AI image detector benchmarks
Benchmarks and evaluation
We will only publish results that say what was tested, when it was tested, and where the tool struggled.
What we will publish
A public benchmark will name the evaluation date, model or provider version, sample source and licensing, image categories, and uncertainty rate. We will not advertise an accuracy figure without explaining the test behind it.
What makes a benchmark useful
A useful set separates real photos, generated images, screenshots, compressed files, and digital art. It reports false positives, false negatives, and uncertain results, not just one flattering score.
Where we are now
The first release comes before our own frozen evaluation set. Until that work is published, we describe the limits of the tool rather than making a broad performance claim.