Can any tool really tell whether an image is AI-generated?
Not from the pixels alone, and not with certainty. The reliable evidence is metadata: content credentials, generator names, and embedded generation parameters. When that evidence is present the answer is close to conclusive. When it has been stripped — which is what social platforms, messaging apps and privacy tools do — a generated image and a real photo that went through a phone editor become very hard to separate. This tool reports the evidence and its confidence instead of inventing a percentage.
Why is there no percentage score?
Because a percentage would be made up. The statistical measurements most detectors rely on overlap heavily between camera photos and generated images: a real photo of a sunset reproduces the same "anomalous" colour readings as a synthetic one. The measurements are published here with their camera-original reference ranges so you can see how weak they are, and the ones that cannot discriminate are kept out of the verdict entirely.
What are content credentials, and why do they matter most?
Content credentials (C2PA) are a cryptographically signed manifest embedded in the file that records how the content was produced. If the manifest states a digitalSourceType of trainedAlgorithmicMedia, the file itself declares that a model generated it. ChatGPT and DALL·E images ship with such a manifest by default, as do Adobe Firefly exports. A signed manifest is the strongest evidence available — and a ChatGPT image with no manifest at all has had its metadata deliberately removed.
The result says "not verifiable". Does that mean the photo is fake?
No. It means the file no longer carries anything that proves its origin. Stripping metadata is exactly what privacy tools, messaging apps and social networks do to real photographs, so a missing signature is not evidence of generation. It only means the question cannot be answered from this file.
Is my image uploaded anywhere?
No. The file is read and analysed entirely inside your browser using JavaScript. The original never leaves your device and nothing is stored or logged.
Which file formats are supported?
JPG, PNG and WebP receive the full analysis, including codec fingerprints such as Huffman tables, chroma subsampling and chunk inventory. Other formats such as GIF or AVIF receive the provenance and metadata layers only, because codec fingerprinting is implemented per container.
Why does the report say my screenshot cannot be trusted?
Screenshots are re-scaled and re-compressed, which erases sensor noise and fine detail — exactly the measurements that would otherwise hint at a generated image. When the tool detects screenshot-like UI strips it warns you explicitly instead of drawing a conclusion from distorted data.
How is this different from other "AI detector" websites?
Those sites mostly run a neural network trained on older diffusion models. Measured against files with known provenance, such detectors both miss modern generators and flag real photographs as AI. This tool deliberately does not use them. It reports verifiable, explainable evidence — signed credentials, generator markers, file structure — and is explicit about what it cannot determine.