It's been my understanding since GPT-3 that tools that claim to identify LLM authorship of arbitrary text are basically unreliable. This apparently is not true

It's been my understanding since GPT-3 that tools that claim to identify LLM authorship of arbitrary text are basically unreliable. This apparently is not true anymore: Pangram has a false positive rate below 0.5% and is robust to various adversarial techniques.