According to 3 out of 11 tested detectors, the other 8 giving a human rating. The only reasonable conclusion is that those 3 tools are broken, cannot conclude for others. I'll keep trusting Pangram.

It's good to have this kind of data in any case. I assumed it was easy to measure accuracy with test sets for this kind of ML problem, so the detectors cannot be too wrong if evaluated properly, and that seems to be the case for the 8 non broken ones.

"But it does not predate the writing style — my very specific writing style — that LLMs themselves were trained on.

Because I’ll have you know that I have been blogging on the worldwide web since 2004"

Having your works being injested for LLM regurgitation should cause them to be a more likely match?

[flagged]