Anthropic is moving Claude toward built-in AI content labeling. Starting in August 2026, new Claude models will include machine-readable marks for generated content, following Anthropic's signing of the EU AI Act Code of Practice on transparency for AI-generated content.
The change is not limited to Europe. According to the source, Claude models that launch in the EU on or after August 2, 2026, will include the labeling, and Anthropic plans to apply the requirement globally across Claude products.
What Anthropic Plans To Label
The rollout covers two broad types of output: text and files. For generated text, Claude will apply an invisible watermark. For supported files, including .svg, .png, and .jpg images, Claude will attach signed provenance metadata.
Anthropic says the text watermark is designed to avoid changing the meaning, quality, or readability of the output. It is applied at the model level, which means the label is meant to travel with Claude-generated text regardless of which Claude product created it.
The policy is expected to cover the API, Claude, Claude Code, Claude Cowork, and Claude Tag. The source also says text watermarks should work through cloud partners such as AWS, Google Cloud, and Microsoft Foundry, although those platforms may not support signed metadata.
Files are handled differently. Instead of an invisible text mark, supported generated files will carry digitally signed provenance metadata based on the open C2PA standard, developed by the Coalition for Content Provenance and Authenticity. The signature can show that Claude processed a file and can indicate later tampering.
Why The Global Scope Matters
The EU AI Act Code of Practice is the trigger described in the source, but Anthropic's implementation goes beyond a narrow regional compliance step. New Claude models will be labeled globally, not just when used inside the EU.
That choice matters for developers and organizations using Claude in products. Anthropic says developers that integrate Claude into their own services must determine which Article 50 requirements apply to those services. In practice, that means the model provider is adding labels, while downstream builders still need to understand their own obligations.
Existing models are not treated the same way as new launches. The source says they get a transition period under the law, while Anthropic says it is already working on retrofitting them.
Anthropic also plans to release verification tools so users and third parties can check the labels. The timing for those tools has not been announced in the source.
What A Claude Watermark Can And Cannot Prove
The most important caveat is that a detected watermark is not the same thing as proof that Claude authored the content from scratch. The source notes that people often use Claude to proofread, translate, or summarize human-written material. In those cases, the text may carry a watermark even when the ideas came from a person.
Anthropic also acknowledges that the absence of a watermark does not settle the question. There are several reasons AI-involved content may not show a detectable label:
- The model may have shipped before the watermarking rollout.
- The text may have been heavily edited or translated.
- The passage may be too short for reliable detection.
- File metadata may have been removed through format conversion or a screenshot.
That makes the watermark a signal, not a complete answer. It can help identify some Claude-processed content, but it cannot describe the full human and AI workflow behind a document.
"may persist through some editing"
That phrase from Anthropic is central to the issue. The source says copied text should keep its invisible watermark, and that the mark may survive some edits. But it also says heavy editing, translation, and format changes can remove or weaken the evidence.
How This Fits The Wider AI Detection Debate
Anthropic is not the only AI company working on content identification. The source says Google Deepmind open-sourced its SynthID watermarking system and built it into the Gemini models. SynthID adjusts probability values during token prediction to create a watermark without degrading text quality, works across languages, and struggles with text edited after generation.
The source also says OpenAI has had a text detector with 99.9 percent accuracy for about two years and has not released it. Reasons described include the ease of defeating detection through translation or rewriting, the risk of stigmatizing certain groups, and likely business concerns.
That sensitivity is especially visible in education. The source points to the risk that unreliable detectors can produce false cheating allegations. At the same time, it says there are valid reasons to understand when and how much AI was used, including concerns that heavy reliance on AI tools can weaken critical thinking and writing skills, particularly among students who use them as a shortcut rather than a learning aid.
The issue also reaches beyond classrooms. The source describes scammers enrolling fake students at US colleges, using AI to move through coursework, and collecting financial aid.
The Business Stakes For Claude
More reliable labeling could affect how some users choose AI tools. The source says Claude is popular for knowledge work, especially among school and college students, because even older models produce fairly natural prose.
If schools and universities continue to debate AI use in academic work, stronger detection signals could change Claude's appeal for users who want AI-generated writing to be hard to identify. At the same time, the same labels could make Claude more useful for organizations that need clearer provenance around AI-assisted content.
The larger question is whether these marks hold up in ordinary use. Copying, editing, translating, converting, and sharing content are normal parts of digital work. Anthropic's watermarking plan will be judged not only by whether labels are added, but by how well they remain useful after content leaves the original Claude interface.