Research map / Large language models / Language, safety and interpretability
Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| paraphrase AI text with DIPPER (60L,60O) | watermark detector · open-ended generation GPT2-1.5B | detection accuracy: 100.0 -> 57.2 | Hurt |
| paraphrase with DIPPER (60L,60O) | DetectGPT · open-ended generation GPT2-1.5B | detection accuracy: 70.3 -> 4.6 | Hurt |
| paraphrase with DIPPER (60L,60O) | OpenAI classifier · open-ended generation GPT-3.5 | detection accuracy: 30.0 -> 15.6 | Hurt |
| paraphrase with DIPPER | RankGen detector · open-ended generation GPT2-1.5B | detection accuracy: 13.5 -> 28.5 | Mixed |
| use retrieval (BM25) over generation corpus instead of watermark detection | AI text detection · long-form QA GPT2-XL paraphrased | detection accuracy: 55.8 -> 97.8 | Helped |
| use BM25 instead of P-SP retriever (ablation) | retrieval detector · long-form QA GPT2-XL 9K corpus | detection accuracy: 75.4 -> 95.2 | Helped |
| enlarge retrieval corpus from 3K to 9K pooled generations (ablation) | BM25 retrieval detector · long-form QA GPT2-XL | detection accuracy: 97.8 -> 95.2 | Hurt |
| add 43K ShareGPT responses to retrieval corpus (ablation) | BM25 retrieval detector · long-form QA GPT-3.5 | detection accuracy: 96.2 -> 95.5 | No clear effect |
| add left/right context to paraphraser training (ablation) | DIPPER · open-ended GPT2-XL generations | GPT-3.5 perplexity preference: 71% preferred | Helped |
| increase paraphrase lexical/order diversity | DIPPER · open-ended generation | semantic similarity: 99.2 -> 94.3 | Hurt |
Closest papers by meaning
- Adversarial Paraphrasing: A Universal Attack for Humanizing AI-Generated Text (NeurIPS 2025 · similarity 0.93)
- Position: On the Possibilities of AI-Generated Text Detection (ICML 2024 · similarity 0.88)
- DNA-DetectLLM: Unveiling AI-Generated Text via a DNA-Inspired Mutation-Repair Paradigm (NeurIPS 2025 · similarity 0.88)
- Watermarking Conditional Text Generation for AI Detection: Unveiling Challenges and a Semantic-Aware Watermark Remedy (AAAI 2024 · similarity 0.87)
- DNA-GPT: Divergent N-Gram Analysis for Training-Free Detection of GPT-Generated Text (ICLR 2024 · similarity 0.87)
- Raidar: geneRative AI Detection viA Rewriting (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.