注：本人今后统一使用“计算性智能”一词，以强调这类系统的计算基础，避免“人工智能”一词引起它具有或模仿人类思维的误解。

计算性智能辅助写作的合理使用与如实说明

本文由 ChatGPT 根据博客发布者提出的问题撰写，经发布者审阅后发表。

读到达特茅斯学院教务长 Santiago Schnell 使用 AI 辅助写作引发的争议后，有人提出一个切身的问题：自己以往发表的文章，有些得到 ChatGPT 较多帮助，有些较少。究竟可以使用多少帮助，才不会被认为有不当行为？

这个问题没有一个适用于所有文章的百分比。关键是计算性智能具体参与了哪些工作，发表文章的平台有什么规定，以及这一过程是否得到如实说明。

修改语法或翻译已经写好的文章，与根据一个人的想法起草段落，参与程度不同。如果计算性智能还提出了论点、例子或结论，参与程度又有所不同。若它起草了大段文字，仅把这种帮助称为“校对”，便不能准确反映写作过程。

以自己名义发表文章的人，应当理解并认可文中的论点，核实事实和引文，仔细审阅最终文本，并对发表的内容负责。向学术期刊、报纸或其他媒体投稿时，还须遵守该平台关于计算性智能的具体规定。

使用说明可以简短，但应与实际情况相符。例如：“文章的主要观点由本人提出；写作过程中使用 ChatGPT 协助整理结构、起草部分文字和修改语言；终稿经本人审阅。”如果只用了语言修改，就只说明语言修改；如果大部分文字由 ChatGPT 起草，也应如实说明。

Schnell 事件说明，写作过程比检测分数更值得关注。《达特茅斯报》报道，他的多篇作品获得了很高的“AI 写作”检测分数。他则表示，思想和初稿由自己完成，AI 用于语言方面的协助，最终文本由自己负责。检测工具无法单凭成文后的文字还原整个写作过程，也无法衡量一个人的思想贡献。清楚说明各方实际做了什么，比争论某个百分比更有意义。

保留提纲、草稿和修改记录，有助于日后说明自己的贡献。最直接的原则是：怎样使用了工具，就怎样说明。

关于本文的说明：博客发布者提出问题并决定发表；正文由 ChatGPT 起草。本文不应被视为发布者亲自撰写的文章。

参考资料：《达特茅斯报》关于 Schnell 事件的报道；国际医学期刊编辑委员会关于使用 AI 的建议。

Provost of Dartmouth busted using AI for his own writing in newspapers and academic journals

Responsible Use and Disclosure of Computational Intelligence in Writing

ChatGPT drafted this article in response to a question from the person publishing it on this blog. That person reviewed the article before publication.

After reading about the controversy involving Dartmouth College provost Santiago Schnell and his use of AI in writing, someone raised a personal question: ChatGPT had helped with some of their previously published articles more than with others. How much assistance can a person use without being accused of wrongdoing?

There is no percentage that applies to every article. What matters is what the computational tool actually did, what rules the publication has, and whether its role is described accurately.

Correcting grammar or translating a completed article differs from drafting paragraphs based on someone’s ideas. If the tool also proposes arguments, examples, or conclusions, its role is different again. When it drafts substantial passages, describing its contribution merely as “proofreading” would misrepresent the process.

A person publishing an article under their own name should understand and endorse its arguments, verify factual claims and references, review the final text carefully, and take responsibility for it. When submitting to a journal, newspaper, or other publication, that person must also follow the publication’s rules on AI use.

A disclosure can be brief, provided it is accurate. For example: “I developed the article’s main ideas and used ChatGPT to help organize the structure, draft some passages, and edit the language. I reviewed the final text.” If the tool only edited the language, say only that. If it drafted most of the text, say so.

The Schnell controversy shows why the writing process matters more than a detector score. The Dartmouth reported high “AI-written” scores for several of his works. Schnell said that he developed the ideas and drafts, used AI for language assistance, and took responsibility for the final text. A detector cannot reconstruct the entire writing process from a finished article or measure a person’s intellectual contribution. A clear account of what each participant did is more useful than a disputed percentage.

Keeping outlines, drafts, and revision records can help explain one’s contribution if questions arise later. The simplest principle is: Describe the tool’s role as it actually occurred.

About this article: The person publishing it raised the question and decided to share the response. ChatGPT drafted the text. It should not be presented as an article personally written by the person publishing it.

Sources: The Dartmouth’s reporting on Schnell; ICMJE guidance on AI use in publications.