How watermark removal and text rewriting work.

Text-watermark evidence is statistical, not visible markup

A generation-time text watermark can influence choices among plausible next tokens using a keyed, context-dependent rule. A compatible detector accumulates evidence over a passage; it does not locate a “watermarked word,” hidden Unicode string, or metadata field inside pasted prose. The general red-green-list construction and related schemes illustrate why a matching detector needs knowledge that is not visible in the text alone. Kirchenbauer et al.

Ordinary AI-authorship detection is different: it uses observable patterns in text to estimate similarity to a training distribution. It cannot recreate a provider's private key or production detector. Google's SynthID-Text publication describes a distinct watermark system, not a public specification for Claude. SynthID-Text publication

A trained classifier runs in the browser

The AI Detector uses the fine-tuned DACTYL BERT classifier, not a hand-written sentence-length formula. The pinned checkpoint was trained for binary AI-text classification and its model card reports results across several model and writing domains. This deployment uses a validated unsigned 8-bit ONNX export of that checkpoint so inference can run locally without a detector API call. DACTYL model card and evaluation

The full passage and its sentence windows pass through the same trained model. The document score comes from the document inference, while highlights and composition come from the individual sentence inferences. Only passages with actual high model scores appear as supporting evidence. The current detector supports English. Related shared-task work continues to document how generated-text detectors are evaluated across changing generators and domains. PAN-CLEF generated-content analysis

The result is a classifier output, not a provider attribution or a Claude watermark result. A trained detector can still make errors when text differs from its training data, is heavily edited, or is too short.

The watermark detector remains unavailable

Anthropic's transparency documentation says it has worked with industry and academia on watermarking developments and is preparing for applicable requirements. It does not provide this app with a public Claude watermark endpoint, request schema, credential type, or detection key. Anthropic transparency documentation

The Watermark Detector mode therefore stays disabled until Anthropic publishes an official integration. The app does not relabel a generic AI-writing score as a verified Claude watermark result.

Recompose from meaning, not from synonyms

Watermark removal uses one rewrite-provider request. The writer first forms an internal content blueprint covering claims, support, facts, names, values, links, terms, relationships, tone, audience, and intent. It then composes fresh prose from that blueprint rather than editing each source sentence in place.

The rewrite preserves the original tone automatically while protecting meaning, material facts, values, names, links, conditions, and technical detail. It is not guided by detector score and is never repeated until a score crosses a target.

Why a complete rewrite is used

Proofreading and light polishing retain much of the original sequence. Translation, summary, and paraphrase create a new passage, but their effect depends on the watermark design and the amount of text the detector evaluates. Anthropic's public transparency documentation describes ongoing watermarking work, not a Claude-specific removal procedure. Anthropic transparency documentation

The public evidence supports deep semantic recomposition rather than synonym replacement. A watermark reliability study found that a known watermark could persist after strong paraphrasing when enough text was available. The SynthID-Text publication separately reports that substantial LLM paraphrasing weakens its token-level signal. Watermark reliability study SynthID-Text study

The product therefore changes wording, sentence construction, transitions, and paragraph organization while preservation checks protect facts and technical details. This can change the text pattern measured by this application without making a guarantee about an undisclosed provider watermark.

Quality checks happen before the final comparison

Before a rewrite is returned, the server protects and restores numbers, dates, currencies, URLs, email addresses, code, and citations. It measures extractable factual values and named terms, phrase overlap, wording change, sentence/paragraph restructuring, and length change. These metrics report rewrite quality without rejecting valid prose because it missed an arbitrary similarity threshold.

After a successful rewrite, the browser runs the same AI-pattern check over the original and updated text. That comparison stays separate from watermark verification and never triggers an automatic second rewrite.

Extensive editing can affect text-watermark detection systems generally; Google makes this point in its SynthID documentation. Google SynthID documentation

Focused processing, clear boundaries