Why this is different
There are a dozen AI image detectors. Most work the same way: you upload your picture to their server, a model looks at the pixels, and you get a number back. Here is what happens here instead — and, further down, where that approach is worse.
| Most detectors | SynthCheck | |
|---|---|---|
| Where does your picture go? | To their server, to be processed | Nowhere. It is read on your device |
| What comes back? | A confidence percentage | The evidence found, graded by how much it can carry |
| Can you see how it decides? | No — the model is theirs | Yes. The whole thing is public, and it is the same code your browser runs |
| Signed Content Credentials | Read and repeated, or handed to an uploader | Signature and image binding checked on your device |
| Cost and limits | A few free checks a day, then a subscription | Free, no account, no limit — nothing is processed on our side to pay for |
| Works with the internet off? | No | Yes, and that is the easiest way to prove nothing is uploaded |
Where a classifier does better
A machine-learning detector looks at the pixels. We do not. On a picture stripped of everything — screenshotted, re-uploaded, re-saved — a good classifier can still see traces in the image itself that we have no way to read. We fall back to the shape of the file, which catches a great deal but not everything. If a check here comes back “can’t tell”, a pixel-based tool is a reasonable second opinion. We would rather say that than pretend the gap does not exist.
Why there is no percentage
A number feels like an answer, and that is the problem. The same detector that scores in the nineties on an easy test set drops towards a coin flip on images built specifically to look authentic — and the number it shows you looks exactly the same in both situations. Evidence does not have that property: you can see what was found, and see when nothing was.