Watermark pressure
—LOCAL RESEARCH ESTIMATE
0%
How far this document's token statistics sit from chance, on the local research estimator. 50% is exactly chance. It is not a verdict from Anthropic's detector.
LOCAL RESEARCH ESTIMATOR · ANTHROPIC API READY
Claude Watermark Detector & Text Cleaner
Analyze statistical watermark pressure locally today. When Anthropic releases its public watermark detection API, FreeAIText is designed to add real API verification with clearly labeled before-and-after results.
Current status: Research estimator only
FreeAIText does not currently have access to Anthropic's private watermark key or detector. Today's score is a local statistical estimate, not an Anthropic verdict.
- Local mode keeps your document in this browser
- Results in under a second
- Every score labelled with its source
Analyze your text
Paste your text or upload a file. The analyzer scores every segment locally and isolates the ones that contribute most to the local research estimate.
- Paste your text Or load the sample
- Analyze it Free, no account
- Apply targeted edits 1 credit per run
Local mode: your document remains in this browser.
API mode: the text you submit will be transmitted securely to Anthropic for watermark analysis. Review the Privacy Policy before continuing. Read the Privacy Policy.
Quick human check
This runs once per session and keeps automated traffic from draining the detection budget.
Ready to edit
These segments contribute most strongly to the local estimator's score. This is not token attribution from Anthropic's detector.
Estimator statistics
- Green-token rate
- —
- Estimator z-score
- —
- Tokens analyzed
- —
- Segments scanned
- —
- High-contribution segments
- —
- Signal concentration
- —
Highest-contribution segments
These segments contribute most strongly to the local estimator's score. This is not token attribution from Anthropic's detector.
Marked text
Non-highlighted text contributed less to this local estimate. It should not be interpreted as confirmed watermark-free text.
Editing run
Each targeted segment is rewritten several ways, every version is re-scored, and only the version that lowers the score is kept. Stubborn segments get another round with more words in play.
Revised text
Targeted words replaced with meaning-preserving alternatives, so you can review targeted edits and observe how they affect the local research estimate.
Review every replacement (0)
How to read these numbers: this is a local research estimator running in your browser. It reproduces the green-list z-test used in watermarking research, but with a key that is not any vendor's, so the score estimates statistical watermark pressure — it is not a verdict from Anthropic's detector or from any other provider. Use the tool on text you own or are permitted to edit.
The method
Three steps, one second
Statistical signal is not spread evenly. A handful of token choices carry most of the measurable evidence — so that is where the tool works.
Score every segment
Your text is split into segments and each token is hashed against its predecessor to determine green-list membership — the construction used in the watermarking literature. Each segment gets a z-score from the local estimator.
Isolate the highest contributors
The three to five highest-scoring segments are highlighted. These segments contribute most strongly to the local estimator's score. This is not token attribution from Anthropic's detector.
Edit, re-score, keep the winner
Inside those segments only, several rewrites are generated using part-of-speech-matched synonyms. Every version is scored again, and only the one that lowers the estimate survives — so you can review targeted edits and observe how they affect the local research estimate.
What this does not do: it does not tell you which words carry Claude's
watermark. That attribution requires Anthropic's key, and
the detection API has not been released.
Capabilities
Know what each tool can actually verify
Three categories of tool get compared to each other constantly. They do not do the same job, and only one row here is about Anthropic's actual detector.
| 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 |
Full write-up with sources: Claude watermark tools compared.
Why this tool
Built to be surgical, not destructive
Most rewriters shred your whole document. This one edits a few dozen words and leaves your writing intact.
Targeted editing
Only 3–5 segments are touched. Structure, tone, terminology and argument stay exactly as you wrote them.
Real statistics, labelled
Green-token rate, z-score and signal concentration are computed live, and every figure carries a badge naming the engine that produced it.
Meaning preserved
Part-of-speech matching, capitalization retention, and a stop-word blacklist prevent nonsense substitutions.
Private by design
Local mode processes your document in this browser tab. If an API mode is ever enabled, you are told before submission and asked to consent.
Full change log
Every substitution is listed with its original word, so you can review or reverse anything you dislike.
Import & export
Drag in a .txt or .md file, and copy or download the result in one click.
3–5segments targeted per pass
<1saverage analysis time
0bytes of your text stored in local mode
Free local analysis · API-verified checks will use a free and paid credit model when available
Answers
Frequently asked questions
Does FreeAIText use Anthropic's watermark detection API?
Not today. Anthropic has announced watermarking for Claude and has said a detection API is coming, but it has not been released. FreeAIText does not have Anthropic's private watermark key or detector, so the score you see now is a local statistical estimate produced in your browser. The integration is built and disabled behind a feature flag; the current state is published on the Anthropic API status page.
What is a Claude text watermark?
A statistical bias in word selection. Where several continuations are equally valid, the model's choice is driven by a keyed source of randomness, so someone holding the key can test whether a sequence is consistent with the choices the model would have made. Anthropic states that nothing is added to the text and there are no hidden characters, and that a watermarked response reads identically to an unwatermarked one.
What does the local research estimate actually measure?
It reproduces the green-list z-test from the published watermarking literature and reports how far the token statistics of your text sit from chance. Because it does not hold any vendor's private key, it measures general statistical watermark pressure — not Claude's watermark specifically. It is a research tool, not a verdict.
Do Unicode cleaners remove a Claude watermark?
No. Unicode cleaners strip invisible characters, which is useful for other reasons, but Anthropic states that its watermark adds no hidden characters. The two operate on different layers: one on characters, the other on word choice.
What does a positive watermark result prove?
That the watermarked model was probably involved in producing or substantially editing the text. It does not establish authorship — a person can write the substance and have a model rephrase it. A negative result proves even less: it does not show human authorship, because unwatermarked models, older text, short samples and heavy editing all return nothing.
Is my text uploaded anywhere?
In local mode, no. Analysis and editing happen in your browser and the document never leaves it. The only outbound requests are single-word lookups to the public Datamuse synonym dictionary, which you can switch off entirely. If an Anthropic API mode is ever enabled, the interface will say so before submission and ask for your explicit consent first.
Is it free?
Analysis is free, unlimited and needs no account — paste text and run the estimator as often as you like. Cleaning, the pass that applies targeted edits, uses one credit per run, and every new account is granted enough credits for 2 full cleaning runs. More credits can be bought with PayPal.
Do I need an account?
Only to clean text. Analysis works signed out. Signing in is an email address and a one-time code — no password, and no Google or other third-party account.
What happens if a cleaning run fails?
Your credit is returned automatically. You are only charged for a run that actually produced edits, and a run that fails or finds no safe substitution costs nothing.
Will editing change the meaning of my text?
The editor is deliberately conservative: content words only, part of speech matched, names, numbers, acronyms and function words skipped, and a hard cap of a few substitutions per segment. Every change is listed so you can review or reverse it. Re-read the result before you use it.
Can you guarantee text will not be detected?
No, and neither can anyone else. Nobody outside Anthropic can currently run Claude's detector, so no third party can verify what its result would be. Editing changes token statistics; that is a mechanism, not a guarantee, and we will not describe it as one.
Is FreeAIText affiliated with Anthropic?
FreeAIText is an independent service and is not affiliated with, endorsed by, or sponsored by Anthropic.
More detail: what a Claude watermark proves · AI detector vs watermark detector · Anthropic API status
Analyze your text now
No sign-up. Local analysis runs in your browser, and every score says where it came from.