They trained a model to detect AI-written text.

It works fine on mainstream models where the output hasn't been edited.

It doesn't work on humanizers or custom model output.