real and generated areas can pull a score either way.
• Unfamiliar generators. A model can miss a generator it was not built to recognize, or flag a real photo that happens to share some of those patterns.

If the decision matters, keep the original file, check the source, and do not treat one score as the whole answer. For a screenshot, a compressed copy, or an edit, see Can an AI image detector check a screenshot?

Check an image in your browser

When you want a second opinion, open the checker, add one JPEG, PNG, or WebP, and read the label beside the image. Then go back to the source if the result would change what you publish, pay, or believe.

Questions