The watermark, in three facts
What Anthropic actually shipped, without the myths.
Claude now watermarks AI text. Here is what that actually means.
In August 2026 Anthropic announced that new Claude models embed a machine readable watermark in the text they generate. It survives copy and paste, applies worldwide, and follows the transparency rules of the EU AI Act.
The surprising part: the watermark contains no hidden characters at all. Paste your text below to see every invisible character it does contain, clean them out, and learn how the real watermark works.
Inspect your text
Visualizes invisible Unicode, special spaces, direction marks and typography artifacts. One click gives you a cleaned copy.
Your text, made visible
Characters found
Cleaned text
Important: Claude's actual watermark is a statistical pattern in word choice. It contains no hidden characters, so no character level tool (this one included) can detect or remove it. What this inspector finds are invisible Unicode characters and formatting artifacts, which are a separate and fixable issue.
- Aug 2026 Claude models launched from August 2, 2026 carry the watermark from day one Anthropic announcement
- Art. 50(2) the EU AI Act transparency article behind the change, applied worldwide EU AI Act Code of Practice
What the inspector checks
Four categories of characters that hide in copied text.
-
Check 1
Invisible characters
Zero width spaces and joiners, soft hyphens, word joiners, variation selectors and Unicode tag characters. Completely invisible, and a common source of broken code, failed searches and messy diffs.
-
Check 2
Special spaces
Non breaking spaces, narrow no break spaces, hair spaces and the rest of the Unicode space family. They look like normal spaces but behave differently.
-
Check 3
Direction marks
Bidirectional control characters that can silently reorder or disguise text. Rare in honest documents, and worth removing on sight.
-
Check 4
Typography
Smart quotes, en and em dashes and ellipsis characters. Perfectly legitimate typography, flagged separately with optional one click normalization to plain characters.
Character cleaners vs the real watermark
Most tools sold as watermark removers clean hidden characters. Claude's watermark is a different animal.
| What character cleaners find | How the Claude watermark works |
|---|---|
| Zero width spaces and joiners | No hidden characters at all |
| Special Unicode spaces | A statistical bias in word choice |
| Individual suspicious characters | A pattern spread across the whole text |
| Removable with find and replace | Only heavy rewriting degrades the signal |
| Detectable by any script, including this one | Verifiable only with detection tools Anthropic will publish |
Private by design
Runs entirely in your browser
Nothing you paste is uploaded, logged or stored. The inspector is plain client side JavaScript with no backend at all.
No signup. No tracking. No text leaves your device.
Beyond the tool
Building something with the Claude API?
I integrate AI APIs into WordPress and custom sites. Fixed scope, fixed price, delivered in a sprint.
See API integration servicesBackground reading
How AI text watermarking actually works
A deeper dive on statistical watermarks, the EU AI Act and what survives editing.
Read the post