Anthropic宣布在其Claude AI产品中部署隐形水印系统，用于检测学生提交的AI生成文章[1]。该水印被嵌入生成的文本中，即使文章被复制粘贴或部分编辑后仍可保留，并可通过检测机制识别[1]。根据该公司的承诺，这一系统已纳入欧洲新法规的合规要求，将在全球范围内适用[1]。
OpenAI和Google等公司也已开发了类似的文本水印检测系统[1]。然而，这一举措面临来自用户和批评者的质疑[1]。批评人士将其比作"毒药的解药由同样获利的公司拥有和销售"[1]，担忧该技术可能影响代码质量且容易被逆向工程破解[1]。Anthropic也承认，使用Claude检查人工创作的文章可能导致这些作品被错误地标记为曾被AI接触过[1]。
Anthropic has announced the implementation of an invisible watermarking system in its Claude AI product to detect artificially generated essays submitted by students.[1] The watermark is embedded within the generated text itself, remaining intact even after copying, pasting, and partial editing, allowing detection mechanisms to identify AI-authored content.[1] The company has committed to incorporating this system as part of compliance with new European regulations, with the technology set to be deployed globally.[1]
The move mirrors similar initiatives by other major technology companies; OpenAI and Google have also developed comparable text watermarking detection systems.[1] However, the approach has drawn criticism from both users and observers. Critics have characterized the solution as "the poison's antidote owned and sold by the same companies that profit from the poison itself."[1] Additionally, Anthropic has acknowledged a potential drawback: using Claude to check human-created work could result in those works being flagged as having been in contact with AI.[1]
Skepticism about the watermarking approach extends to technical concerns as well. Users in the Claude AI community have expressed doubts about the system's effectiveness, suggesting that the watermarks could impact code quality and are vulnerable to reverse engineering attacks.[1]