规范 · 1.2.0
它如何被定义,以及它如何与其他一切对话。
这是一份声明标准,不是一份证明标准。它无法由密码学来验证,也并不打算做到这一点。署名同样做不到。
判定流程
按顺序回答五个问题。第一个使流程终止的回答会给出级别。
| # | 问题 | 是 | 否 |
|---|---|---|---|
| 1 | 这件作品的制作过程中,使用过生成式 AI 吗? | 下一题 | 0 |
| 2 | AI 是否产出了出现在发布作品中的新内容——而不只是机械地处理你已经做好的东西? | 下一题 | 1 |
| 3 | 最终成品的大部分是你自己做的吗——大部分文字、像素或声音? | 2 | 下一题 |
| 4 | 实质是你的吗——知识、数据、论证、方向? | 3 | 下一题 |
| 5 | 发布之前,有人类审阅过它,并为它所说的内容负责吗? | 4 | 5 |
局部
可选。一个数字描述作品的整体。当其中某些部分差别很大时,就把它们分别声明出来。徽章上永远只出现一个数字;细分放在链接背后。
主级别规则。 主级别描述的是承载作品意义的那一部分。理性的读者不会认为对作品具有实质影响的装饰性素材,不决定主级别——把它们声明在这里。
可声明的局部: textimageaudiovideocodedata
{
"aiUsageScale": "1.0",
"level": 3,
"surfaces": { "text": 3, "image": 5, "audio": 0 }
} 边界情形
下面每一条,都来自试图把决策树问垮的努力。在这里给出定论,是为了让两个诚实的人得到同一个数字。
- 翻译
-
忠实的翻译沿用原作的级别,并附上一条翻译说明。用模型翻译一篇第 1 级的文章,发布出来仍是第 1 级,注明由 AI 翻译。
实质与结构没有改变。惩罚翻译只会让世界上的文字更难被读到,而那恰恰与本标准的初衷背道而驰。
- 转写
-
把你录下的讲话转成文字,属于机械性处理。它不会提高级别。
话是你说的。模型只是把它记了下来。
- 实质来自 AI,文字出自你
-
如果模型想出了实质性决定作品说什么的东西——论点、论证、结构——那它就产出了留在作品里的新材料,即使每个字都是你写的。那是第 2 级或以上,绝不是第 1 级。反馈、批评以及经你核实的事实资料搜集仍是第 1 级:编辑的建议不会让编辑变成作者。
想法就是材料。用自己的话复述并不能取消它们被想出来这件事——而第 1 级的承诺——模型什么也没想出来——必须说到做到。
- 装饰性素材
-
一篇人写的文章配上一张生成的题图,不会让这篇文章变成第 4 级。请在「局部」中声明这张图。
主级别描述的是承载意义的那一部分。否则,一张图库配图就会吞掉整份声明,而人们会干脆不再声明。
- 衍生作品
-
声明你自己那部分贡献的级别;若来源作品已经声明过级别,就把它一并注明。
- 实时生成
-
按需生成、未经审阅就呈现给用户的内容是第 5 级——哪怕系统提示词是由人精心写就的。
- 非生成式机器学习
-
自动对焦、降噪、放大、色彩匹配和分类,凡是不凭空造出任何东西的,都不是生成式 AI。它们不会让你离开第 0 级。
如果你怀疑某个声明
声明是可以公开证伪的。核对链接和元数据;把声明的级别与作品对照,若有出入,再与「局部」细分对照;然后去问作者——在评论里、在评审里、在 issue 里。大多数出入是对边界的诚实解读,而上面的边界情形能解决其中大多数。
如果声明是明知故犯的假话,这套标准的回答是能规模化的那一种:声明曾是公开的、可链接的,如今被公开证伪。这里刻意没有检测器,也没有登记处。声誉就是执行机制——和署名一模一样。
互操作性
这把标尺与溯源标准相辅相成。它映射到既有的 IPTC 词汇,并试验 W3C AI Content Disclosure Community Group 正在讨论的候选语法;该语法并非 W3C 标准。
W3C AI Content Disclosure Community Group
The W3C AI Content Disclosure Community Group is discussing candidate syntax and a four-part model, but has not published a W3C standard. This scale currently emits experimental ai-disclosure metadata aligned with that model; consumers must not treat it as standardised W3C markup.
IPTC Digital Source Type NewsCodes
Every level maps to an IPTC term, so a declaration can travel into C2PA manifests, XMP, and the metadata Meta, Pinterest, and Google already read. The mapping is lossy in one direction only: IPTC cannot tell Levels 3, 4, and 5 apart.
C2PA / Content Credentials
Complementary, not competing. C2PA provides tamper-evident, signed provenance about an asset and its processing history. Its core specification does not attribute content to individuals or organisations, and provenance alone cannot determine whose thinking shaped a work. This scale is a declaration about contribution and review, designed to work for web text as well as media files.
EU AI Act (Regulation (EU) 2024/1689), Article 50
Article 50(4) switches the disclosure duty for public-interest text off the moment a human reviews it and holds editorial responsibility — the same question this scale asks to separate Level 4 from Level 5. The Code of Practice publishes three EU icons, and specifies the basic one as a first layer expecting an interactive second layer behind it. A declaration on this scale is that second layer: a mark anyone can read, linked to a public definition. The icons are free to use and require no attribution; using them is optional, while the Article 50 duties are not.
Measures for Labeling AI-Generated Synthetic Content, and GB 45438-2025
China requires two layers at once: an explicit label a reader can see, and an implicit label in the metadata carrying the nature of the content — confirmed, possible or suspected AI-generated — plus the service provider and a content ID. That two-layer shape is what a badge plus ai-usage metadata already produces. The duties fall on internet information service providers and content distribution platforms, not on individual authors.
California AI Transparency Act (SB 942, as amended by AB 853)
California pairs a manifest disclosure on the surface with a latent disclosure inside the asset: provider name, system name and version, a timestamp, and an identifier linking the content back to the system that made it. AB 853 moved the operative date to 2 August 2026 to line up with Article 50; hosting-platform duties follow on 1 January 2027. The duties fall on large generative-AI providers and platforms, not on individual publishers.
级别映射
| # | Experimental ai-disclosure | IPTC |
|---|---|---|
| 0 | none | digitalCreation |
| 1 | ai-assisted | algorithmicallyEnhanced |
| 2 | ai-assisted | compositeWithTrainedAlgorithmicMedia |
| 3 | ai-generated | trainedAlgorithmicMedia |
| 4 | ai-generated | trainedAlgorithmicMedia |
| 5 | autonomous | trainedAlgorithmicMedia |
这套映射只在一个方向上有损,而这处损失正是关键所在:IPTC 分不清第 3、4、5 级。它没有词来表达实质出自谁,也没有词来表达是否有人读过。而这恰恰是读者真正在意的两个问题,也正是这把标尺存在的理由。
机器可读形式
权威英文定义集中在一个文件中,并用于生成徽章和机器可读输出。本地化的说明文案另行维护,并接受结构一致性检查。
GET /zh/levels.json · CC0
<!-- AI Usage Scale proposal; custom metadata, not a registered standard -->
<meta name="ai-usage" content="3">
<meta name="ai-usage-standard" content="https://usagescale.org">
<!-- Experimental W3C Community Group alignment; not a W3C standard -->
<meta name="ai-disclosure" content="ai-generated">
<link rel="ai-disclosure" href="https://usagescale.org/3"> 网站也可以在 well-known 路径发布全站声明,爬虫无需提示就知道去那里查看:
GET /.well-known/ai-disclosure.json
{
"aiUsageScale": "1.0",
"level": 3,
"surfaces": { "text": 3, "image": 0 },
"definition": "https://usagescale.org/3"
} 前人的工作
这里没有任何前所未有的东西,假装有,本身就是一种小小的不诚实。
AI Assessment Scale (AIAS)
Five non-hierarchical levels, adopted by hundreds of institutions in 30+ languages. AIAS describes what a student is permitted to do. This scale describes what an author did. The debt is direct: the principle that no level ranks above another is theirs, and it is the most important rule here.
Creative Commons
The three-layer model — a mark anyone can read, a page that explains it, a specification that pins it down — is theirs.
Human Provenance in Film
Three tiers, free, CC BY 4.0. Proof that a graded standard can launch in a hostile industry.