top | item 45043237

Apparently it's easy to detect LLM-generated text now

5 points| jcbe | 6 months ago |papers.ssrn.com

2 comments

order

jcbe|6 months ago

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.

allears|6 months ago

It offered a free demo, but it wanted me to subscribe first, so I declined.