三十所学校里，只有三所给出了明确的分级框架——Johns Hopkins 三档、UNC 四档、Brown 五档。另有十所用非正式的分类，十七所完全没提。本校采用四档，命名为 AI-0 到 AI-3，短到可以印在作业单的角落。

Only three of the thirty publish an explicit graduated framework — Johns Hopkins with three levels, UNC with four, Brown with five. Ten more use informal categories; seventeen say nothing. We use four, labelled AI-0 through AI-3, short enough to print in the corner of an assignment sheet.

AI-1 之所以成立，芝加哥大学说得最清楚：

The University of Chicago states the reasoning behind AI-1 most clearly:

The principle here is that if generative AI is performing a task that other technologies (or people) currently perform, and such uses are not a violation of academic integrity, then those uses of generative AI do not violate the default policy.

University of Chicago Law School · univ/10-UChicago.md

采用该默认值的理由是可预测性，而非严格程度。将默认规则交由任课教师逐门确定，在大学环境中可行——三十所中有八所学校就是这么规定的——其前提是教师普遍会在课程说明中写明 AI 条款。该前提在高中并不稳定成立。当默认规则为「由教师确定」而教师未作说明时，规则处于空缺状态：使用可能构成违规，不使用则可能在同等要求下处于不利位置。以 AI-0 为默认值可消除该空缺——未获明确许可即不得使用，需要使用则事先询问。

The reason for this default is predictability, not severity. Leaving the default to each instructor is workable at university — eight of the thirty rely on exactly that — because instructors there generally publish an AI clause in the course policy. That condition does not hold reliably in high school. Where the default is "the instructor decides" and no instructor has decided, the rule is simply absent: using AI may be a violation, and not using it may leave you at a disadvantage under the same requirements. An AI-0 default removes that gap — without explicit permission you may not use it, and if you need to, you ask first.

当你拿不准某个用法算不算过线时，斯坦福给了一把最好用的尺子：

When you cannot tell whether a particular use crosses the line, Stanford offers the most useful yardstick:

Absent a clear statement from a course instructor, use of or consultation with generative AI shall be treated analogously to assistance from another person.

Stanford Board on Conduct Affairs · univ/01-Stanford.md

该类比的作用，是把一个缺乏经验的判断转换成一个已有判断标准的问题。多数人难以判断「AI 参与到何种程度构成违规」，但对「由他人代写这一段是否可接受」 有稳定的判断。斯坦福的表述要求沿用后一判断。

The purpose of the analogy is to convert an unfamiliar judgment into a familiar one. Few people can say how much AI involvement crosses the line, but almost everyone has a settled view on whether it would be acceptable for a classmate to write that paragraph. Stanford's formulation asks you to apply the second judgment to the first question.

采用强制而非建议，理由是执行成本：大学单门课程规模常达数百人，高中班级约三十人，逐份查阅披露声明在后者可行。可执行性是规则能否实际生效的前提。

Required rather than recommended, on grounds of enforcement cost: a university lecture may hold several hundred students, a class here about thirty, and reading thirty statements is feasible. Enforceability is a precondition for a rule taking effect at all.

AI 使用声明
工具：ChatGPT (GPT-5)
环节：初稿构思；语法检查
程度：AI-1 —— 仅用于语法纠错，未生成任何段落

AI Use Statement
Tool:    ChatGPT (GPT-5)
Where:   brainstorming the outline; grammar check
Extent:  AI-1 — grammar correction only; no text generated

三件套：什么工具、用在哪一步、用到什么程度。这三项分别有 9 所、14 所、 14 所学校要求，是跨校统计出的最小集合。采用固定格式而非自由撰写，是为降低填写成本：结构化表单只需补全字段，开放式说明则要求使用者自行决定披露到何种程度。

Three things: which tool, at which step, to what extent. Nine, fourteen, and fourteen of the thirty require these respectively — this is the minimum set the data supports. The format is fixed to lower the cost of complying: a structured form only asks you to fill in fields, whereas an open-ended statement requires you to decide for yourself how much to say.

提示词日志只在 AI-2 及以上要求。八所学校要求记录 prompt。但若全班每份作业均附提示词记录，实际上无法逐份查阅；长期不被查阅的要求会失去约束力，并连带削弱本规范其余条款的可信度。

Prompt logs are required only at AI-2 and above. Eight of the thirty ask for prompts. But if every student attaches one to every assignment, none of them can actually be read; a requirement that is never checked loses force, and weakens the credibility of the rest of this document with it.

A student cannot avoid limits on the use of generative AI that would otherwise apply by attributing their work to generative AI. Under the default policy, using generative AI to compose all or part of a paper, even if the use of generative AI is fully documented in properly placed footnotes, is a violation.

University of Chicago · univ/10-UChicago.md

华盛顿大学圣路易斯与南加州大学各有一句同义的规定。这一点上没有分歧。

WashU and USC each state the same rule in their own words. There is no disagreement on this point.

一 · 核实

1 · Verification

这是三十校中分歧最大的一条（一致度 39%），但在高中场景下取向较为明确。该条款的作用是排除一类免责主张，即以「内容由 AI 生成」为由推卸对错误的责任。若不作明确规定，本规范其余条款均可被「不知其为误」的说法规避。用的是北卡罗来纳大学的口径——该校要求学生对最终成品负 "ultimately 100% responsible" 的责任，并把它和另一句话放在一起：

This is the most contested provision in the set (39% agreement), though in a high-school setting the direction is fairly clear. Its function is to rule out one class of defence: disclaiming responsibility for an error on the ground that AI produced it. Without it, every other rule here can be evaded with "I did not know it was wrong." The wording follows UNC's, which holds students "ultimately 100% responsible" for the final product and pairs that with a second line:

AI should help you think, not think for you.

University of North Carolina at Chapel Hill · univ/25-UNC-Chapel-Hill.md

二 · 隐私与数据

2 · Privacy and data

这个选择本身并不异常——三十所学校里有十九所同样没有指定任何工具。但值得注意的是，在表态的十一所里，七所提供机构版工具、四所发布批准清单，理由都是数据保护。我们不做这件事，所以那份保护得由你自己提供。

The choice is not unusual — nineteen of the thirty likewise designate nothing. What is worth noticing is that among the eleven that do address it, seven provide an institutional tier and four publish an approved list, and in every case the reason is data protection. We are not doing that, so the protection has to come from you.

伯克利把这条边界写得最简洁：

Berkeley draws the boundary most concisely:

no personal, confidential, proprietary, or otherwise sensitive information may be entered into or generated as output.

University of California, Berkeley · univ/16-UC-Berkeley.md

而在没有任何工具经过审查的前提下，加州理工那把尺子就不是补充建议，而是本节的操作规则：

And with no tool here vetted by anyone, Caltech's yardstick is not supplementary advice — it is the working rule of this section:

Unless you are using a GenAI application that explicitly guarantees data separation and confidentiality, uploading information into GenAI tools should be considered a public disclosure.

California Institute of Technology · univ/12-Caltech.md

应将任何上传视同公开发布。提交前的判断标准是：该内容若出现在公开网页上是否可接受。不可接受，则不应输入。

Treat any upload as publication. The test before submitting is whether the content would be acceptable on a public web page. If it would not, it should not be entered.

三 · 版权

3 · Copyright

不得上传或转发受版权保护的材料、教师自制课件、购买的教辅、未获授权的电子书。在三十所学校里，提到这一条的十七所全部有明确规定，没有一所例外—— 这是整批数据里唯一零分歧的条款。

Do not upload or redistribute copyrighted material, a teacher's own slides, purchased workbooks, or unlicensed e-books. Of the thirty, seventeen address this and all seventeen have an explicit rule — the only provision in the entire dataset with zero disagreement.

这一点上三十所学校几乎没有分歧：二十七所把违规使用定性为学术诚信问题，其中十六所归入学术不端、九所明确等同抄袭。没有一所为 AI 另建一套制度。

The thirty barely disagree here: twenty-seven classify unauthorized use as an academic-integrity matter — sixteen as misconduct, nine as plagiarism outright. Not one builds a separate regime for AI.

Plagiarism includes, but is not limited to, the unauthorized use of generative artificial intelligence to create content that is submitted as one's own.

Northwestern University · univ/09-Northwestern.md

本节其余部分规定认定程序中的证据标准。仅规定学生义务而不规定认定方式的文件，其约束是单向的；载明举证要求之后，双方才受同一套规则约束。

The rest of this section sets the evidentiary standard for a finding. A document that states only student obligations binds one side; once it also states what must be shown, both sides are held to the same rules.

Current guidelines indicate that a report to the Undergraduate Honor Council cannot be based solely on an artificial intelligence detector score.

Vanderbilt University · univ/15-Vanderbilt.md

该条款对本校学生具有特殊意义。范德堡在同一份文件中指出，检测工具对 非母语英语写作者存在系统性偏差。二语学习者的书面表达通常具有若干共同特征：词汇选择趋于保守、句式结构较为规整、较多使用课堂习得的固定搭配。这些特征与检测工具用以判定「疑似机器生成」的指标高度重合，但其反映的是语言学习的阶段性状态，而非文本的实际来源。

This provision carries particular weight for students here. Vanderbilt's own guidance records that detection tools are systematically biased against non-native English writers. Second-language writing tends to share certain features: conservative word choice, regular sentence structure, frequent use of phrasing learned in class. Those features overlap closely with what detection tools read as machine-generated — but what they actually indicate is a stage of language learning, not the origin of the text.

可提交的证明材料

Evidence you can provide

写作过程本身构成证据，且无需额外操作即可留存。Google 文档与 Word 均自动保存版本历史。版本历史、草稿、笔记，以及能够口头复述自身思路，这几项合并使用，其证明力高于任何检测器分数。

The writing process is itself evidence, and it is retained without any extra step. Google Docs and Word both keep version history automatically. Version history, drafts, notes, and being able to talk through your own reasoning carry more probative weight together than any detector score.

The Council suggests using document editing history to help establish student-produced work and designing prompts that require engagement with course material and critical thinking.

Rice University Honor Council · univ/18-Rice.md

三十所学校里有二十所对「过程留证」未作表述，是沉默率最高的条款之一。本规范纳入该条款，因为它是其中少数能够为守规者提供保护、而不只用于认定违规的规则。

Twenty of the thirty say nothing about keeping process evidence — one of the most silent provisions in the set. It is included here because it is one of the few rules that can work in a compliant student's favour rather than only serve to establish a violation.

这一节处理的两个问题，本项目采集的三十份校内 AI 政策均未涉及。我们对这三十份原文做了关键词扫描：

None of the thirty campus AI policies collected here addresses the two matters in this section. We ran a keyword scan across those thirty source documents:

原因是这些政策的适用对象不同：大学政策的默认读者为已被录取的成年学生，且其英语水平已通过入学要求的检验。这三项前提对国际部高中生均不成立。

The reason lies in who those policies address. A university policy assumes a reader who is an adult, already admitted, and whose English has already satisfied an entrance requirement. None of the three holds for students here.

6.1 以英语能力为评估对象的作业

6.1 Assignments that assess English itself

在大学，AI 润色属于边缘问题：教师假定学生英语已达要求，所评估的是论证本身。国际部的情形不同——相当一部分写作作业所评估的正是英语产出能力。因此同一操作在两种场景下的性质并不相同。

At university, AI polishing is a marginal question: the instructor assumes English is already adequate and is assessing the argument. Here the situation differs — a substantial share of written work is assessed on the ability to produce English. The same operation therefore does not have the same character in the two settings.

因此本规范不以所用工具划分界限，而以所修改的对象划分。工具持续更替，该判准不随之改变。

The line is therefore drawn not at which tool was used but at what was changed. Tools are replaced continually; this test is not affected by that.

翻译需单独说明。先以中文完成、再由 AI 译为英文提交，所提交的文本未经过任何一次英语层面的组织。本规范将其归入「改内容」：其性质不是润色，而是由他人完成表达，区别仅在于所依据的原稿出自本人。

Translation needs a note of its own. Composing in Chinese and having AI render the result in English produces text that was never organised in English at any point. These guidelines classify that as changing content: in character it is not editing but expression carried out by another party, differing only in that the underlying draft is your own.

此处需要说明的是后果，而非评价。以 AI 修改后的文本呈现自身英语水平，在单次评分上可能获益；但大学阶段的写作要求与此相同或更高，且课堂讨论、限时写作等场景不提供使用外部工具的条件，也不接受事后声明。两者之间的差距会在那时显现。

What needs stating here is the consequence, not a judgment. Presenting AI-corrected text as your own level of English may raise a single grade. University writing demands are the same or higher, and settings such as seminar discussion and timed writing allow no external tool and accept no statement after the fact. The gap becomes visible at that point.

6.2 大学申请文书

6.2 College application essays

这一节的写法和全文不同。本校无权规定申请文书的 AI 规则——那些规则由申请平台和每一所目标大学各自制定，而且每年在变。所以这里只做三件事：告诉你规则在哪、本校的底线是什么、以及风险在哪里不对称。

This section works differently. This school does not get to set the AI rules for your applications — the platforms and each individual university set their own, and they change from year to year. So this section does only three things: tell you where the rules live, state this school's floor, and explain where the risk is asymmetric.

两类认定的后果并不对称。课程作业被认定违规，影响范围为该门课程的成绩与校内纪律记录；申请材料被认定违规，可能导致录取资格被撤销，入学后被发现亦可能导致学位被撤销。此外，申请材料的审查由申请平台和招生院校各自规定，不适用本规范 §5 的程序；部分平台设有书面告知和陈述机制（如 Common App 在书面通知后 10 个工作日内接受陈述），但这不等于每所大学都提供同样的复核。

The consequences of the two are not symmetrical. A finding on coursework affects that course's grade and an internal disciplinary record. A finding on an application can result in an offer being withdrawn, and if discovered after enrolment, in a degree being revoked. Further, an admissions review is governed by the application platform and each university's own rules, and the procedure in §5 does not apply. Some platforms give written notice and a chance to respond (Common App, for example, accepts a written statement within 10 business days of its written notice), but that does not mean every university offers the same review.

规则未被遵守，原因通常不是不知晓，而是知晓与实际使用发生在不同时点。本卡片供后一时点使用。

Rules are usually broken not because they are unknown but because the point of knowing and the point of use are different moments. This card is for the second of them.

AI 使用声明
工具：_______   环节：_______
程度：AI-__ —— _______

AI Use Statement
Tool: _______   Where: _______
Extent: AI-__ — _______

规则处理不了的，判例可以。每条按「情境 → 判定 → 为什么」写。

What rules can't settle, cases can. Each one gives the situation, the ruling, and the reason.

本规范的依据来自三十所美国大学公开发布的官方 AI 政策，按 12 项条款编码，共 360 格。其中 234 格有明文规定，126 格未提及；全部 234 条引文均经精确字符串比对，逐字匹配其所属原文。政策图谱中的任一单元格均可查看原文与官方链接。政策图谱其后按同一编码规则扩充至 52 所大学；本规范中的数字均以上述 30 所为准。

This document rests on the published AI policies of thirty U.S. universities, coded against twelve provisions across 360 cells. 234 cells have an explicit rule; 126 are not addressed by the source. All 234 quotations were matched verbatim against the file they came from. Any cell in the policy atlas shows the original wording and its official link. The atlas has since been extended to 52 universities under the same coding rules; the figures in this document refer to the thirty above.

说明：其一，三十校政策为 2026 年 8 月的快照，此后可能修订，本规范将随之更新。其二，对政策的分类与强度分级是本项目的判断，不代表任何一所大学的立场。

Note: First, the thirty policies are an August 2026 snapshot and may since have been revised; this document will be updated accordingly. Second, the classification and strength grading placed on them are this project's judgment and represent no institution's position.

→ 打开政策图谱，核验每一格 → Open the policy atlas and check any cell