Can you detect or remove Claude's text watermark?
Not with a generic hidden-character scan. Anthropic has identified the method as a keyed version of the SynthID-Text approach: the signal lives in Claude's word choices, not in invisible characters. Its matching detection API has not been released yet.
The method is now public, but the usable detector is not. The text checker therefore reports Claude as not testable while still checking literal Unicode and exact C2PA text credentials separately.
How Claude's text watermark works
Anthropic says its implementation is a version of the SynthID-Text approach published by Google DeepMind. During generation, a secret key and the preceding words influence the random choice among several similarly suitable next words. Across a long enough passage, those choices form a statistical pattern that a detector using the same key can score.
Nothing is appended to the response: there are no hidden characters, metadata tags, extra tokens, user identifiers, organization identifiers, or chat identifiers in the text watermark. Copying the text preserves the words—and therefore may preserve the statistical pattern—but a Unicode scanner has no literal payload to reveal.
| Claude output | Marking | What can verify it |
|---|---|---|
| Generated text | A keyed statistical pattern across Claude's word choices. | Anthropic's matching key and forthcoming detection API. |
| Supported PNG, JPEG, or SVG files | A cryptographically signed C2PA Content Credential in file metadata. | A compatible C2PA validator; this is separate from the text watermark. |
Where detection is stronger—or weaker
Longer passages give the detector more word choices to measure. Short answers, factual passages, exact expressions, and code provide fewer low-stakes choices, so the watermark can be sparse or undetectable. Light proofreading of human text may also leave too little Claude-authored material to register. A translation produced by Claude is different: Claude chooses every translated word, so Anthropic says translations carry the watermark.
| Text type | Expected signal | Why |
|---|---|---|
| Long Claude-written prose | Stronger | More eligible word choices provide more evidence. |
| Short or factual answer | Weaker | There are fewer choices without changing accuracy. |
| Code | Generally weaker | Many tokens must be exact, though comments may offer choices. |
| Light proofreading | Possibly undetectable | Most words still come from the human-authored input. |
| Claude-produced translation | Watermarked | Claude chooses the words in the translated output. |
What Check AI Watermarks reports today
| Claude-related layer | Can check here? | Result language |
|---|---|---|
| Claude SynthID-Text variant | No | Not testable without Anthropic's key or API. |
| Literal hidden characters | Yes | Exact code point and cleanup policy; never attributed to Claude. |
| C2PA 2.4 text wrapper | Yes, structurally | Exact wrapper decoded; never relabeled as Claude. |
| C2PA structured-text block | Yes, structurally | Manifest reference decoded locally; external URLs are not fetched. |
| C2PA in PNG or JPEG | Yes | Manifest, asset integrity, signature, and configured trust are checked. |
| Claude-generated SVG credential | Not in this release | SVG is outside the current upload formats. |
Can the Claude watermark be removed?
There is no attached character for this app to delete. Anthropic says light editing probably will not remove the signal completely, while a complete rewrite can. That is not deterministic cleanup: it changes the words, and until Anthropic releases its API there is no authoritative way here to confirm the result.
What a positive—or negative—result means
Anthropic says its detector can estimate the likelihood that Claude partly wrote or processed a passage. It cannot prove human authorship, identify a different AI provider, distinguish “Claude wrote this” from “Claude heavily edited this,” or trace the text to a user, organization, or chat.
A missing signal would not prove that Claude was uninvolved. The passage may be too short, contain mostly fixed factual or code tokens, come from light proofreading, have been thoroughly rewritten, or come from an older model that has not yet transitioned. A zero in this app is narrower still: no supported literal signal was found.
Primary sources and related checks
- Anthropic: How Claude marks AI-generated content
- Nature: Scalable watermarking for identifying large language model outputs
- Why keyed SynthID detectors are provider-specific
- How C2PA image validation works
Frequently asked questions
Does Claude add a watermark to generated text?
Anthropic says future Claude models will use a keyed version of the SynthID-Text approach, with older models transitioning over the coming months. It is applying the watermark globally at launch.
Can a hidden-character checker detect Claude's watermark?
No. Anthropic explicitly says nothing is added to the text and there are no hidden characters. Detection needs Anthropic's matching key or its forthcoming detection API.
Can this tool remove the Claude text watermark?
No verified Claude-specific removal is available here. The signal is distributed through Claude's word choices rather than attached as a removable character. Rewriting may weaken it, but changes the text and cannot currently be verified without Anthropic's detector.
What would a detected Claude watermark prove?
Only that Claude was likely involved in writing or processing the passage. It does not prove authorship, ownership, a specific user, organization, or chat, and cannot distinguish writing from heavy editing.