EditLens: Quantifying the extent of AI editing in text (2025)
11 hours ago
- #computational linguistics
- #AI detection
- #text editing
- The paper introduces EditLens, a method to quantify and detect AI editing in text, distinguishing it from human-written and AI-generated content.
- Lightweight similarity metrics are proposed to measure the extent of AI editing, validated by human annotators.
- EditLens, a regression model using these metrics as supervision, achieves state-of-the-art performance in binary (94.7% F1) and ternary (90.4% F1) classification tasks.
- The research highlights applications in authorship attribution, education, and policy, and includes a case study on AI-edits from Grammarly.
- The models and dataset will be publicly released to encourage further research.