이미지: TechCrunch AI
Summary
- Some users pushed back against Anthropic's Claude watermark policy on Reddit
- Claims emerged that students and office workers would be caught using AI, but other users pushed back
- The watermark was rolled out across all Claude products worldwide following Anthropic's signing of the EU AI Act's transparency code of practice in August
- 논쟁 발단
- 레딧 이용자 'visionode'(계정 개설 3주)가 워터마크를 '음모'라고 주장
- 정책 도입 배경
- EU AI Act 투명성 규약(Code of Practice) 서명에 따른 조치
- 적용 범위
- 8월 2일 이후 EU 출시 신규 클로드 모델부터, 전 세계 API·Claude·Claude Code·Cowork로 확대
- 텍스트 워터마크 특징
- 복사·붙여넣기 후에도 남는 보이지 않는 표식
- 커뮤니티 반응
- 다수 레딧 이용자가 비판 게시글에 반박 댓글을 달았다
Backlash erupts on Reddit
After Anthropic began embedding invisible watermarks in Claude's outputs, some users vented their frustration on Reddit. A user named 'visionode,' whose account was only three weeks old, called the policy "a coercive conspiracy aimed at good-faith chatbot users." They wrote, "The student who just wanted help tidying up a paragraph, the journalist who asked for a summary of a 200-page transcript, the writer stuck in a creative slump who asked for a synonym—all of them come out with a digital brand on their forehead."
Another user called the watermark "unethical," asking, "AI is just a tool, so what exactly is it claiming credit for?" But the Reddit response was not sympathetic. Many replies pushed back, arguing, "That's not about claiming credit. It's about making it possible to detect the risks that AI-generated content can create."
What's going on here
This watermark stems from Anthropic's global labeling policy, which METAL LAB reported on August 11. After signing the EU AI Act's transparency code of practice, Anthropic made labeling a default feature for new Claude models launched in the EU starting August 2, and later expanded the scope beyond the EU to all products worldwide, including the API, the Claude app, Claude Code, and Claude Cowork. Text outputs carry an invisible watermark that survives copy-and-paste, while files such as images carry signed provenance metadata based on the C2PA standard.
The intent behind the policy is clear. Regulators required that AI-generated content be identifiable, and Anthropic simply embedded the watermark to meet that requirement. The issue is that the watermark ends up "exposing" behavior that was already problematic to begin with, such as plagiarism or cheating. TechCrunch pointed out that it was already unethical for a student to paste Claude's answer directly into an essay, or for a journalist to drop an unverified summary straight into an article. The watermark doesn't create a new problem so much as make an existing one harder to hide.
So what changes
Since the watermark was introduced, the risk of getting caught copying a chatbot's response verbatim has increased. Conversely, the debate has made clear that users who honestly use AI as a drafting or organizing tool face no real impact. In the end, what this policy tests isn't the watermarking technology itself, but the habits users have developed in how they've been using AI all along.
