Unmasking the ‘Concentration’ of AI Edits! New Model EditLens Identifies Human and AI Mixed Text with Over 90% Accuracy

📰 News Overview

• Proving the Uniqueness of AI-Edited Text: It has been revealed that text edited by AI possesses distinct characteristics that set it apart from purely human-created or AI-generated texts.
• Development of the New Model “EditLens”: A lightweight similarity metric is employed as an intermediary monitor to build a regression model that predicts the amount of AI editing within the text.
• Remarkable Identification Accuracy: Achieved a top-notch performance with an F1 score of 90.4% in a three-class classification of human, AI, and mixed (AI-edited) texts.

💡 Key Points

• Quantifying “Rewriting”: The innovative aspect lies in its ability to quantify the degree of editing by AI, rather than simply detecting whether it was AI or human-generated.
• Practical Validation: Analyzed the editing effects of the popular writing assistant tool “Grammarly” as a case study, demonstrating its effectiveness in real-world applications.
• Open Source Commitment: The team has pledged to release the model and dataset to foster the development of the research community.

🦈 Shark’s Eye (Curator’s Perspective)

EditLens brings a groundbreaking gradient to detection technology, which previously relied on a binary choice of “AI or human,” by introducing the question of how much AI has intervened!

Particularly impressive is the implementation that combines lightweight similarity metrics to learn patterns unique to AI editing. This model can detect the workflow where “human drafts are polished by AI” with over 90% accuracy, marking a game changer for education and publishing industries. The ability to quantify the impact of common tools like Grammarly is a testament to its high practicality!

🚀 What’s Next?

With the quantification of text “purity,” we may soon see specific standards like “AI usage rate below X%” established for assignment submissions in educational institutions and ensuring the credibility of news articles. Additionally, developers of AI writing tools will have to shift to mimicking editing quirks that are closer to human behavior.

💬 Haru Shark’s Takeaway

The era where “I just had a little help from AI” will no longer hold water! There’s no escaping the jaws of the shark! 🦈🔥

📚 Terminology

• EditLens: A regression model developed to predict the extent of AI editing contained in text.

• F1 Score: A metric used to evaluate the accuracy of the model, combining precision and recall, where closer to 1 indicates higher performance.

• Authorship Attribution: Identifying who wrote a text (or how much AI was involved).

• Source: EditLens: Quantifying the extent of AI editing in text (2025)