Each finding includes a Unicode name, code point, category, count, and line-column location. You can reproduce the result locally from the same input.
Evidence-first inspection
Claude Watermark Detector & Hidden Character Check
Find concrete hidden Unicode and formatting anomalies with exact locations. The report is real and reproducible; it is not Anthropic’s unpublished statistical watermark detector.
Analyze + Clean stay local Nothing pre-selected No fake detection score
Exact Unicode inspection
Analyze first. Clean only what you select.
Analyze + Clean stay local
- 1AnalyzeFind exact UnicodeLocal
- 2CleanRemove selected typesLocal
- 3RewriteChange visible contentOff
Scope: Claude’s model-level watermark is not hidden Unicode and is not detected by this inspector.
0 characters
Inspection report
Characters found
0flagged characters
0character types
0lines checked
No inspected artifacts found
This checks known invisible Unicode and formatting controls only. It is not a detector for Claude’s model-level statistical watermark.
0 types selected
Selective Cleaner result
Review every change
Only the exact code points you selected were removed. No normalization, rewriting, or other replacement was applied.
BeforeRemoved markers
AfterCleaned output
Step 3 · Optional server processing
Rewrite is unavailable
Unlike Analyze and Clean, a configured rewrite would send your content to a model provider. No provider or server endpoint exists in this MVP.
Rewriting changes visible text or code. It may reduce a detectable statistical signal, but this tool cannot verify or guarantee that outcome.
Detection boundary
Observable evidence is not an official Claude detection result
This inspector can prove that a specific Unicode code point is present. Anthropic has not published the technical mechanism needed to reproduce its model-level watermark detection.
This tool does not assign an AI-authorship score or declare text ‘watermarked.’ A hidden character may come from an editor, website, language system, or clipboard.
How to detect a Claude watermark without inventing evidence
Anthropic says it is working to let users and third parties detect supported Claude marks, but the company has not yet published the technical detection mechanism. Until that changes, an independent page cannot responsibly reproduce an official Claude watermark result.
What this detector can verify now
The local inspector identifies concrete Unicode artifacts and reports their exact names, code points, counts, and locations. That is useful for debugging pasted text, typography, source code, and publishing workflows—but it is not an AI-authorship judgment.
How to interpret the report
- A flagged character means the exact code point exists at the reported location.
- No flagged characters means this inspector found none of its supported artifacts.
- Neither result confirms whether Claude generated or processed the text.
Anthropic also warns that a detected mark is not complete provenance proof, while a missing mark does not prove the content was never AI-generated or processed.
Clear answers, narrow claims
Frequently asked questions
Is this Anthropic’s official Claude watermark detector?
No. Anthropic says detection details will appear in forthcoming technical documentation. This independent tool checks only observable Unicode and formatting artifacts.
How can I detect a Claude watermark today?
You cannot independently reproduce Anthropic’s unpublished model-level detection mechanism from this site. You can inspect exact hidden characters now and use an official detection mechanism if Anthropic releases one later.
Do hidden characters prove that Claude wrote the text?
No. Editors, websites, clipboard transformations, typography, accessibility tools, and many other systems can introduce hidden Unicode. A finding proves only that the reported code point is present.
Does a clean scan mean there is no Claude watermark?
No. A clean scan means none of this inspector’s supported Unicode artifacts were found. Claude’s documented watermark is model-level, so absence of hidden characters is not proof that no statistical mark exists.