OpenAI 将在欧盟上线 ChatGPT 文本隐形水印以遵循《人工智能法案》

OpenAI 周一在一篇官方博文中表示，将开始在欧盟地区为 ChatGPT 和 Codex 生成的文本添加隐形水印，以遵守欧盟《人工智能法案》（EU AI Act）。

欧盟《人工智能法案》的透明度规则已于 8 月 2 日生效，要求 AI 公司必须以其他系统能够识别的方式对 AI 生成的内容进行标记。

OpenAI 表示，该水印将在未来几周内向欧盟地区所有套餐的符合条件的 ChatGPT 和 Codex 用户推出，但仅限欧盟范围内。全球任何地区的 OpenAI API 开发者从今天起也可以为选定模型启用该功能；该功能默认处于关闭状态。OpenAI 指出，上线初期不会将文本水印设为全球默认开启。

该水印并非实际可见的符号，而是通过微调模型的选词逻辑来运作，从而留下人类读者无法察觉、但检测器可以捕获的特征模式。因为这种印记融入了文字本身，所以用户在复制和粘贴文本时，水印也会随之一同转移。OpenAI 表示，该水印不会识别用户身份，且在开启水印后，其模型的性能并未出现明显变化。

在发布声明的同时，OpenAI 还发布了该技术方案的技术报告，该方案名为 textGrain。这份由宾夕法尼亚大学和耶鲁大学研究人员共同撰写的报告，演示了如何使用密钥对预测的下一个词进行排序以完成句子的示例。将数百次微调累加后，检测器仅凭文本和密钥即可识别出 AI 生成的内容。

那么，通过编辑能够抹去水印吗？OpenAI 的测试表明确实可以。在一项测试中，若将 10% 的词语替换为同义词，检测率就会从约 92% 降至 66%。该公司还表示，短文本段落、数学答案以及翻译后的文本更难被检测出来。

图片来源：OpenAI（在新窗口打开）

“这些局限性促使我们决定，最初仅向获准的研究人员和专业机构开放检测器的访问权限，由他们协助我们评估其可靠性与负责任的使用方式，”OpenAI 表示。

OpenAI 还提醒道，未检测出水印“并不代表该文本是由人类创作”。文本可能过短、被深度编辑，或者来自其他公司的 AI。

“（水印）可以表明 OpenAI 系统生成或处理了段落的某一部分，但无法说明其中包含了多少人类的判断、编辑或创造力，”该公司表示。

在此次发布前两个月，Anthropic 宣布将为其模型 Claude 生成的文本添加水印，并且是在全球范围内实施。这一举措引来了部分 Claude 用户的强烈反对，他们认为自己提供了“指令、上下文和决策”，而 Claude 仅仅是“工具”。

据《华尔街日报》2024 年报道，OpenAI 早前就已研发出文本水印技术，但一直推迟发布，部分原因是担心用户会流向不加水印的竞争对手产品。

目前，Anthropic、Google、Meta、微软和 OpenAI 均已承诺遵守欧盟关于 AI 生成内容的实践准则。

OpenAI will start adding an invisible watermark to text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act, the company said Monday in a blog post.

The EU AI Act’s transparency rules, which took effect on August 2, require AI companies to mark AI-generated content in a way other systems can identify.

OpenAI said the watermark will roll out over the coming weeks to eligible ChatGPT and Codex users on all plans, but only in the EU. Developers using OpenAI’s API anywhere in the world can turn it on for select models starting today; it’s off by default. OpenAI said it is not making text watermarking a global default at launch.

The watermark is not an actual symbol, but works by subtly shaping the model’s word choices, leaving a pattern readers can’t see, but a detector can pick up. Because it lives in the words themselves, it travels with the text when it’s copied and pasted. OpenAI said the watermark doesn’t identify the user, and that it saw no meaningful change in its models’ performance with it switched on.

OpenAI also published a technical report for its method, called textGrain, alongside the announcement. Co-written with researchers from the University of Pennsylvania and Yale, it walks through an example of using a secret key to sort next-word predictions to finish the sentence. Add hundreds of these nudges together, and the detector can spot AI-generated content using only the text and the key.

Can the watermark be removed by editing? OpenAI’s tests suggest yes. In one test, replacing 10% of words with synonyms dropped detection from about 92% to 66%. The company also said short passages, math answers, and translated text are harder to detect.

Image Credits:OpenAI (opens in a new window)

“These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses,” said the company.

OpenAI also cautioned that a missing watermark “does not prove human authorship.” The text could be too short or too heavily edited, or it could come from another company’s AI.

“[Watermarks] can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it,” the company said.

The announcement comes two months after Anthropic said it would watermark text generated by Claude, a move it’s applying worldwide. That decision drew backlash from some Claude users, who argued they had supplied “the instructions, context, decisions” while Claude was just “the tool.”

OpenAI had built a text watermark before but held off on releasing it, partly over concerns that users would switch to rivals that didn’t watermark, The Wall Street Journal reported in 2024.

Anthropic, Google, Meta, Microsoft, and OpenAI are among the companies that have committed to following the EU’s code of practice on AI-generated content.