AI-generated text detection on IntelLabs peer-review withheld pre-ChatGPT (test)
100TPR @ 0.1% FPR (GPT-4o)Vanilla
Evaluation Results
| Method | Links | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| VanillaBackbone=RoBERTa-base, Fine-tuning Protocol=Fully fine-tuned2026.07 | 100 | 100 | 100 | 99.99 | 100 | 100 | 99.98 | 100 | 100 | 95.37 | 98.74 | 99.39 | 99.85 | 99.96 | 99.99 | |
| AnchorBackbone=Specialized detector, Fine-tuning Protocol=Original implementation by Yu et al. (2025)2026.07 | 63.5 | 83.7 | 88.8 | 59.7 | 80.3 | 86.5 | 59.6 | 75.8 | 81.8 | 13.6 | 28.5 | 36.8 | 72.5 | 88.4 | 92.1 |