이미지: METAL LAB 생성
Summary
- Anthropic has signed the EU AI Act's transparency code of practice and will build labeling into new Claude models by default starting August 2026
- Generated text will carry an invisible watermark that survives copying, while files such as images will carry signed provenance metadata based on the C2PA standard
- The rollout extends beyond the EU to all Claude products worldwide, including the API, Claude Code, and Claude Cowork
- 규약 서명
- EU AI Act 투명성 코드 오브 프랙티스(Code of Practice) 서명
- 적용 시점
- 2026년 8월 2일 이후 EU에 출시되는 신규 Claude 모델부터 기본 탑재
- 적용 범위
- 전 세계, API·Claude·Claude Code·Claude Cowork·Claude Tag 포함
- 텍스트 라벨
- 눈에 보이지 않는 워터마크, 복사·붙여넣기 후에도 유지 가능
- 파일 라벨
- C2PA 표준 기반 서명된 출처 메타데이터, .svg·.png·.jpg 지원
- 기존 모델
- 법상 전환 기간 부여, 앤트로픽은 소급 적용 작업 진행 중
- 검증 도구
- 제3자 검증용 도구 제공 예정이나 구체적 시기는 미정
A mark that survives copying
Anthropic has begun embedding invisible markers into text written by Claude. Even if a document is copied in full and pasted into another file, the marker survives. According to the company, the watermark has no effect on the meaning, quality, or readability of the text, and "may remain present through some editing." In other words, it's a model-level signal that is never visible to the human eye.
Two-track labeling: watermarks for text, signatures for files
Anthropic is running two approaches in parallel. Text generated by Claude carries the invisible watermark described above, while files such as .svg, .png, and .jpg carry signed provenance metadata built on the C2PA (Coalition for Content Provenance and Authenticity) standard. C2PA is an international standard designed to track when and with what tools digital content such as photos and videos was created and edited, and it can also reveal whether an image has since been manipulated. The company noted that text watermarking still works through cloud partner channels such as AWS, Google Cloud, and Microsoft Foundry, but signed metadata support may not be available there.
An EU-driven obligation, applied globally
The move follows Anthropic's decision to sign the EU AI Act's code of practice on transparency. New Claude models released in the EU after August 2, 2026 will ship with this labeling built in by default. However, Anthropic has chosen not to confine the obligation to the EU alone, extending it uniformly to all products worldwide, including the API, Claude, Claude Code, Claude Cowork, and Claude Tag. Models already released are granted a legal transition period, but the company said it is working to apply the feature retroactively to existing models as well. Developers who integrate Claude into their own products must determine for themselves which provisions of Article 50 of the AI Act apply to their services.
An admission: "presence of a watermark can't be guaranteed"
Anthropic also acknowledged the limits of this approach. Because people often use Claude only for tasks like proofreading, translation, or summarization, detecting a watermark doesn't necessarily mean the ideas in the text were AI-generated. Conversely, the absence of a watermark offers no assurance either. The text could have been produced by a model released before watermarking was introduced, heavily edited or translated, too short to detect reliably, or stripped of its metadata entirely through a screenshot or format conversion. In the end, whether this technology proves genuinely useful will depend on whether the mark survives editing, restructuring, and translation.
So what actually changes
Attempts to distinguish AI-generated text from human writing have surfaced repeatedly, but have long been dogged by low detection accuracy and easy workarounds. Anthropic's move differs in that it isn't a standalone detection tool but a label embedded at the point of model release, placed on top of a legal framework created by EU regulation. Anthropic has recently been shoring up its infrastructure more broadly, including forming its own chip design team, while also addressing safety issues in succession—disclosing in early August, for instance, that a Claude model gained access to real corporate systems due to a misconfigured evaluation environment. Content provenance labeling reads as a trust-building measure in that same vein. Still, when and in what form a public watermark verification tool will be released remains to be seen.
