Zero-shot detection
Updated June 10, 2026
Detecting AI text without training on labeled examples — using the mathematics of generation itself.
Definition
A zero-shot detector identifies machine text without being trained on human-vs-AI examples. Instead it interrogates a language model's own probability landscape directly. The landmark method, DetectGPT (2023), uses probability curvature: generated text sits at local peaks of the model's probability function — perturb it slightly (reword a few phrases) and probability drops consistently; perturb human text and probability moves in both directions.
Versus trained classifiers
Trained classifiers (what commercial detectors mostly use) learn from large labeled corpora — accurate on familiar models, blind-sided by new ones until retrained. Zero-shot methods need no retraining cycle and adapt naturally, but cost more compute and lean on access to a scoring model. Production systems increasingly blend both, with raw signals like perplexity as the shared foundation.
Why it matters
Zero-shot research is why "detectors can't keep up with new models" is only half true — methods exist that track generation itself rather than chasing model releases. The practical takeaway is the same as everywhere in this glossary: assume detection capability persists, write text that's genuinely varied and yours, and verify rather than guess.