Anthropic includes invisible, machine-readable watermarks in text generated by its Claude AI models. According to its August 11, 2026, announcement, the move reflects the company’s commitments under the European Union’s AI transparency guidelines. Notably, the rollout followed the implementation of the EU AI Act’s Transparency Code, which came into effect on August 2, 2026.
The watermarks are designed to help users and other systems identify Claude-produced content. The marks will not be visible in the text but will be embedded within it. So, even if users copy or paste the text, the watermark will travel with it, and it may persist through some editing.
As per Anthropic, the marks will not affect meaning, quality, or readability. The company has also committed to marking supported files with signed provenance metadata. It will use the Coalition for Content Provenance and Authenticity (C2PA) standard for those files.
Why are Watermarks for Claude-Generated Text Introduced?
Anthropic believes watermarking will make it easier to understand where content comes from and improve transparency around AI-generated content. This will help people quickly identify useful, context-rich content when seeking the information they need. Anthropic says Claude models launched on August 2, 2026, or after will support watermarking.
The markings will apply across supported Claude products, including Claude, Claude Platform (API), Claude Code, Claude Cowork, and Claude Tag. Claude’s cloud partners, including AWS, Google Cloud, and Microsoft Foundry, will also embed watermarks. Additionally, Anthropic says marking will be applied globally wherever Claude is accessible.
What Does the EU AI Act Suggest About AI Transparency?
According to Article 50 of the EU AI Act, the increased use of generative and interactive AI models has also made it difficult to differentiate between human-created and AI-generated content. This has also contributed to the rise in misinformation, fraud, manipulated content, impersonation, and deepfakes.
However, the Act sets out guidelines that AI providers and deployers must comply with to sustain trust and integrity in the information ecosystem. Here is what AI providers and deployers must be aware of:
- AI providers must ensure that individuals are clearly informed when they are interacting directly with an AI model.
- AI providers must include machine-readable watermarks which will help detect AI-generated and manipulated content.
- AI deployers must inform individuals whenever they are exposed to emotion recognition, biometric categorization tools, deepfakes, and text publications without human review or editorial control.
Claude is Adopting Two Major Ways to Mark Content
Anthropic is introducing two machine-readable methods to identify content processed by Claude.
- Embedded Watermarks for Claude-generated Text: Claude will embed an imperceptible watermark directly into generated text. Users will not see the watermark in their responses. Anthropic says it will not affect meaning, quality, or readability.
The watermark is embedded at the model level. It will therefore apply across supported Claude products and services. The mark will also travel with text after users copy and paste it elsewhere. Anthropic says it may remain detectable after some editing. However, the company has not explained how much editing can remove or weaken it.
- Signed Provenance Metadata: Anthropic will use signed provenance metadata for supported files processed by Claude. Supported formats include SVG, PNG, and JPG files. The metadata follows the Coalition for Content Provenance and Authenticity, or C2PA, standard. C2PA is an open standard designed to record information about digital content provenance. For this purpose, a valid signed metadata label can show that a file was processed by Claude.
How to Detect Watermarks for Claude-Generated Text?Â
Anthropic is developing tools that will help users detect Claude’s marks. The company also plans to support third-party detection. However, it has not released a detailed detection mechanism yet. The company will possibly publish technical documentation explaining the detection process. Reportedly, the mechanism will check whether text or files contain supported Claude marks. In this regard, a detected mark will indicate that the content may have been processed by Claude.
Anthropic cautions that its marking system has important limitations. A detected watermark does not conclusively prove that Claude created the underlying content. The absence of a watermark also does not prove that content was written without AI. Heavy editing, paraphrasing, translation, and short passages can impact successful watermark detection. Notably, file metadata can also disappear during format conversions and other processing.
The inclusion of watermarks for Claude-generated text aims to provide greater transparency as the distribution of AI-generated content becomes prevalent online. Claude users will observe the change without experiencing any visible interface changes. Instead, it will serve as another signal for tracking AI-generated content for businesses and researchers.
Stay ahead in the tech-first world with expert-led insights and the latest breakthroughs brought to you by YourTechDiet.
Recommended For You:
Claude Haiku 4.5 vs Claude Sonnet 4.5: Picking Your Perfect AI Partner