OpenAI to Add Invisible Text Watermarks in the EU

01The Story新闻事实

OpenAI announced that it will add machine-readable, invisible watermarks to text generated by its AI models, including ChatGPT and Codex. The feature will initially be opt-in globally, but eligible outputs in the European Union will become automatically watermarked in the coming weeks under the EU AI Act, while the United States has no comparable federal requirement.OpenAI宣布，将为其人工智能模型生成的文本添加可由机器识别的不可见水印，涵盖ChatGPT和Codex。该功能最初将在全球范围内以自愿启用方式提供，但未来几周内，欧盟符合条件的输出将根据《欧盟人工智能法案》自动添加水印；美国目前没有类似的联邦要求。

The watermark will be invisible to the naked eye but detectable by machine.水印肉眼不可见，但可以通过机器检测。
OpenAI says the watermark will apply to text generated by ChatGPT and its coding agent Codex.OpenAI表示，水印将应用于ChatGPT和其编程代理Codex生成的文本。
Users globally can opt in at first, while eligible ChatGPT and Codex text outputs in the EU will receive the watermark automatically in the coming weeks.起初，全球用户都可以选择启用水印；未来几周内，欧盟符合条件的ChatGPT和Codex文本输出将自动添加水印。
Researchers and expert organizations will receive detector access on a case-by-case basis because of concerns about false positives.由于担心误判，研究人员和专业机构将按个案获得检测工具访问权限。
A Rainey Center poll cited in the article found support for AI watermarking among nearly eight in 10 people, including 82% of Democrats and 75% of Republicans.文章引用的Rainey Center民调显示，近八成受访者支持人工智能水印要求，其中民主党支持者的比例为82%，共和党支持者为75%。

The rollout illustrates how a legal requirement in the EU is producing a different AI-content transparency standard from the one currently in place in the United States.这一部署表明，欧盟的法律要求正在形成与美国现行标准不同的人工智能内容透明度规范。

024 Dimensions维度分析

Cause因果关系

Regulation drives implementation监管推动技术落地

OpenAI’s EU rollout is presented as a direct response to the EU AI Act’s requirement for a machine-readable identifier on model-generated text. The feature is therefore mandatory for eligible EU outputs but optional for users elsewhere at launch.文章将OpenAI在欧盟的部署描述为对《欧盟人工智能法案》要求的直接回应，该法案要求模型生成的文本包含可由机器读取的标识。因此，欧盟符合条件的输出必须添加水印，而其他地区的用户在推出初期可以自愿启用。

The EU AI Act requires model-made text to include a machine-readable identifier, according to the article.文章称，《欧盟人工智能法案》要求模型生成的文本包含可由机器读取的标识。
Eligible EU outputs will receive the watermark automatically, with no off switch.欧盟符合条件的输出将自动添加水印，且没有关闭选项。

Comparison对比

EU-US policy gap欧美政策差距

OpenAI’s policy differs by jurisdiction: mandatory watermarking is planned for eligible EU outputs, while the company is not applying the requirement to the United States. The article attributes this difference to the absence of a comparable US federal mandate.OpenAI将根据司法辖区采取不同做法：欧盟符合条件的输出计划强制添加水印，而美国不适用这一要求。文章将这种差异归因于美国缺乏相应的联邦强制规定。

Democratic support for AI watermarking民主党支持者对人工智能水印的支持率

Recent Rainey Center pollRainey Center近期民调82 %

Recent Rainey Center pollRainey Center近期民调75 %

The article says the United States has no federal requirement for frontier AI companies to add such watermarks.文章称，美国没有要求前沿人工智能公司添加此类水印的联邦规定。
The cited poll found that nearly eight in 10 people support AI watermarking requirements.文章引用的民调显示，近八成受访者支持人工智能水印要求。

Stakeholder impact利益相关方

Detection and verification检测与验证

OpenAI will not make its detector publicly available at launch. Instead, researchers and expert organizations may receive access case by case, reflecting the company’s concern that unrestricted use could produce false positives.OpenAI在推出初期不会向公众开放检测工具，而是按个案向研究人员和专业机构提供访问权限，原因是公司担心不受限制的使用可能产生误判。

The detector is intended to identify the invisible indicator embedded in generated text.该检测工具用于识别嵌入生成文本中的不可见标识。
OpenAI says access will support evaluation by researchers and expert organizations.OpenAI表示，访问权限将用于支持研究人员和专业机构开展评估。

Historical context历史脉络

Industry-wide direction行业整体趋势

OpenAI is not the first company cited as using machine-readable markers. The article also points to Anthropic’s similar marker and Google DeepMind’s watermarking work, although it notes that Google has announced removal options for some media in countries where watermarking is not required.据文章引用的信息，OpenAI并不是首家使用机器可读标记的公司。Anthropic也推出了类似标记，Google DeepMind则开展了水印工作；不过，文章指出，Google已宣布在不要求水印的国家为部分媒体提供移除选项。

Anthropic launched a similar machine-readable marker in August.Anthropic于8月推出了类似的机器可读标记。
Google DeepMind has been watermarking outputs for a couple of years, according to the article.文章称，Google DeepMind为输出添加水印已有数年。

03Timeline时间线

2026-10-05 OpenAI announced its invisible text-watermarking plan, according to the report.据报道，OpenAI宣布了为文本添加不可见水印的计划。

Key takeaway核心结论

The EU’s legal requirement is turning AI-generated text provenance from an optional company feature into a mandatory condition for eligible outputs in the region.欧盟的法律要求正将人工智能生成文本的来源标识，从企业自愿提供的功能变为该地区符合条件的输出必须满足的条件。

04What to Watch后续关注

How the automatic watermarking rollout for eligible EU ChatGPT and Codex outputs proceeds in the coming weeks.未来几周欧盟ChatGPT和Codex符合条件的输出将如何推进自动添加水印。
Which researchers and expert organizations receive case-by-case access to OpenAI’s detector.哪些研究人员和专业机构将按个案获得OpenAI检测工具的访问权限。
Whether the United States introduces a federal requirement comparable to the EU approach.美国是否会推出与欧盟做法相似的联邦要求。
How users and platforms respond to the availability of machine-readable indicators for AI-generated text.用户和平台将如何应对人工智能生成文本中机器可读标识的出现。

05Uncertainties不确定因素

The article does not specify which outputs will qualify as “eligible” for mandatory watermarking in the EU.文章没有说明哪些输出将被列为欧盟强制添加水印的“符合条件”内容。
The accuracy of the detector and the scale of potential false positives remain unresolved.检测工具的准确性以及可能出现误判的规模仍未确定。
The article does not state when, or whether, the United States will adopt a comparable federal policy.文章没有说明美国何时或是否会采取类似的联邦政策。

06Sources信息来源

OpenAI Is Adding Text Watermarks in the EU Because Regulation WorksOpenAI Is Adding Text Watermarks in the EU Because Regulation Works Gizmodo · 2026-10-05

Share this analysis 分享此深度分析

X Facebook LinkedIn WhatsApp