Get in Touch
AI content toolsfrom Opace

Opace research · 26 August 2026

Claude SynthID text watermark what can be verified?

Public watermark research can be reproduced when its scheme, key and configuration are known. It does not establish that Claude production text carries the same signal or that rewriting has removed it.

Claude production verification: NOT AVAILABLE

Research evidenceBrowser-only
SourcePublic fixture boundaryReproducible inputs only
SignalsProduction claim heldNo public Claude verifier exists
Watermark scanKnown keysReal SynthID mathematics; Anthropic production keys stay private
Open the readiness lab

The current product status is unsupported. Anthropic's 14 August 2026 announcement describes watermarking for future supported Claude models and a detector planned for its API. No official production detector is available to the Opace browser tools, so there is no score, threshold or pass result to report.

01 / Public research boundary

Public experiments answer a narrower question

The Google DeepMind SynthID Text reference implementation can show how a documented scheme behaves with a known configuration and key. A reproducible record also needs the source commit, licence, configuration hash, transformation and detector version. This public implementation is not Anthropic's production verifier.

02 / Transformation evidence

Why rewriting is not proof of removal

Google's SynthID guidance says detection confidence can fall after thorough rewriting or translation, but a changed public detector response does not establish transfer to Claude production output. A rewrite may also alter names, figures, links, quotations, citations or meaning, so signal movement and factual damage need separate measurements.

03 / Production evidence

The evidence needed for a production claim

An official method would need to identify the production scheme and accept the matching input, key or verification route. The test would then need recorded before-and-after outputs, matched versions and an independently checkable result. Until that route exists, “Claude verified” and “Claude watermark removed” are unsupported claims.

04 / Opace reporting method

How Opace will report a future method

The Claude Watermark Readiness Lab keeps public-fixture evidence apart from production availability. The AI Content Integrity methodology defines the canonical statuses and prevents unavailable checks from appearing as passes.

Current status

Test public evidence without overstating the result.