The developer behind the Clew-by-Custos AI detection tool discovered that its semantic layer has minimal separating power on real-world text. While the structural layer successfully filtered out non-redundant content, the semantic layer struggled to distinguish between genuinely different outputs on the same topic, as they shared too much vocabulary. This finding contradicts previous synthetic tests and suggests a need for more diverse real-world data across various topics to accurately train and validate the detection system. AI
IMPACT Highlights limitations in current AI detection methods, suggesting a need for more robust semantic analysis and diverse training data.
RANK_REASON Developer reports on limitations found during testing of their AI detection tool using real-world AI output. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
