Anthropic and Claude checks
Claude watermark detector
Check Claude-related watermark, provenance, and AI-writing risk signals with current Anthropic support limitations.
Run a signal check
No detector model loads until you press Check.
Current support
Future Claude models will use SynthID-Text-style watermarking, but Anthropic's detection API is not public yet.
What this checks
- Anthropic says Claude text watermarking changes the source of generation randomness, not hidden characters, metadata, or extra tokens.
- Anthropic says it will offer a detection API, but public implementation details are still pending.
- Claude-generated files can carry C2PA content credentials when the file type is supported.
What this does not prove
- A Claude watermark can only indicate likely Claude involvement, not authorship or whether Claude wrote versus heavily edited the text.
- Short, factual, code-heavy, or lightly proofread passages may carry too little watermark signal to verify reliably.
- Heavy editing or complete rewriting can weaken or remove the watermark.
FAQ
- Can this prove whether content was written by AI?
No. AI Detect reports signals such as provenance metadata, watermark support status, and statistical text risk. It does not claim legal or authorship proof.
- Does text leave my browser during a local check?
Browser-default checks run local text features and provenance checks after you press Check. Server enhanced checks are separate and will be labelled before use.
- Why can a result be unknown?
Many providers do not expose public watermark detectors, and metadata can be absent or stripped. Unknown is a valid result when no reliable public signal exists.
Provider registry
Current watermark support
| Provider | Text watermark | Detection | File provenance | Tamper risk | Updated | Sources |
|---|---|---|---|---|---|---|
| Anthropic / Claudeanthropic, claude, claude-sonnet, claude-opus | plannedSynthID-Text-style statistical watermark | plannedAnthropic has described Claude text watermarking and says a detection API is planned, but no public Claude verification adapter is connected yet. | C2PAsupported_when_present | weak against rewriteClaude text watermarking is described as a statistical wording signal. Substantial rewriting can weaken it; C2PA file credentials are a separate metadata/provenance layer that can be stripped or lost. | 2026-08-17 | |
| Google / Geminigoogle, gemini, synthid | researchSynthID Text | availableSynthID Text is documented for compatible generation pipelines, but it is not a universal detector for arbitrary pasted text. | n/a | weak against rewriteSynthID-class text signals are statistical and should be treated as probabilistic. Media and file provenance can involve separate metadata or pixel-domain marks with different removal risks. | 2026-08-12 |
Removal pressure
Watermark tamper-risk registry
AI Detect uses remover research as adversarial context only. It does not run removal code, and it treats missing watermark or provenance evidence as an unknown state.
| Risk class | Target | Mechanism | Risk | How to read absence | Sources |
|---|---|---|---|---|---|
| Invisible Unicode and carrier cleanuplayer-a-invisible-unicode | Text with zero-width characters, bidi controls, tag characters, exotic spaces, or edit-based carriers. | Deterministic text normalization can remove visible-invisible carrier characters without needing a provider detector. | high | If this class of carrier is absent, the text may be clean, already normalized, or previously scrubbed.Whether obvious text-carrier artifacts are found when a detector is connected. | |
| Statistical text watermark weakeninglayer-b-statistical-rewrite | Token-sampling watermarks such as SynthID-style, Kirchenbauer green-list, keyed-Gumbel, or similar wording-distribution marks. | Substantial rewriting, paraphrase, or regeneration can move token choices away from the original marked distribution. | medium | A weak or missing statistical watermark can mean unsupported detector access, too little text, normal edits, or an intentional rewrite.Provider support status for public watermark verification. | |
| C2PA and file metadata strippingfile-provenance-metadata | Images, PDFs, Office files, HTML/Markdown, audio/video containers, and other files carrying provenance metadata. | Container cleaners can remove C2PA manifests, EXIF/XMP hints, document properties, YAML front matter, and similar metadata layers. | high | No C2PA manifest found means no local manifest was found; it does not prove the file was never AI-generated.Whether a browser-side C2PA manifest is present in supported file checks. | |
| Pixel and media watermark uncertaintymedia-pixel-watermarks | Images, generated media, or audio/video that may include imperceptible in-content watermarks. | Removal risk depends on the watermark family, transform robustness, regeneration, compression, cropping, and available verifier. | varies | Metadata absence does not answer whether an in-content watermark or soft-bound provenance exists.C2PA/provenance metadata when present. |
Reviewed source: https://github.com/guillaumemeyer/watermarks-remover at bcf497260af4.
Workflow upgrade
Need repeatable reviews?
Move from one-off checks to saved reports, batch review, and API-ready signal records when the workflow needs it.