I passed all 5 of my real-AI tests. The most useful thing I found is that half my detector barely works.
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.