扫描 13500 篇 arXiv 论文：38% 疑似由 AI 撰写
13500 篇论文的实测数据，Computer Science 高达 62%，比泛泛而谈有说服力得多。漏检率 20% 说明真实数字还要更高，做科研或投稿的人该看看这个趋势。
一项覆盖 2021 至 2026 年共 13500 篇 arXiv 论文的扫描显示，2026 年 8 月至 10 月采样的论文中有 38% 被 AI 检测器判定为疑似 AI 撰写，而 ChatGPT 发布前的基线仅为 0.4%。自 2023 年起这一比例逐年上升。
学科分布差异明显：Computer Science 以 62% 居首，Quantitative Biology 为 58%，Electrical Engineering 为 51%，Math 仅 3.7%，原因在于公式密集的论文可供检测器分析的文本较少。该检测器为压低误报率而调校，漏检率约 20%，因此真实比例可能更高。
这一数据与 Dwarkesh Patel 在播客中对 AI 研究员参与科研加速的判断相互印证。对期刊编辑和研究者而言，论文写作流程正在被大模型重构，检测与署名规范将成为新的议题。
AI is changing scientific research. We already know that. But those numbers are interesting: More than a third (38%) of new arXiv papers now read as AI-written, according to a scan of 13,500 papers from 2021 to 2026.
Before ChatGPT, the detector flagged just 0.4% (that's where the threshold was set). For papers sampled in August to October 2026 it's 38%, and the curve has gone up every year since 2023.
- Computer Science is highest at 62%, then Quantitative Biology at 58% and Electrical Engineering at 51%. Math sits at 3.7%, though equation-heavy papers give the detector less prose to work with
- It's tuned to avoid false positives and misses about 20% of AI text, so the real number is probably higher
Last night, I watched Dwarkesh Patel’s latest podcast episode on RSI. He seems quite convinced that AI researchers will increasingly engage in research next year. Consequently, I expect the figures to rise significantly. This raises the question of how we will generate knowledge in the future and what the substance of science will become.
This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include arXiv, AI detection, research, ChatGPT.
Source attribution: curated from kimmonismus (@kimmonismus).
