Anthropic to Add Invisible Watermarks to Claude Text Output
The watermarking is a probabilistic signal, not definitive proof, meaning text that a human wrote and then had Claude proofread or translate could be flagged as AI-generated, raising concerns about how detection results will be used by platforms and regulators.
Key Facts
- Anthropic will embed invisible watermarks in text generated by Claude models, applying globally including Japan and the United States.
- The watermarking is tied to the EU AI Act, which took effect on August 2, 2026, and requires generative AI providers to mark AI-generated content.
- Watermarks will apply to text generated through Claude's web app, API, Claude Code, Claude Cowork, and Claude Tag, as well as via AWS, Google Cloud, and Microsoft Foundry.
- For image files such as PNG, JPEG, and SVG, Anthropic will add C2PA-compliant provenance metadata.
- Anthropic states that the watermark indicates the possibility that Claude processed the content, not that it created it from scratch, and detection methods will be disclosed later.
Reporting from 3 sources: ASCII.jp, GIGAZINE, GameBusiness.jp.
Anthropic has announced plans to embed invisible digital watermarks into text generated by its Claude AI models, a move tied to compliance with the EU AI Act. The watermarking will apply globally, including in Japan and the United States, and will cover text produced through Claude's web app, API, Claude Code, Claude Cowork, and Claude Tag, as well as via cloud providers like AWS, Google Cloud, and Microsoft Foundry. The watermarks are designed to survive copy-pasting and some editing, allowing detection of whether text was processed by Claude. For image files such as PNG, JPEG, and SVG, Anthropic will add C2PA-compliant provenance metadata. The company notes that the watermark indicates the possibility that Claude processed the content, not that it created it from scratch. Text that is proofread, translated, or summarized by Claude may also carry the watermark. Detection methods will be disclosed in future technical documents. The EU AI Act, which took effect on August 2, 2026, requires generative AI providers to mark AI-generated content. Anthropic has signed the associated Code of Practice, along with Google, OpenAI, Meta, Microsoft, Mistral, and Cohere.
- Anthropic's watermarking is implemented at the model level, so it applies to any route that generates text, not just the chat interface.
- Claude's main models, including Fable 5, Opus 5, and Sonnet 5, were released before the EU AI Act took effect, so they are not yet obligated to carry watermarks, but Anthropic says it will add support progressively during the grace period.
- Google's SynthID-Text, which slightly biases word-selection probabilities to create a detectable statistical pattern, is a prior example of text watermarking, and Google has opened the technology to Apple, NVIDIA, and OpenAI.
- The watermark is a probabilistic signal, not definitive proof. Anthropic's support document says text that is proofread, translated, or summarized by Claude may carry a watermark, so the presence of one cannot determine who wrote the text.
- Heavily edited or paraphrased text, or text too short for statistical patterns to be detected, may not be detectable. For image files, C2PA metadata can be lost through format conversion, re-saving, or screenshots.
- The GameBusiness.jp piece flags a risk: if "AI-likeness" is linked to penalties or algorithmic disadvantages and operated as a black box, text that a human wrote and then had Claude finalize could be judged AI-generated and rejected.
- The same article notes the rise of "AI slop" and witch-hunt-like reactions where content is assumed to be AI-generated without evidence, with LLM-typical phrasing and em dashes becoming memes as "typical AI speech."
Synthesized by Yomimono from the 3 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.