New generators multiplied after 2025, and models such as Seedream (ByteDance), ERNIE-Image (Baidu) and Z-Image (Alibaba) are widely used for photo-like results. Detection varies by generator, so this page shows the measured number for each one.
Results differ by generator
The same detectors catch generators resembling what they saw in training and miss those that differ. As the box above shows, some generators were caught almost entirely while Seedream and Reve stayed around half. We write the weak spots down instead of hiding them.
Traces left by labelling rules
Since September 2025 services in China must add visible and metadata labels to AI-generated content. When those labels survive in the original they are read as provenance evidence; when a copy has lost them, the detectors decide.
What a miss looks like
When the detector score is below the threshold, the result is inconclusive unless human-side evidence is strong. Inconclusive leaves room for the image to come from such a generator; it does not confirm a person made it.
How it improves
We retrain the detectors on public samples from the newest generators and ship only after separately confirming they still keep false alarms on real photos within budget. When the numbers change, the box on this page changes with them.
Frequently asked
Why do the numbers differ by generator?
Because each generator resembles what the detectors learned to a different degree. That is why we show per-generator numbers rather than one average.
Are such small samples meaningful?
Each generator has only a handful of samples, so the error is large. Read them as direction, and read individual results with the ledger.
Are originals from Chinese apps caught better?
If mandated labels survive they are strong provenance evidence. Once a copy loses them, only the detector result remains.
What about generators not listed?
They are scanned by the same detectors, but we have no separate measurement, so we promise no number.