AI-generated text detection on Wikipedia GPT2-XL generations (+60L,600 attack)
98Accuracy (1% FPR)BM25
Evaluation Results
| Method | Links | |
|---|---|---|
| BM25Corpus Size=3K generations from model itself2023.03 | 98 | |
| SPCorpus Size=3K generations from model itself2023.03 | 81.5 | |
| BM25Corpus Size=9K generations pooled from all three models2023.03 | 78.7 | |
| SPCorpus Size=9K generations pooled from all three models2023.03 | 63.2 | |
| Watermark2023.03 | 57.2 | |
| OpenAI2023.03 | 14.8 | |
| DetectGPT2023.03 | 4.6 |