Sources

S1 — GIJN《记者识别 AI 生成内容指南》

URL: https://gijn.org/resource/guide-detecting-ai-generated-content/

• 权威: 权威教程/书
• 支撑: 支撑“没有单一铁证、要叠加多个信号”的核心心态，以及手指/文字等经典破绽正在消退。
• 关键事实: Treat every tell as a hint, not proof. Classic giveaways like malformed hands and garbled text have become far less reliable as models improve; combine multiple signals and verify context.

S2 — Kellogg Insight：一张照片是 AI 生成的 5 个迹象

URL: https://insight.kellogg.northwestern.edu/article/ai-photos-identification

• 权威: 权威教程/书
• 支撑: 支撑第 2、3 节看图破绽清单：解剖、饰品交界、皮肤蜡感、光影不一致、背景拼贴。
• 关键事实: Clues include waxy/shiny skin, oversaturated color, a face that is “a little too perfect”, mismatched lighting between face and background, smudgy patches, and backgrounds patched together from different scenes.

S3 — C2PA 官方：What is C2PA

URL: https://c2pa.ai/what-is-c2pa

• 权威: 官方文档
• 支撑: 支撑第 5 节“内容凭证（Content Credentials）= 图片的营养标签”“C2PA 本身不检测 AI”。
• 关键事实: C2PA (Content Credentials) is tamper-evident provenance metadata; it records a signer’s assertion about whether AI was used. It does NOT itself detect AI or deepfakes, and a missing credential says nothing definitive.

S4 — Content Authenticity Initiative：How it works

URL: https://contentauthenticity.org/how-it-works

• 权威: 官方文档
• 支撑: 支撑“凭证可被截图/上传剥离”“可用 Verify 网站查看签名者与是否标注 AI”。
• 关键事实: Content Credentials travel with a file and can be inspected with a verifier to see who made it and what tools/AI were involved; metadata can be stripped by screenshots or re-uploads.

S5 — 中央网信办：《人工智能生成合成内容标识办法》答记者问

URL: https://www.cac.gov.cn/2025-03/14/c_1743654685896173.htm

• 权威: 官方文档
• 支撑: 支撑第 5 节“中国新规：显式标识+隐式标识”“2025 年 9 月 1 日施行”。
• 关键事实: 办法规定标识分两类——显式标识（文字/声音/图形，用户可明显感知）与隐式标识（嵌入文件元数据、不易感知）；自 2025 年 9 月 1 日起施行。

S6 — 四部门联合发布《人工智能生成合成内容标识办法》

URL: https://www.cac.gov.cn/2025-03/14/c_1743654685899683.htm

• 权威: 官方文档
• 支撑: 支撑“服务提供者须加显式标识、平台须核验并对疑似内容加风险提示、禁止删除篡改标识”。
• 关键事实: 服务提供者应对文本/图片/视频等添加显式标识并在元数据中加隐式标识；传播平台须核验标识、对未标识或疑似生成内容添加风险提示；任何人不得恶意删除、篡改、伪造、隐匿标识。

S7 — Google DeepMind：SynthID

URL: https://deepmind.google/models/synthid/

• 权威: 官方文档
• 支撑: 支撑第 5 节“隐形水印”“SynthID 把标记藏进像素、肉眼看不见”。
• 关键事实: SynthID embeds an invisible, machine-readable watermark directly into AI-generated images, audio, text and video at generation time, designed to remain detectable through common transformations.

S8 — Google 官方博客：SynthID Detector

URL: https://blog.google/innovation-and-ai/products/google-synthid-ai-content-detector/

• 权威: 官方文档
• 支撑: 支撑“水印是厂商专属、只认自家：Google 工具查不出别家 AI”。
• 关键事实: SynthID Detector identifies content watermarked with Google’s own AI tools (Gemini, Imagen, Lyria, Veo); it does not detect content from other companies’ models.

S9 — University of San Diego 图书馆：AI 检测器的误报与漏报

URL: https://lawlibguides.sandiego.edu/c.php?g=1443311&p=10721367

• 权威: 权威教程/书
• 支撑: 支撑第 5 节“检测器不可全信”“会把真人写的判成 AI（假阳性），对非英语母语者尤甚”。
• 关键事实: AI text detectors produce both false positives (human text flagged as AI) and false negatives; studies found detectors misclassified a large share of essays by non-native English speakers as AI-generated. OpenAI withdrew its own classifier for low accuracy.

S10 — 华盛顿邮报：如何识别 ChatGPT 文本（破折号等迹象）

URL: https://www.washingtonpost.com/technology/interactive/2025/how-detect-chatgpt-em-dash/

• 权威: 权威教程/书
• 支撑: 支撑第 4 节“AI 文章的味道”：破折号偏好、套话开头、整齐排版、不敢下结论。
• 关键事实: A Washington Post analysis of 330,000 ChatGPT messages found telltale language patterns—heavy em-dash use, favorite emoji, and stock phrasing—though none is proof on its own.

S11 — The Conversation：SynthID 与 AI 水印真的有用吗

URL: https://theconversation.com/googles-synthid-is-the-latest-tool-for-catching-ai-made-content-what-is-ai-watermarking-and-does-it-work-257637

• 权威: 博客
• 支撑: 支撑“水印不是万能、可被破坏或绕过；生态碎片化，需多个工具”。
• 关键事实: Watermarking like SynthID is not foolproof—it can be weakened or bypassed by heavy edits, and different companies use incompatible schemes, so no single detector covers everything.