Exploring the internal representations of Pangram 3.3.2

I use Pangram quite extensively (burning through my 600 token allowance every month). They managed to get their false positive rate impressively low: if Pangram says something is 100% AI-written, you can trust that.

But they need to improve their humanizer dataset. Right now, most models can be given system prompts which cause them to emit text classified as 100% human. It looks like their automated humanizers do worse than these system prompts. Or (alarming if so) they chose not to include ones that would make their product look unreliable.

I wonder if they had enough material from individual humans if they could've distinguished between them as well? It really seems like their model is learning to recognize some general form of writer's "voice", so to speak (and I assume their final layer just knows which voices are supposed to be tagged as what).

Hoping for a follow-up with Sparse Autoencoders.