2026 的寫作生態
2023-2026 之間，學術寫作的工具鏈經歷了三波變革：(1) generative AI (ChatGPT, Claude, Gemini) 進入寫作流程；(2) Preprint 從可選變成標準（bioRxiv 已超過 30 萬篇）；(3) Reproducibility tools (Quarto, Snakemake, Docker) 從「nice to have」變成「期刊要求」。
這頁告訴你 2026 主流期刊到底要什麼、不要什麼，以及如何在合規的前提下用 AI 加速你的寫作。
From 2023 to 2026, the academic writing toolchain saw three shifts: (1) generative AI (ChatGPT, Claude, Gemini) entered the workflow; (2) preprints went from optional to standard (bioRxiv exceeds 300k papers); (3) Reproducibility tools (Quarto, Snakemake, Docker) shifted from "nice to have" to "journal-required."
This page covers what 2026 mainstream journals actually require, what they prohibit, and how to use AI compliantly.
一、AI 政策共識
截至 2026/05，Elsevier, Springer Nature, Wiley, Taylor & Francis, SAGE, Cell Press 等主流期刊的 AI 政策已收斂為以下五點。
As of 2026/05, AI policies across Elsevier, Springer Nature, Wiley, Taylor & Francis, SAGE, Cell Press, and other major publishers have converged to the following five points.
| 政策 | 說明 |
|---|---|
| ① AI 不能掛 author | 所有主流期刊禁止 ChatGPT, Claude, Gemini 等列為 author 或 co-author。理由：AI 無法承擔 ICMJE 第 ④ 條 "accountability"。All major publishers prohibit ChatGPT, Claude, Gemini etc. as authors. Rationale: AI cannot bear ICMJE criterion ④ "accountability". |
| ② Generative 內容必須揭露 | 若 AI 用於草擬實質內容，必須在 Methods 或 Acknowledgements 揭露：工具名稱 + 版本 + 用途 + 用在哪個段落。If AI drafted substantive content, disclose in Methods or Acknowledgements: tool name + version + purpose + where used. |
| ③ 文法/翻譯/校對例外 | 用 Grammarly, DeepL, Claude 純粹做語法 / 翻譯 / 校對，多數期刊不要求揭露 (但 APA、Cell 仍鼓勵揭露)。Using Grammarly, DeepL, Claude purely for grammar/translation/proofreading typically doesn't need disclosure (but APA, Cell still encourage it). |
| ④ 圖片生成/變造嚴禁 | Cell Press 等明文禁止 generative AI 製作或修改科學圖片 (DALL-E, Midjourney, Stable Diffusion)。Conceptual graphical abstract 可能例外但要揭露。Cell Press etc. explicitly prohibit generative AI for creating/altering scientific images (DALL-E, Midjourney, Stable Diffusion). Conceptual graphical abstracts may be exempt but require disclosure. |
| ⑤ Reviewer 不得用 AI | Cell, NEJM, JAMA, NIH 明文禁止 reviewer 把 confidential manuscript 上傳到 ChatGPT 等公開 AI (洩漏 confidential research)。違反 = peer review breach。Cell, NEJM, JAMA, NIH explicitly prohibit reviewers from uploading confidential manuscripts to public AI (ChatGPT, Claude). Doing so = peer review breach. |
二、AI 揭露範本
# Generic acknowledgements style During the preparation of this manuscript, the author(s) used Claude 3.7 Sonnet (Anthropic, 2025-Q4) to improve language and check grammar in the Discussion section. After using this tool, the author(s) reviewed and edited the content as needed and take full responsibility for the content of the publication.
# Methods section AI use Initial drafts of the data preprocessing pipeline (Snakemake rules and R scripts) were generated with ChatGPT-5 (OpenAI, January 2026) and Claude Sonnet 4.6 (Anthropic, March 2026). All generated code was independently tested, validated against benchmark datasets (GSE298765), and manually corrected before use. Authors take full responsibility for code correctness and have deposited the final verified pipeline at github.com/charlene/tnbc-pipeline.
# Cell Press 2026 disclosure (most stringent) Declaration of generative AI in scientific writing: During the preparation of this work the author(s) used [Tool name] [Version] for [specific use: e.g., language polishing of the Introduction; structuring the Discussion; literature retrieval]. The authors reviewed and edited the content and take full responsibility for the content of the publication. # Cell Press additionally disallows AI for image creation/editing
三、Preprint 工作流
2026 年 preprint 已成主流。Cell Press, Nature, PNAS, eLife 等主刊均明確支援 preprint 與後續正式發表並行 (post-bioRxiv 不被視為 prior publication)。
建議流程：
- 論文寫好後 → 投 bioRxiv (生命科學) / medRxiv (臨床) / arXiv (計算/物理) / Research Square (跨領域)。
- 同時 / 之後投正式期刊。
- Preprint DOI 立刻引用，可用於 grant application、conference talk、招聘材料。
- 正式被接受後，bioRxiv 自動連結到正式版本。
By 2026, preprints are mainstream. Cell Press, Nature, PNAS, eLife etc. explicitly support preprints alongside formal publication (post-bioRxiv ≠ prior publication).
Recommended workflow:
- Draft complete → post to bioRxiv (life sciences) / medRxiv (clinical) / arXiv (computational/physics) / Research Square (interdisciplinary).
- Simultaneously / subsequently submit to a formal journal.
- Preprint DOI is immediately citable — useful for grant applications, conference talks, recruitment.
- Once accepted, bioRxiv auto-links to the formal published version.
| Preprint server | 領域 | 特色 |
|---|---|---|
| bioRxiv | 生命科學Life sciences | CSHL 維護；多數生醫主刊皆認可CSHL maintained; recognized by most biomedical journals |
| medRxiv | 臨床醫學Clinical medicine | 嚴格 screening (避免直接影響病人決策)Stricter screening (to avoid direct patient impact) |
| arXiv | 物理 / 數學 / CSPhysics / math / CS | 最早的 preprint server (1991-)Oldest preprint server (since 1991) |
| chemRxiv | 化學Chemistry | 由 ACS, RSC, GDCh, CSJ 共同維護Co-maintained by ACS, RSC, GDCh, CSJ |
| Research Square | 跨領域Interdisciplinary | 與 Springer Nature 整合 (In Review feature)Integrated with Springer Nature (In Review feature) |
四、Quarto / R Markdown / Jupyter
2026 reproducibility 的標準做法：論文與程式 / 圖表 / 數據在同一個檔案內，確保 every figure 都可以從原始數據一鍵重生。
2026 reproducibility best practice: paper, code, figures, data live in one file — every figure regeneratable from raw data with one command.
📘 Quarto (2022+)
R Markdown 的下一代，支援 R + Python + Julia + Observable。可一鍵 render 成 PDF / HTML / docx / pptx。Posit (前 RStudio) 維護。2026 推薦首選。
R Markdown's successor — supports R + Python + Julia + Observable. One-click render to PDF / HTML / docx / pptx. Maintained by Posit (formerly RStudio). 2026 top pick.
📗 R Markdown
knitr-based。R 為主、Python via reticulate。已被 Quarto 取代但仍廣泛使用。bookdown 適合 thesis-length。
knitr-based; R-first, Python via reticulate. Superseded by Quarto but still widely used. bookdown is suited to thesis-length docs.
📙 Jupyter (nbconvert)
Python-first。Jupyter Book 適合做 large reproducibility report。但同行常因 cell 順序問題 reproducibility 不穩。
Python-first. Jupyter Book suits large reproducibility reports. But cell-order issues hurt reproducibility in practice.
五、2026 工作流
專案管理
用 Snakemake / Nextflow 管理 pipeline；GitHub repo 從 day 1 開始；用 renv (R) / uv (Python) lock 套件版本。
Manage pipelines via Snakemake / Nextflow; start GitHub repo from day 1; lock package versions with renv (R) / uv (Python).
文獻 + 寫作
Zotero (with Better BibTeX) → Quarto + LaTeX/Word；可在 Quarto chunk 內直接放分析 code，避免 figure 與 paper 不同步。
Zotero (with Better BibTeX) → Quarto + LaTeX/Word; Quarto chunks hold analysis code so figures and paper never desynchronize.
AI 輔助
Claude / ChatGPT 用於：(a) 結構回饋、(b) 語言潤稿、(c) 程式 debug。不用於：直接生成最終敘述、引用、圖片。每次 AI 生成後手動驗證並修改。
Claude / ChatGPT for: (a) structural feedback; (b) language polishing; (c) code debugging. Not for: directly generating final prose, citations, or images. Always manually verify and revise.
Preprint + 投稿
同時送 bioRxiv + 投正式期刊。Code 同步 deposit GitHub + Zenodo (frozen DOI)；data 送 GEO/Zenodo/Figshare。
Post to bioRxiv simultaneously with formal submission. Deposit code on GitHub + Zenodo (frozen DOI); data on GEO/Zenodo/Figshare.
Revision + 接受
Revision response 用 Quarto track-changes；接受後在 ORCID 上發布、社群分享 (X, BlueSky, Mastodon, Slack)、寫 lay summary 給 institutional press office。
Quarto track-changes for revision response; on acceptance update ORCID, share on social media (X, BlueSky, Mastodon, Slack), write lay summary for institutional press office.
六、AI 安全使用決策樹
🌳 AI 使用紅綠燈
📝 自我檢測
1. 你用 ChatGPT-5 草擬 Discussion 的 2 段，再大幅修改。2026 主流期刊的標準做法是？
1. You used ChatGPT-5 to draft 2 paragraphs of Discussion then heavily revised. 2026 standard practice?
2. 關於 preprint (bioRxiv) 與正式期刊，下列何者正確？
2. Which is correct about preprint (bioRxiv) and formal journals?
3. 你被邀請當 reviewer，可不可以把 manuscript 上傳到 ChatGPT 幫你寫 review？
3. As a reviewer, can you upload the manuscript to ChatGPT to help write the review?