[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-llm-assisted-writing-biomedical-papers":3,"news-related-c0ca1295-8b69-4e4f-b29d-24de3bf08d7e":35},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"c0ca1295-8b69-4e4f-b29d-24de3bf08d7e","生物医学论文 89% 带 LLM 痕迹:方法部分也不再是净土","arXiv 预印本用词频漂移的无偏估计扫描 PubMed Central:2025 年底 89% 的开放获取生物医学论文出现 LLM 关联词汇过量,全年比例 77%(2024 年为 52%),连方法部分使用率也超过 50%。","过去几年,学术圈对「论文里有多少是 AI 写的」一直停留在猜测阶段:检测工具不可靠,问卷调查又依赖自愿承认。8 月 11 日提交到 arXiv 的预印本 arXiv:2608.10715 给出了迄今最硬的答案:到 2025 年底,PubMed Central 收录的开放获取生物医学论文中,89% 出现 LLM 关联词汇过量。Nature 的报道直接把「Staggering 90%」写进标题,并指出这一数字远高于此前对 LLM 使用比例的估计。\n\n## 词频漂移,而不是逐篇检测\n\n这项工作来自 Lena Holzwarth、Rita González-Márquez、Dmitry Kobak 三位作者。他们没有再造一个「AI 检测器」去逐篇判定真伪,而是提出一种基于词频变化的无偏估计方法:把整个语料当作流行病学样本,统计 LLM 时代词频的系统性偏移,从分布层面估算被 LLM 改写过的文本比例。作者在摘要里直言,尽管这个方向近期有进展,但此前没有任何现有方法能给出可靠估计——先有可信的测量,才谈得上政策,这是整项工作的起点。\n\n## 三个坐标轴上的数字\n\n- **时间轴**:2025 年全年英文论文的 LLM 使用比例达到 77%,2024 年为 52%;2025 年 12 月发表的论文,近九成带 AI 辅助写作痕迹。\n- **章节轴**:讨论部分段落的使用率约为 68%,接近方法部分段落(32%)的两倍;但即便在方法部分,整体使用率也超过 50%。摘要、引言、讨论等章节的痕迹比方法、结果部分更常见。\n- **自认轴**:2025 年的一项调查显示,71% 的研究人员承认在用 AI 辅助写作,实际比例可能更高——与词频估计的结果互相印证。\n\n## 真正的警报:方法部分过半\n\n讨论部分被大量改写,多数人不会意外——那本来就是「发挥」的章节。扎眼的是方法部分:方法章节存在的意义,是精确记录实验到底怎么做的,它直接决定可复现性。当描述事实的段落也开始被模型重写,读者面对的就不只是文风问题,而是这份记录离原始操作有多远的问题。\n\n对期刊与审稿体系而言,这项研究把一个模糊的担忧变成了可测量的基线。作者也点明了双面性:LLM 写作降低了语言壁垒,对非英语母语作者是实打实的好处,但学术不端与造假风险同步放大。可以预期的下一步,是披露要求与方法章节写作规范收紧——连「方法部分过半」这种数字都有了,继续装作没看见的成本只会越来越高。\n\n参考:[arXiv:2608.10715](https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.10715) · [Nature 报道](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fd41586-026-02551-z)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.10715","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"c33b1bbc-d6ce-4f61-9d5d-1a0704a6a09b","ai-policy",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"dacb5b0b-a927-4ff1-b9a3-1324038062f8","en","89% of biomedical papers now show LLM writing traces","An arXiv preprint tracks word-frequency drift in PubMed Central papers: 89% show excess LLM vocabulary by late 2025, and even Methods sections top 50%.","For the past few years, the question of how much of the scientific literature is written by AI has been stuck at the guessing stage: detectors are unreliable, and surveys depend on self-reporting. A preprint submitted to arXiv on August 11 (arXiv:2608.10715) now offers the hardest answer yet: by the end of 2025, 89% of open-access biomedical papers archived in PubMed Central showed an excess of LLM-associated vocabulary. Nature's coverage put \"Staggering 90%\" in the headline and noted the figure is far higher than previous estimates of LLM use.\n\n## Word-frequency drift, not per-paper detection\n\nThe work comes from Lena Holzwarth, Rita González-Márquez, and Dmitry Kobak. Rather than building yet another per-document \"AI detector,\" they propose an unbiased estimation method based on changing word frequencies: treat the whole corpus as an epidemiological sample, measure the systematic drift in word frequencies of the LLM era, and estimate the share of text altered by LLMs at the distribution level. The authors state plainly in the abstract that despite recent progress, no existing method could produce reliable estimates — trustworthy measurement has to come before policy, and that is the starting point of the whole work.\n\n## The numbers on three axes\n\n- **Time**: across English-language papers published in 2025, LLM usage reached 77%, versus 52% in 2024; papers published in December 2025 showed traces in nearly nine out of ten cases.\n- **Section**: paragraph-level usage in the Discussion section (68%) is about twice that of the Methods section (32%); yet even inside Methods, overall prevalence exceeds 50%. Abstracts, introductions and discussions show traces more often than methods and results.\n- **Self-report**: a 2025 survey found 71% of researchers admit using AI-assisted writing, with actual usage likely higher — consistent with the word-frequency estimates.\n\n## The real alarm: Methods past the halfway mark\n\nHeavy LLM rewriting of Discussion sections surprises few people — that is where authors \"perform.\" What stings is the Methods section: its entire purpose is to record precisely how the experiment was done, and it is what reproducibility rests on. When paragraphs that describe facts start being rewritten by a model, the reader faces not a style question but a distance question — how far this record sits from what was actually done.\n\nFor journals and review systems, this study turns a vague worry into a measurable baseline. The authors also note the double edge: LLM-assisted writing lowers language barriers — a real benefit for non-native English authors — while the risks of misconduct and fraud grow in step. The predictable next step is tighter disclosure requirements and stricter writing norms for Methods sections — once numbers like \"more than half of Methods paragraphs\" exist, the cost of looking away only goes up.\n\nReferences: [arXiv:2608.10715](https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.10715) · [Nature coverage](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fd41586-02551-z)","llm-assisted-writing-biomedical-papers","2026-08-24T21:40:00Z","2026-08-24T15:07:54.170171Z","2026-08-24T15:07:54.170180Z",true,"agent",40,{"items":36},[37,42,47,52,57,62],{"id":38,"title":39,"news_slug":40,"published_at":41},"99916419-0f68-4a6a-a4cf-8bbe353b4d75","康涅狄格法官开出美国首例 prompt injection 制裁令:法庭文件里的隐藏 LLM 暗口令","us-court-prompt-injection-sanctions","2026-08-18T03:00:00+00:00",{"id":43,"title":44,"news_slug":45,"published_at":46},"144fa9dc-de03-4972-a695-3d392f334772","PubMed 中央库研究:2025 年生物医学论文 77% 有 LLM 写作痕迹","pubmed-77-percent-llm-writing-2025","2026-08-26T01:00:00+00:00",{"id":48,"title":49,"news_slug":50,"published_at":51},"ae3f239d-dc29-4ec0-a823-446f463e6bab","皮尤扫了 49 万网页:ChatGPT 之后发布的页面,三分之一带 AI 痕迹","pew-ai-authorship-web-study","2026-08-25T13:08:45+00:00",{"id":53,"title":54,"news_slug":55,"published_at":56},"f9cf9f03-6aca-4d29-94d3-5c6acfeaf435","匿名模型 OX Alpha 短暂登顶 OpenRouter 编码榜:研究者推测底座指向智谱 GLM-5.x","ox-alpha-stealth-openrouter-glm-5-zhipu","2026-08-24T03:00:00+00:00",{"id":58,"title":59,"news_slug":60,"published_at":61},"a64d03b9-1d07-404b-9231-d434c65c44ce","OX Alpha 免费一周:模型页说不训练,EULA 却保留训练权","ox-alpha-stealth-eula-retention-conflict","2026-08-23T13:10:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"43eda321-b0b7-4df7-b20e-9758cbab42c9","记忆越完整,眼前题越做不对:MemTrapBench 把 LLM 长期记忆框架打回原形","memtrapbench-llm-memory-cognitive-traps","2026-08-22T04:00:00+00:00"]