Paper Overview
Research Area: NLP Authors: Sondos Mahmoud Bsharat, Jiacheng Liu, Xiaohan Zhao Published: 2025-06-11 arXiv: 2506.08272
Summary
As AI writing assistants become increasingly integrated into real-world drafting and revision workflows, many documents are no longer purely human-written or AI-generated, but the result of progressive human-AI co-editing. However, existing AI text detection benchmarks focus mainly on final outputs, offering little understanding of how AI authorship signals emerge, accumulate, or disappear during the revision process.
The authors introduce OpAI-Bench, an operation-guided benchmark that studies progressive human-to-AI text transformation across multiple granularities: document, sentence, token, and span.
Key Findings
- AI text detectability is not determined solely by the proportion of AI-edited content.
- Detectability is also influenced by the type of editing operations performed.
- The domain of the text affects how detectable AI contributions are.
- The cumulative revision history of a document plays a role in detection outcomes.
Original Abstract (excerpt)
> As AI writing assistants become increasingly integrated into real-world drafting workflows, many documents result from progressive human-AI co-editing. We introduce OpAI-Bench, an operation-guided benchmark for studying progressive human-to-AI text transformation across document, sentence, token, and span granularities.
---
*Auto-collected on 2025-06-11*