AI Image Detector — Free AI Image Checker

Add it to the box at the top of this page. The tool reads the file's Content Credentials and AI labels first and looks for the hidden ChatGPT and Gemini watermark. If none of those settles it, an AI model on your device estimates how likely the image is AI-generated and draws a heatmap of where it saw AI patterns.

Yes, in both directions. In our October 2026 test 10 of 69 real photos landed between 25% and 75%, and 3 of them scored 50% or more. Of 114 AI images, 4 got past both the file check and the model. Photos re-sent through chat apps are where it slips most often, so a percentage on its own shouldn't be used to accuse anyone of using AI.

Sometimes. Google and OpenAI hide an invisible SynthID watermark in images their AI makes, and this tool looks for the fixed pattern it leaves. In our tests it found the pattern in all 19 original ChatGPT images and in 73 of 76 re-saved copies, while for Gemini it worked only on untouched files: 19 of 20 originals and 0 of 60 re-saved copies. Google's SynthID Detector and OpenAI's checker read the watermark itself, though each of them uploads your image to its company's servers.

No. The file check, the watermark check and the AI model all run inside this browser tab, and the only network requests fetch the model and its helper files from this site. Open your browser's network tab during a check and you'll see nothing carrying your image away.

Only that the model found little sign of AI in the image's structure. No detector can prove a photo is real, so even a low result leaves room for an AI image the model doesn't recognize.

Only when the file says so. If its Content Credentials or generation settings name the tool, you'll see that name. We don't guess it from how the image looks, because the best-known detector in our tests named the wrong generator for 3 of 8 AI images.

A signed label that some cameras and AI tools write into an image file, also known as C2PA. A valid signature from a signer on the official trust list proves which company signed the file and that it hasn't changed since. Re-saving, resizing and screenshots usually strip it, which is why a missing label tells you nothing either way about where the image came from.

Because the file and the pixels can disagree. A photo can carry camera data while the model still sees signs of AI in it, and we show both and change neither. Camera data is easy to edit, and detectors misjudge some real photos.

Only when the file says so. Some phones and editing apps record an AI edit in the photo's Content Credentials, and then you'll see “AI-edited”. We tested browser models that mark edited areas, and the best one found about 9% of the edits while leaving 95% of untouched photos unmarked, so we don't mark edits. The heatmap answers a narrower question: it shows where the model reacted on a resized copy of the image. That doesn't prove those areas were made or edited with AI, and on a real photo the model gets wrong it glows just as brightly.

The AI model (about 126 MB) downloads once and stays in your browser, so later checks start right away.

No. This tool checks still images, one at a time. For a video, take a screenshot of a frame and check that, though a screenshot loses the file's labels.

Yes, with no sign-up and no daily limit. The checks run on your device, so there's no per-image server cost to pass on to you.

Not on its own. An estimate is a probability from a model that makes mistakes, so don't use it alone to accuse anyone. A Content Credential from a trusted signer is stronger evidence, because a known company signed it.

Yes, with less to go on. Screenshots and chat-app copies lose the file's labels, so only the watermark check and the estimate are left. On copies of AI images resized and recompressed the way chat apps do it (1600 px, JPEG quality 70), the model caught 79 of 114.