LLM is Large Language Model, the fancy word for ChatGPT-like AI's where you can ask it a question or give it a topic, and it will output a fluent-sounding but factually incorrect answer.
Newer users who don't know our rules have been creating articles using AI. One AFC reviewer reports having seen 60 drafts that were LLM-generated.
There are open source detectors for LLM that are very accurate They'll give a probability, and many of the matches are above 99%. We should look into the feasibility of adding a feature to PageTriage that tags articles with LLM-generated text.
Technically, there's several approaches:
- We could create something similar to the pagetriagetagcopyvio API
- We could find an open source one and incorporate its code, assuming it is not resource-intensive, i.e. requiring a farm of AIs or something.
- We could create a third party tool similar to Earwig's copyvio detector
Credit to S0091 for the idea.