Every major generation service marks its output somehow. The trouble is that each one does it differently: some write metadata, some embed a mark in the pixels, some print words in a corner, and each needs its own reader. This product reads everything that public specifications and public models allow in one pass, and says plainly what it cannot read.
The first is metadata: the AI disclosure assertion in a C2PA manifest, the IPTC digital source type, and the AIGC label of China's GB 45438 standard. The file wrote these itself, so they are easy to read and disappear with a single re-save. The second is an invisible watermark in the pixels. The fixed watermark of the official Stable Diffusion pipeline and Adobe TrustMark are read locally with published code and models. The third is visible text: published label phrases such as Imagined with AI, AI-generated content, AI生成 and its Korean equivalent are read from the image edge with text recognition.
A mark left by the generating service itself is decisive: the C2PA AI disclosure, the GB 45438 AI-generated label and the Stable Diffusion watermark all are, and the provenance axis records them as AI-generated, self-declared. Visible text can be typed by anyone, so it is moderate evidence and never makes a verdict on its own. Labels that reflect a user declaration or a platform's guess are weak. Adobe TrustMark is not AI evidence at all: it is embedded together with Content Credentials in real photographs too, so it is used only as provenance information.
A missing mark is not evidence, but a mark that should be there and is not is different. When the pixels still carry a TrustMark watermark and the C2PA manifest is gone, the file once had credentials and the metadata was stripped. The same holds when the file names a tool that attaches credentials to every output yet has no manifest, or when the GB 45438 label slot is present but empty. In those cases the provenance state is recorded as stripped, while the AI likelihood verdict does not move.
Google SynthID is widely used for images, audio, video and text, but its detector is not public, so it can only be confirmed through each company's own service or an API that requires an access application. The text watermarks of Gemini and Claude use private keys and cannot be read locally either. Even readable marks are lost along the way: the Stable Diffusion watermark disappears after one JPEG re-compression, TrustMark mostly survives re-compression and halving, and visible text survives a screenshot but not a crop. A result that found no mark therefore always means unknown, never human-made.
This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.