EditLens: Quantifying the Extent of AI Editing in Text

TLDR

• ArXiv paper trains EditLens, a regression model that detects and quantifies how much AI editing occurred in human-written text, achieving F1=94.7% binary and F1=90.4% ternary classification.

Key Takeaways

• EditLens uses lightweight similarity metrics as intermediate supervision to predict AI-edit magnitude, not just binary AI-vs-human detection.
• Ternary classification (human / mixed / AI-generated) hits F1=90.4%, a meaningful step beyond prior fully-AI-generated detection approaches.
• The paper validates similarity metrics against human annotators before using them as training signal, grounding the regression targets empirically.
• Case study applies EditLens to Grammarly-edited text, showing the approach works on real-world popular writing tools, not just synthetic datasets.
• Models and dataset will be publicly released, making this directly usable for authorship attribution, education integrity, and policy tooling.

Hacker News Comment Review

• No substantive HN discussion yet.

Original | Discuss on HN