Record the test details
Keep the scheme, public key, configuration, sample and detector version with the result.
Open source · Free · Public keys
AI Watermark & SynthID Checker
An AI watermark checker with its limits in the open: run the published SynthID-Text calculation against three public Opace demo keys, entirely in your browser. Production watermarks, Gemini's SynthID included, can only be checked by their own provider: no public AI watermark detector reads them, and this page never pretends otherwise.
Runs in your browser
Try it yourself
Try the SynthID checker yourself
Load a sample or paste your own text to run the published watermark calculation in this browser. The lab checks three public Opace test keys and shows the score for each one.
The demo detector is a 1.7 MB module (mostly the GPT-2 vocabulary). It loads once, on request, and everything then runs locally in this tab. Nothing you paste is sent anywhere.
Passage, coloured by per-token g-value
Each token shows its mean g-value across the six tournament layers. High values carry watermark evidence (green tint, solid overline), low values count against it (red tint, dotted underline), grey tokens are unscored (no full context window, or masked as repeated context).
Pick a sample to score it
Same text, different key: the signal vanishes. A watermark result is specific to the key used when the text was generated.
- Mean g
- –
- Weighted mean g
- –
- z-score
- –
- p-value
- –
- Scored positions
- –
- Key · depth
- –
Check the public test keys
Does this text match one of Opace's three public test keys?
Local in this browser. Your text is tokenised and checked in this tab. It is not stored or sent to analytics or any server.
Useful resultA match can recover the exact Opace test key used by one of this lab's controlled GPT-2 fixtures.
Expected result for other textNo match means only that the text was not written with these three Opace keys.
Result scopeThe score covers the three named Opace test keys and is separate from an authorship check.
What editing does to the signal
The fixture set includes degraded copies of two watermarked passages, scored here under their own generation keys. Truncation keeps most of the per-token signal because the score needs no document structure, while word substitution replaces the very token choices that carry the watermark, so the mean falls towards 0.5; Google's published guidance says detection confidence can drop further under thorough paraphrase or translation.
| Passage | Tokens | Scored | Mean g | z | Outcome |
|---|
Provider comparison
Is there a Claude watermark? What each provider publishes
See, source by source, what OpenAI, Google and Anthropic say about marking text and whether they offer a public text check.
Every cell verified against the provider's own page on 29 August 2026. Re-date on republish. If this date is older than the page carrying it, treat the panel as unverified.
Marks its own text?What the provider has published
Can anyone outside verify it?Whether a public check exists for text
Google Gemini Gemini app and web
Published as marked
Google says Gemini's text output carries SynthID-Text, and has done since 14 May 2024.
deepmind.google/science/synthid/ai-generated-text/
Verified 29 August 2026 · review monthly
No public text verifier
Google's public checking covers images, video and audio. Text is not an accepted input.
deepmind.google/models/synthid/
Verified 29 August 2026 · review monthly
Anthropic Claude All shipping models
Committed, not demonstrated
Anthropic says Claude models launched on or after 2 August 2026 support marking at launch, and that it is working to add marking to models released earlier.
No Claude model has launched on or after that date. Opus 5 launched 24 July 2026, Sonnet 5 on 30 June. Anthropic publishes no per-model status.
support.claude.com … how-claude-marks-ai-generated-content
Verified 29 August 2026 · review monthly
Announced, not released
Anthropic says it will soon offer a watermark detection API. It has not shipped one.
anthropic.com/news/claude-text-watermark
Verified 29 August 2026 · review monthly
About 2 August 2026. That is the date the EU AI Act's Article 50 transparency obligations began to apply. Anthropic inherited it as a legal boundary rather than choosing it as a rollout date. It is not a date on which Claude output began carrying a mark, and no such date has been published. digital-strategy.ec.europa.eu … guidelines-ai-transparency-obligations · Verified 29 August 2026
OpenAI ChatGPT Text output
Not publicly documented
OpenAI's provenance documentation covers images and audio only; no text watermark is publicly documented.
developers.openai.com/api/docs/guides/content-provenance
Verified 29 August 2026 · review monthly
No public text verifier
Its provenance check endpoint accepts image and audio files. Text is not an accepted input.
developers.openai.com/api/docs/guides/content-provenance
Verified 29 August 2026 · review monthly
Before and after editing
Compare the SynthID watermark score before and after editing.
Google's SynthID guidance says rewriting or translation can reduce detection confidence. The lab shows the score change separately from any changes to names, figures, links and quotations.
Compare every test key
See how the same passage scores under the matching key and the two comparison keys.
Protect important details
Check names, figures, links, quotations and meaning separately from the watermark score.
Source and method
Reproduce the SynthID watermark test
The Google DeepMind SynthID Text reference repository provides the published known-key calculation. The Opace research note provides the test setup and measured results, and the methodology defines the result labels used by this lab. Check a draft with the free AI content checker, or a file's signed credentials with the free AI image checker.
Public watermark-key questions
AI watermark checker questions
What does this AI watermark checker actually check?
It runs the published SynthID-Text calculation against three public Opace demo keys and shows the score for each key. It demonstrates how a text watermark check works; it does not read any provider's production watermark.
Can this SynthID checker detect SynthID in Gemini or ChatGPT text?
No, and no public AI watermark detector can. A production watermark can only be checked by its own provider with its own key, Google for SynthID, while OpenAI publishes no text watermark at all; this lab shows what a watermark check can really prove, using keys we publish ourselves.
Is there a Claude watermark detector?
Not a public one. Anthropic says Claude models launched on or after 2 August 2026 support marking at launch, no Claude model has launched since that date, and its detection API is announced rather than released, so nothing on this page can tell you whether a piece of Claude text carries a mark.
Does my text leave the browser?
No. Samples and pasted text are tokenised and scored in the current browser tab. Nothing is sent to Opace or added to analytics.
Why switch between the keys?
A matching public key finds the controlled test signal; the other keys do not. That comparison shows why a watermark result is tied to a specific key.
Where can I read the testing rules?
The AI detector methodology defines the result labels, protected-content checks and evidence required to support a claim.
Created by Opace Digital Agency
Need help putting content checks to work?
Opace designs practical AI and publishing workflows for organisations in Birmingham, across the UK and internationally.
Free · Runs in your browser · Nothing sent
Test the published calculation with public keys.
Load a sample or paste your own text, switch between the three public test keys and inspect the score for each one.