Check whether an image or video shows signs of AI generation or manipulation. The detector reads Content Credentials (C2PA), metadata, compression traces, sensor noise, spectral patterns and — for video — frame-to-frame coherence, and runs a trained detection model locally. Everything happens in your browser: a file you drop is never uploaded. The answer is always a probability with an interval and a confidence level, never a flat yes or no — because no analysis can prove where an image comes from.
Analysez une image ou une vidéo
Glissez-déposez, ou cliquez pour choisir un fichier
8 images · 40 Mo1 vidéo · 300 Mo
Rien n’est envoyé : l’analyse s’exécute dans votre navigateur, un fichier déposé ne quitte pas cet appareil. Pour un lien, l’adresse seule peut passer par notre serveur si le site refuse l’accès direct.
How to check an image in 3 steps
Drop the file — the original if possible
Drag in an image or video, or paste a link. The original file gives the most reliable result: screenshots and messaging copies have already lost most of the evidence.
Let the analysis run on your device
Each signal — provenance, metadata, compression, noise, spectrum, and a locally-run detection model — is checked in your browser and shown separately, with its own status.
Read the verdict with its reserves
You get a probability, its interval and a confidence level, plus a plain instruction. When signals are missing, the answer is “uncertain” — an honest answer, not a failure. A PDF report carries the same reserves.
Why use this AI image detector
- The engine runs inside your browser: no upload, no storage, no signup. For a link, only the address may pass through our server — never your files.
- Declared provenance, metadata, compression, noise, spectrum, temporal coherence — plus a trained model (Community Forensics, CVPR 2025) executed locally. Each signal is displayed on its own.
- Always a probability, an interval and a confidence level — never “100% AI”. When evidence is thin, the tool answers “uncertain” instead of inventing a certainty.
- The full analysis exports as a PDF that keeps every reserve and every limit — so a result is never quoted without what it does not prove.
Frequently asked questions
Can it prove an image was made by AI?
- No analysis can prove the origin of a file — this one included. The result is an estimate built from technical clues: a high score does not establish fraud, and a low score does not guarantee authenticity.
Why is the result often “uncertain” for WhatsApp or Facebook images?
- Those platforms recompress every image, and the fine traces the analysis relies on disappear at the first share. “Uncertain” is then the honest answer, not a defect. For a useful result, provide the original file — never a screenshot or a messaging copy.
Does missing metadata mean the image is fake?
- No. Almost every image in circulation has lost its metadata — that is the normal case, not a red flag. Absence never raises the score; it only lowers the confidence.
Are my images uploaded to a server?
- No. The engine and the detection model run in your browser, so a dropped file stays on your device. When you analyze a link, only the address may transit our server — never a file from your device.
Can it detect Google's SynthID watermark?
- No third-party tool can: Google has not published its detector. Only an optional mention in the metadata is readable — and its absence proves nothing.
Does it detect deepfakes — swapped faces?
- Not as such: no dedicated face detector is embedded, so a video manipulated only locally can pass the global checks. That limit is written in the report, before anyone relies on it.
What are Content Credentials (C2PA) worth here?
- They carry decisive weight only when the manifest's hard binding to the file's exact bytes is verified locally — a manifest copied onto another image is caught. The signer's identity, however, is not verified, and the screen says so.