Skip to main content
AI Watermark Removal

Independent evidence tracker

Which AI watermark claims are confirmed, contested, or unsupported

And which you can test yourself. Anthropic says Claude watermarks its text and has published no way to check. Google's own pages disagree with each other about Gemini. OpenAI still calls text provenance a goal. This site tracks what each provider actually documents, runs its own tests where a question can be measured, and says plainly what nobody outside those companies can verify.

Open the status databaseRead the original researchAll sources verified since 2026-09-03
New · Sept 4, 2026Anthropic told Claude for Work administrators that Claude Opus 5 gets the text watermark from Sept 9, 2026, a detail its public help page has not been updated to match as of Sept 5, 2026.Read the breakdown

Start with your problem

  • 01

    I pasted Claude text somewhere and something looks off

    Scan it for invisible characters in the browser. Nothing is uploaded.

    Open the checker
  • 02

    I need to know whether a specific provider marks its output

    18 product surfaces across 11 providers, each cell sourced and dated.

    Open the status database
  • 03

    I have an EU AI Act compliance question

    Article 50 splits into a provider marking duty and a deployer disclosure duty. They are not the same.

    Article 50 explained
  • 04

    I have a file and want to know if it carries provenance data

    C2PA manifests and embedded watermarks fail in opposite ways. Check for both.

    What survives what
  • 05

    Someone told me my writing was flagged as AI

    A detector score is not a watermark. It is a guess about style, and it is wrong often enough to matter.

    False positives

Figure 1

The five things people mean by AI watermark

Ordered by who can actually check them. Only the top two are verifiable by an ordinary person, and the one below the line is not a watermark at all.

Visible label

Painted on the pixels

Disclosure rules

On the image itself, or printed beside the text

Anyone

A crop, or a screenshot of the middle

C2PA metadata

A signed manifest, attached

How manifests work

Alongside the file, not inside the picture

Anyone

Almost any upload, re-encode, or screenshot

Embedded media watermark

Inside the pixels

SynthID across media

In the pixels or the samples themselves

Gated

Heavy compression, cropping, regeneration

Statistical text watermark

In the word choices

How detection scores work

In which words the model picked, nowhere visible

Nobody

Heavy rewriting, translation, short samples

Below this line it is not a watermark

AI detector score

Nowhere. It is a guess

Why this is different

Nowhere at all. A classifier's opinion about prose

Anyone, badly

Nothing. It was never a signal to begin with

Method Compiled from provider documentation and the C2PA specification. Order reflects who can verify a signal, not how strong it is.

Checked 2026-08-13

Provider status snapshot

All 18 surfaces
Text and metadata watermarking status for six of the most-asked-about provider surfaces.
SurfaceTextMetadataPublic detectorMechanism publishedVerified
AnthropicClaude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, plus supported models through AWS, Google Cloud and Microsoft Foundry, where the text watermark carries through and C2PA file metadata applies only where that platform offers Claude's file generationYesPartialPartialYes2026-09-05
GoogleGemini app and web experience (consumer)ContestedYesPartialYes2026-09-04
GoogleGemini API, AI Studio and Antigravity (developer)ContestedYesPartialYes2026-09-04
OpenAIChatGPT and API text outputNoN/AN/AN/A2026-09-03
xAIGrok outputUnknownUnknownNoPartial2026-09-04
DeepSeekchat.deepseek.com and Open PlatformNoUnknownNoN/A2026-09-04

Original research

The Lab

  • Invisible-character census in Claude output

    Original data

    No. Across 128 outputs and 26,093 words from four Claude models, zero zero-width or bidirectional-format characters appeared. That now includes Fable 5.1, the first model Anthropic confirms carries the watermark.

    Sample
    128 outputs · 26,093 words · 4 models
    Cannot answer
    Whether Anthropic's statistical text watermark is present. Its only detector is an Anthropic API in private preview, limited today to eligible organizations under EU law and not open to this lab, so nothing we or you can run measures that. The corpus was also captured before 2026-09-09, the date Anthropic's 2026-09-04 administrator notice schedules Claude Opus 5 marking from, a date its public help page had not confirmed as of 2026-09-05.

  • Which invisible characters survive ordinary software

    Original data

    The truly invisible ones survive almost everything; only the space-like ones get cleaned up. 12 of 18 characters came through every transformation that was not deliberately trying to remove them.

    Sample
    18 characters × 12 transformations
    Cannot answer
    Anything about Google Docs, Word, or Notion. Those pipelines cannot be run and verified here, so the matrix covers only transformations that execute in code.