How Claude text watermarks work—and what removal really requires.

What Anthropic announced

Anthropic says Claude models launched on or after August 2, 2026 support text marking at launch, while marking for earlier models is still being rolled out. The mark is applied at the model level across supported Claude products and can persist through copy and paste and some editing.

Anthropic also cautions that a detected mark would indicate likely Claude involvement, not prove authorship or misconduct. Failure to detect a mark would not prove that Claude was uninvolved.

What remains unknown

Anthropic has not released the marking algorithm, detection thresholds, calibration data, model-by-model activation dates, or a public detector. General watermark papers can explain possible designs, but they do not establish how Claude’s production system works.

That means no third-party cleaner can inspect a passage and prove that Claude’s mark was removed. A generic AI detector measures its own classifier, not Anthropic’s undisclosed mark.

Is the Claude watermark made of hidden characters?

Anthropic has not published an implementation that third parties can inspect. Its documentation describes model-level marking that can survive copy and paste; it does not identify zero-width characters as the mechanism.

Unicode cleanup can still remove zero-width spaces, bidirectional controls, unusual spaces, soft hyphens, and other pasted-text artifacts. That is useful text hygiene, but there is no evidence that it removes Claude’s documented model-level mark.

Can you remove a Claude watermark from text?

Anthropic says detection may fail after heavy editing, paraphrasing, translation, or mixing with other text. That describes a limitation of detection, not a guaranteed removal recipe. The company has not published a success threshold or public verification method.

Independent watermark research also finds that editing robustness varies by scheme, threshold, passage length, and depth of revision. A tool cannot honestly display “100% removed” without testing against Anthropic’s detector.

A substantive rewrite changes sentence structure, word choice, rhythm, and often paragraph organization. That depth also creates an editorial risk: meaning, figures, citations, or quoted language can drift. Source and result should always be compared before the rewrite is used.

What our Claude watermark rewrite tool does

Claude Watermark Removal applies a conservative Unicode cleanup, then rebuilds the visible prose through a separate rewrite process. Optional writing samples can guide cadence and vocabulary. The result page keeps the original beside the rewrite and reports factual anchors and revision depth for human review.

We do not expose a fake watermark score, claim a guaranteed detector bypass, or label Unicode artifacts as watermark bits. The output is a substantive editorial rewrite whose watermark status remains independently unverifiable.

What a future detector result would mean

A positive watermark result would suggest that Claude was involved in producing or transforming the passage. It would not establish authorship, intent, misconduct, or the absence of later human editing. A negative result would also be limited: the passage may be too short, too constrained, heavily edited, translated, or produced by a model without an active mark.

Detection should therefore be treated as one provenance signal, not a binary judgment about a person. Schools, employers, publishers, and platforms should evaluate context and other evidence before making consequential decisions.