How we measured AI writing across arXiv, and where the measurement breaks

如何在 arXiv 上测量 AI 写作及其测量失效之处

HN 148 分 · 106 条评论 · 作者 dopamine_daddy · 来源 unslop.run · HN 讨论

【摘要】

A study of 12,750 arXiv papers reveals that approximately one-third of recent submissions appear machine-written, with the detection calibrated to a 0.4% false-positive rate using pre-2023 data as a control. The analysis spans from January 2023 to July 2026, scoring the full body text of the initial PDF versions to ensure accuracy.

⋯ 继续阅读请登录会员 ⋯