因應歐盟 AI 法透明義務，OpenAI 將在歐盟為 ChatGPT 文字加隱形浮水印
OpenAI will start watermarking ChatGPT's text in the EU
摘要分析
2026 年 10 月 5 日 OpenAI 宣布，為符合歐盟 AI 法 8 月 2 日生效的透明義務，將為歐盟使用者的 ChatGPT 與 Codex 生成文字加上隱形浮水印。技術名為 textGrain，與賓州大學、耶魯共同發表，透過細微調整用字讓浮水印在複製貼上後仍保留；OpenAI 稱不會識別使用者、模型效能無明顯變化。功能數週內向歐盟所有方案推出，全球 API 開發者可選擇啟用（預設關閉）。編輯會削弱偵測：同義詞替換 10% 字詞，偵測率從約 92% 降至 66%；偵測器僅開放核准研究者。Anthropic 已於 8 月宣布全球加浮水印。
- 歐盟 AI 法透明義務 8/2 生效，OpenAI 在歐盟加文字浮水印
- textGrain 技術靠調整用字，複製貼上仍保留
- 同義詞改寫 10% 即讓偵測率從 92% 降到 66%
- 浮水印不識別使用者，偵測器只給核准研究者
- Anthropic 8 月已先行全球加浮水印
OpenAI announced on October 5, 2026 that it will add an invisible watermark to text generated by ChatGPT and Codex for users in the European Union, to comply with the EU AI Act's transparency rules that took effect on August 2, 2026 and require AI-generated content to be marked in a machine-detectable way. The watermark, described in a technical report on a method called textGrain co-authored with University of Pennsylvania and Yale researchers, subtly shapes word choices so the mark survives copy-and-paste; OpenAI says it does not identify users and had no meaningful effect on model performance. It rolls out over coming weeks to EU users on all plans, while API developers worldwide can enable it for select models (off by default). Editing weakens detection: replacing 10% of words with synonyms dropped detection from about 92% to 66%, and short, mathematical or translated text is harder to detect. Detector access is limited to approved researchers. Anthropic announced global watermarking of Claude text in August.
- EU AI Act transparency rules in force Aug 2; OpenAI watermarks text in the EU
- textGrain shapes word choice so the mark survives copy-paste
- Replacing 10% of words cuts detection from ~92% to 66%
- Watermark does not identify users; detectors limited to approved researchers
- Anthropic began global watermarking in August
關鍵字：歐盟 AI 法浮水印透明義務OpenAIAI 生成內容
來源原文：https://techcrunch.com/2026/10/05/openai-will-start-watermarking-chatgpts-text-in-the-eu/
本頁為報導內容的摘要與分析，版權屬原媒體所有；請以原文為準。