What a Claude text watermark is
A text watermark is a statistical pattern introduced during generation, not hidden Unicode, metadata embedded in a pasted paragraph, or a list of special words. General watermark schemes alter selection among plausible next tokens using a context-dependent rule, then aggregate evidence across many selections.
The red-green-list proposal is a useful public example of this family of methods. It should not be treated as a disclosed Claude implementation. Watermark for Large Language Models
A high-level generation-time concept, not a claim about Anthropic's private implementation.
Why normal AI detection is different
Generic AI detectors look for observable language patterns or similarity to a labelled data distribution. They can be useful for diagnostics, but they do not recover a provider's secret key, prove an author, or turn text statistics into official watermark verification.
Anthropic's public transparency material discusses watermarking developments but does not publish a Claude detection API, key, or production scoring format for this app. Anthropic transparency documentation
What this experimental simulation does
The detector in this application is claude-wm-sim-v1. It produces an experimental simulation from observable sequence distributions, n-gram features, token transitions, and overlapping windows. A normal English prose passage receives Strong, Moderate, or Weak experimental signal; a short, code-heavy, quote-heavy, or unsupported-language sample receives Inconclusive.
The score is not a watermark probability or an Anthropic confidence value. It is a local simulation score with limited reference confidence, shown to support comparison before and after one deep transformation.
How deep transformation works
Transform Text asks the existing rewrite provider to internally extract a factual content blueprint and compose a new passage from that meaning. The process can rebuild sentence and paragraph structure, change transitions and openings, and make wording more independent without dropping claims or changing material facts.
Names, dates, values, URLs, email addresses, code, citations, and technical identifiers receive deterministic protection checks. The service also measures phrase overlap, wording change, structure, and length. A shallow result can trigger one quality-focused recomposition; it never loops against the experimental detector.
Before and after experimental comparison
After a successful rewrite, the app runs one independent local simulation on the updated text. It shows the original and updated signals beside rewrite quality, fact preservation, wording independence, structural change, and visual text differences. This measures the application's experimental simulation response—not whether a proprietary watermark was definitely removed.
Editing can change text-watermark detector confidence in general. Google documents this limitation for SynthID Text, which is a separate watermarking system. Google SynthID documentation and SynthID-Text publication
Limits and appropriate use
Short, heavily quoted, code-heavy, factual, translated, or unsupported-language text has less usable evidence. The app intentionally returns Inconclusive for poor inputs. Neither an experimental signal nor a rewrite should be used alone for authorship, academic discipline, employment, legal, or other high-stakes decisions.
Ready to scan a passage?
Review sample suitability and the experimental Claude-watermark detector status.