How Pangram uses old school machine learning to detect AI text at very high confidence

This got the juices flowing about deep learning vs LLMs. There seems to be a tendency to latch on to the latest thing as the ultimate silver bullet technology. Right now that's LLMs. Therefore a lot of AI text detection is also LLM (perplexity) based. But those methods don't work that well. A more open-ended custom deep learning approach apparently works much better. Here's pangram's account of why perplexity methods (LLM based LLM detection) don't work: https://www.pangram.com/blog/why-perplexity-and-burstiness-fail-to-detect-ai

I've been thinking about this in the more general context where I find myself skeptical of some of the things people are trying to do with LLMs directly vs custom deep learning or old school ML approaches. We saw the same inappropriate over-application of new things with the crypto/blockchain stuff. The hype is the constant enemy of the sensible engineer. In any case this seems like another good reminder of that.

They claim: "Pangram Text is designed to detect AI-generated content with a near-zero false positive rate."

This is rubbish. We need technologies that will help hold authors accountable -- and technologies that help trace the provenance of information -- not more pretentious spam detectors that centralize 'detection' powers.