GLOSSARY
AI Detectors
Tools that guess whether text or media was machine-made — useful as signals, unreliable as judges, and increasingly entangled with plagiarism accusations.
Detectors look for statistical fingerprints of generation — uniform predictability, unusual word distributions, watermarks when present. Their documented failure mode matters more than their hits: false positives disproportionately flag non-native writers, whose prose is (correctly) more predictable, which has led to students expelled on evidence courts later discarded. OpenAI quietly retired its own text detector over accuracy concerns.
The practical stance: treat detector scores as one signal among several, never as proof. For educators and editors, process beats policing — drafts, version history and oral checks reveal learning better than a percentage. For builders, detection is an arms race that watermarking (proactive) is better positioned to win than classification (reactive).