I've been doing more testing and thinking. I'm now convinced genAI detection is possible and Pangram does it pretty well. But there are a few caveats. First, Pangram likely depends on the limits of current LLMs because it seems to be using n-gram math. 1/
I uploaded a few texts to test Pangram, Unlke other AI detectors, it kind of works. It recognized some of my uploaded texts as 100% human (which they were). It identified genAI texts (although not that they were 100% LLM generated - see below).