A new educational project demonstrates how to build an AI detector from scratch using small language models. The author uses the detector as a verifier to train models to bypass detection. This experiment highlights the inherent limitations of current detection tools. Practitioners can use this workflow to test model robustness against classifiers.