Claude Watermark Tools Compared: What Each Can Actually Verify
Search for a Claude watermark tool and you get three quite different product categories presented as if they were interchangeable. They are not. This is what each one actually does, and the single claim that none of them can honestly make today.
The claim nobody can make
Start here, because it saves time: no third-party tool can currently verify Claude's watermark. The test needs Anthropic's key, and the detection API has not been released. Any product page implying otherwise is describing something it cannot do.
With that established, the categories differ in useful ways.
Category 1: Unicode and invisible-character cleaners
What they do: strip zero-width characters, normalise spaces and punctuation, remove stray control codes.
Worth using: yes. Invisible characters break tooling and a few really are used for fingerprinting.
Relevance to Claude's watermark: none. Anthropic states that no hidden characters are added, so there is nothing at that layer to remove. Details in why Unicode cleaners do not remove Claude watermarks.
Category 2: General AI humanizers
What they do: rewrite text wholesale, usually through another language model, aiming to change how it reads.
Worth using: depends entirely on whether you want your document rewritten. They are blunt: structure, terminology and voice all move, and the edit is rarely reviewable line by line.
Relevance to watermarks: heavy rewriting genuinely disturbs token statistics — but you cannot confirm by how much, and rewriting through a second model may introduce that model's own watermark. Claims of guaranteed removal are not supportable by anyone.
Category 3: Statistical estimators (where FreeAIText sits)
What they do: reproduce the published green-list z-test to estimate statistical watermark pressure, score segments, and target edits at the passages contributing most to that estimate.
Worth using: for understanding how these schemes behave, and for making small, reviewable edits instead of a wholesale rewrite.
Relevance to Claude's watermark: indirect. A local estimator does not hold Anthropic's key, so its score is a research estimate, not a Claude verdict. It is honest about the mechanism; it is not the same test.
Capability comparison
| Capability | FreeAIText now | Unicode cleaners | General AI humanizers |
|---|---|---|---|
| Hidden Unicode cleanup | Yes | Yes | Sometimes |
| Local statistical estimate | Yes, research only | No | Usually no |
| Full list of edits | Yes | Usually | Not always |
| Browser-based processing | Yes | Usually | Usually no |
| Anthropic API verification | Ready to integrate when public | No | Not currently verifiable |
| Genuine API score before and after | Planned, subject to API access and terms | No | Not currently verifiable |
How to evaluate any tool in this space
- Does it say where its score comes from? A tool that will not distinguish its own heuristic from a provider's detector is hiding the most important fact about its output.
- Does it claim guaranteed removal? Nobody can guarantee that, because nobody outside Anthropic can check.
- Does it claim an official relationship? Look for a named, verifiable one. FreeAIText is independent and not affiliated with, endorsed by, or sponsored by Anthropic.
- Can you see every edit? Unreviewable rewriting is a risk to your meaning.
- Where does your text go? Local processing and API processing are different privacy propositions and should be labelled as such.
What changes when the API ships
Once Anthropic publishes its detection API, one row of that table becomes answerable for real: a genuine score from Anthropic's endpoint, labelled as such. If the terms also permit iterative use, a before-and-after pair becomes possible — both figures from the same endpoint, neither of them inferred.
Until then the honest position is the one on this page: local estimates are useful and clearly labelled, and verification is pending. We track it on the API status page.
Sources
- Anthropic, Watermarking Claude's text output.
- Kirchenbauer et al., A Watermark for Large Language Models (2023).