Claude and Anthropic are named here only to describe publicly documented product behaviour. AI Watermark Center is independent and is not affiliated with Anthropic.
Current status
As of 15 August 2026, Anthropic documents machine-readable marking for supported new Claude models. Models launched in the EU on or after 2 August 2026 are described as supporting marking at launch. Anthropic says it is working to add marking to eligible older models during a legal transition period. Model-by-model coverage beyond that launch-date rule is not verified here.
Text: statistical watermark
Official documentation describes a statistical, token-choice watermark: when a supported model generates text, keyed randomness can influence which already-plausible next word is selected. Nothing extra is appended, and Anthropic states there are no hidden characters in that mechanism. Detection, according to Anthropic, depends on a key. A public detection API has been announced as forthcoming, not as a generally released third-party product we can integrate.
Files: C2PA / Content Credentials
For supported generated files such as PNG, JPG, and SVG, Anthropic documents signed C2PA provenance metadata. That metadata is a different mechanism from the text watermark: it sits in file metadata rather than in token choices. Presence of credentials is not an AI/human verdict.
What is officially documented
- Supported new models produce machine-readable marks.
- Text marking is statistical, not visible, and not hidden Unicode.
- Short, factual, proofread, and code-heavy text may carry less usable signal.
- Translations generated by Claude are described as carrying the text watermark because Claude chooses the words.
- The watermark is not described as identifying an individual user.
- A detected mark indicates Claude may have processed the content; it does not prove Claude wrote every word.
What AI Watermark Center can currently inspect
- Hidden Unicode and formatting controls in pasted text.
- Embedded EXIF, GPS, and Content Credentials in JPEG, PNG, and WebP.
What cannot currently be independently verified
Provider detection capability: Announced, not generally released. AI Watermark Center integration: Not integrated. AI Watermark Center does not operate Anthropic’s key and does not claim to detect or remove Claude’s text watermark. Independent tests of copied Unicode are pending collection.
Anthropic states that Claude’s official text watermark is not based on hidden characters. The scanner checks Unicode artifacts only.