Example: Stable Diffusion image
Downloaded straight from a diffusion tool. The generator wrote its own prompt, sampler and seed into the file, so this needs no guesswork at all.
This is a real result from the same pipeline a visitor uses — not a mock-up. Upload the same file and you get the same verdict.
The file itself declares how it was made: The PNG stores a "parameters" block with the prompt and sampler settings used to generate it. This is embedded provenance data rather than a guess, so it is the strongest evidence available — though metadata can be stripped or forged.
What the file says about itself
The strongest evidence available — read straight from the file. It can still be stripped or forged, and we don't verify the cryptographic signature.
- Image contains generation settingspoints to AI · Embedded provenance
The PNG stores a "parameters" block with the prompt and sampler settings used to generate it.
Supporting signals
Individually weak. They shift the estimate but never decide it.
- Dimensions match a standard AI output sizepoints to AI · Statistical hint
1024×1024 is a preset output size for common image generators.
Detection models
The file's own metadata was conclusive, so no model was needed.
Analysed in 0.1s. Detection models are unreliable on content they weren't trained on, and text detectors are documented to misjudge writing by non-native English speakers. Never use this as sole evidence of wrongdoing.
Why this matters. When a file states how it was made, we report that and stop. No model was consulted for this result.