Are AI Detectors Actually Effective?

We did some split tests on many of the popular AI detectors that large companies use when detecting AI images, we had completely AI images pass with no issues by changing the structure of the pixels in the image, we had completely real images detected as 99% AI... extremely inconsistent results.

Do these companies actually know what they are doing? Or are they just figuring it out along the way?

tjpenthouse said:

We did some split tests on many of the popular AI detectors that large companies use when detecting AI images, we had completely AI images pass with no issues by changing the structure of the pixels in the image, we had completely real images detected as 99% AI... extremely inconsistent results.

Do these companies actually know what they are doing? Or are they just figuring it out along the way?

Click to expand...

I don't know which companies you're referring to, but with Genfluence AI I have created and posted AI models's images on Instagram with very good success.

honestly a lot of them are figuring it out as they go. detecting ai images reliably is way harder than generating them. once you slightly alter pixel distribution, compress the image, rescale it, or run minor edits, many detectors lose the signal they rely on.
most of these tools are probabilistic models trained on known generator artifacts. the problem is new models constantly change those patterns, and simple transformations can erase them. so results become inconsistent fast.