怎么给 AI 内容加来源标识How to Label AI-Generated Content
微型指南 · 约 6 分钟Guide · about 6 min · 引用核验 2026-11-06（仅引用链接与日期） · refs checked 2026-11-06 (reference links and dates only)
来源标识分三层：看得见的提示、文件里的凭证、生成侧的水印。这篇讲三层各自加在哪、会丢在哪。Provenance has three layers: the visible notice, the credentials inside the file, and the generator-side watermark. Here is where each is added and where each gets lost.
先定显式标识的位置：图文在正文或图上写明，音视频在开头或画面里说明；只在页面角落留一行小字，读者与审核都可能看不见。
隐式标识写进文件元数据：C2PA 这类内容凭证记录「谁生成、用什么工具、改过哪几步」，但转码、截图、二次导出都可能把它整段丢掉。
分清生成侧水印与自己的标识：模型侧水印（如 SynthID）由生成方注入，不覆盖你后续编辑的部分，也不能替代你该加的标注。
义务按条文核对：《人工智能生成合成内容标识办法》对显式与隐式标识分别有要求，适用范围、例外情形以条文原文与你的业务定性为准。
按内容类型做一张检查表：图片、视频、音频、文本各需要什么标识、由哪个环节加、在哪一步验证，缺一格就会整批漏掉。
举个例子（虚构流程）：输入是一张 AI 生成的活动海报 → 处理是生成工具写入内容凭证、发布系统再叠加画面角标 → 结果判断是把成品下载回来，检查元数据里凭证是否还在、角标是否可见 → 若凭证在压缩后丢失，下一步是把校验挪到发布管道的末端，并要求渠道方保留原始文件，而不是只补一句口头说明。
Fix the position of the visible label first: text and images state it in the body or on the image, audio and video say it at the start or on screen. A single line in a page corner can miss both readers and reviewers.
Implicit labels live in file metadata: content credentials such as C2PA record who generated it, with which tool, and which steps changed it - but transcoding, screenshots and re-export can strip them entirely.
Separate generator watermarks from your own labeling: model-side marks (for example SynthID) are injected by the generating party, do not cover your later edits, and cannot replace the notice you owe.
Check obligations against the text of the rule: China's labeling Measures set separate requirements for explicit and implicit labels; scope and exceptions follow the article and how your service is classified.
Build one checklist per content type: what label an image, video, audio or text needs, which stage adds it, which step verifies it - a missing cell means a whole batch slips through.
Worked example (fictional flow): input is an AI-generated event poster -> processing writes content credentials at generation and overlays a visible mark at publish time -> the check downloads the finished file to confirm the credential is still in the metadata and the mark is visible -> if the credential disappears after compression, next step is to move validation to the end of the publishing pipeline and require channels to keep the original file, rather than adding a verbal note.
例子与判断Example and interpretation
以下是教学情境，不是本站运行或实测记录。A teaching scenario, not a measured or executed site test.
发布一张生成图片前，确认工具是否写入内容凭证、凭证里包含哪些字段、平台是否读取；凭证字段为空时，改用文字说明生成方式，不要声称已标识。Before publishing a generated image, check whether the tool writes content credentials, which fields they carry, and whether the platform reads them. If the fields are empty, state the generation method in text instead of claiming the item is labelled.
做完例子后怎么检查How to check your practice
先用自己的话解释上面的判断，再检查输入条件，最后回原文验证。说不清原因时回概念；缺日期、字段或权限时补输入；结论无出处时先撤回结论。Explain the example in your own words, inspect input constraints, then verify against the source. Revisit concepts if you cannot explain why; supply missing dates, fields or permissions; withdraw unsupported conclusions.
按这篇主题，先从“概念没分清”练习查起For this topic, start with concept confusion · 查看全部六个情境与理由See all six scenarios and their reasoning
水印与内容凭证Watermarks and Content Credentials 深度伪造Deepfake 生成式 AI 合规义务Generative AI Compliance Duties