This is interesting. Do you know what features the classifier is matching on? Like how much does stuff like whitespace matter here vs. deeper code structure? Put differently, if you were to parse the AI and non-AI code into AST and train a classifier based on that, would the results be the same?

Candidly, it's a bit of a black box still. We hope to do some ablation studies soon, but we tried to have a variety of formatting and commenting styles represented in both training and evaluation.