What Building an AI Detector Taught Me About False Positives

Short summary

A builder of an AI content detector shares hard lessons about false positives after a scanned 1973 paper scored 98% AI-generated. Even a 0.1% false positive rate affects thousands of students monthly. The core problem: detectors measure statistical similarity to model output, not authorship, and the gap between user expectations and probabilistic results creates real-world harm in academic settings.

• •AI detectors measure statistical patterns, not authorship — even 0.1% false positive rates impact thousands of users
• •Human writing edited for clarity can trip the same alarms as AI-generated text
• •Multi-language detection significantly increases false positive rates due to different model fingerprints per language

Generated with AI, which can make mistakes.