Cross-Model Meta-Classification for AI-Text Detection

Stacks Binoculars, Fast-DetectGPT, and RoBERTa via an XGBoost meta-classifier. On the RAID benchmark (≈23k samples), it cuts the false-positive rate at 95% recall from ~0.99 to 0.22.

Code · BSc Computer Science (AI), Cardiff Met, 2025–26

Bars are normalised so longer = more AI-like; see "What do the detector bars mean?" below. First inference is slow while models load; later ones are faster.