Page MenuHomePhabricator

Investigate automated detection methods for LLM-generated and machine-translated content
Open, Needs TriagePublic

Description

Wikipedia is increasingly receiving edits that contain AI-generated and blindly machine-translated content. For this there is need for methods for automatically identifying such content and edits. The task is suitable for multiple persons to do simultaneously and seek different solutions for the same problem.

Examples from Finnish Wikipedia

LLM-edited articles with hallucinated sources

Machine-translated articles created simultaneously across multiple Wikipedias

LLM-generated text with unnatural writing style

Crosswiki-spamming using machine translated articles

Tracking categories

Task
Investigate how LLM generated content detection could be done. Provide a proposal or results of the investigation in the comments, and we can write an actual task ticket based on the investigation.

Investigation can focus on single topics such as:

  • Methods for detecting hallucinated citations
  • Cross-wiki edit pattern analysis approaches
  • Detecting if text is directly translated from another Wikipedia

Also note for implementation proposals: Only open source tools should be considered. Closed source LLMs are out of scope.