Anthropic宣布将在其Claude模型生成的所有内容中嵌入不可见的机器可读水印，以符合欧盟《人工智慧法案》的要求。该法案规定，模型提供商必须对合成音讯、图像、影片或文字进行标记，使其可被机器侦测为AI生成内容，否则将面临高达年营业额百分之三的罚款。这项水印技术名为SynthID，由Google开发，透过在Claude的用词和措辞选择中留下人类难以察觉但机器可识别的模式来运作。

开发者Guillaume Meyer在Anthropic公告后不到一天便发布了移除水印的工具，该工具在GitHub上迅速走红，在X平台上被收藏超过两万次，并吸引了上百名贡献者。Meyer的方法利用不带水印的大型语言模型生成多个改写版本，替换同义词并略微重组内容。其他开发者也各自开发了移除工具，例如软体工程师Erik Hughes仅用十五分钟便利用Claude本身打造了一款工具，而牛津大学访问学者Leon Chlon则提出透过翻译成语义差异极大的语言再翻译回来的方法。（关键数字：20,000, 100, 190, a day）

Meyer表示他并非反对透明度或内容归属，但认为水印本身是一个糟糕的解决方案，因为存在误判风险，且无法区分轻度或重度AI使用。他担心雇主可能因水印侦测结果不公平地拒绝求职者，或研究人员仅因侦测器标记就被过度指控使用AI。Anthropic自身也承认水印只能生成文字经Claude处理的机率值，且经过大幅编辑、改写或翻译的内容可能不会携带水印。Anthropic计划发布文字侦测API，届时开发者才能验证其移除方法是否真正有效。

Anthropic announced it would embed invisible, machine-readable watermarks in all content generated by its Claude models to comply with the EU AI Act, which requires providers to label synthetic content or face fines of up to three percent of annual turnover. The watermarking technique, called SynthID, was developed by Google and works by leaving a pattern in Claude's word and phrase choices that is imperceptible to humans but detectable by machines. Some users worry this could degrade the quality of Claude's responses, though Anthropic insists it will not affect meaning, quality, or readability.

Developer Guillaume Meyer published a removal tool within a day of Anthropic's announcement, which quickly went viral on GitHub, was bookmarked over 20,000 times on X, and attracted more than 100 contributors. His method uses a non-watermarking large language model to generate multiple rewrites with synonym substitutions and content reorganization. Other developers have also created their own removal approaches, including translating text through languages with very different semantics and using Claude itself to strip invisible characters and reorder sentences. However, these methods rely on other models not implementing watermarks, which is uncertain given that 190 organizations have signed the EU's transparency code of practice.

Meyer argues that while he supports transparency and content attribution, watermarking is a flawed solution due to the risk of false positives and its inability to distinguish between light and heavy AI usage. He worries that employers could unfairly reject candidates or researchers could face overblown accusations based on probabilistic detection results. Anthropic itself acknowledged that heavily edited, paraphrased, or translated content might not carry a watermark, and plans to release a text-detection API soon. As one industry observer noted, it is unlikely that any watermark can withstand all circumvention attempts.