The checks we actually run
1. Deterministic text inspection (local)
Pasted text is scanned character by character in your browser for a fixed, documented set of Unicode code points: zero-width characters, invisible formatting characters, unusual whitespace, bidirectional controls, C0/C1 control characters, variation selectors and tag characters, plus Unicode normalization state. The scan is exhaustive and repeatable — the same input always produces the same output. Nothing is sent to a server, stored, or included in analytics.
2. C2PA / Content Credentials verification (local)
Selected image files are parsed with the official Content Authenticity Initiative browser SDK (@contentauth/c2pa-web), compiled to WebAssembly and running in your browser. It reads the C2PA manifest if present and validates its cryptographic signature against the file. Files are not uploaded.
3. Provider status (cited)
Statements about what Anthropic, Google, OpenAI and others mark or verify come from a single internal data source, backed exclusively by primary sources (provider documentation and announcements), each with a visiblelast verified date. We do not source provider claims from other checker sites.
What the labels mean
- Verified — a cryptographic validation succeeded (e.g. a C2PA signature validates against the file).
- Detected signal — a deterministic technical finding (e.g. three zero-width spaces at listed positions). A signal is a fact about the content, not a verdict about its author.
- Not detected — the specific checks performed found nothing. Never presented as “clean” or “human”.
- Verification unavailable — a mark may exist, but no mechanism is publicly available for anyone (including us) to check it.
- Inconclusive / Check not completed — a parse error or unsupported input. A failed check is reported as a failed check, never as “not watermarked”.
What we will not do
- No statistical “AI probability” scores. Classifier-based AI detectors guess from writing style and are wrong often enough to be dangerous; that is a different product, deliberately not this one. See watermark checker vs AI detector.
- No simulated provider detection. Until Anthropic or Google ship public verification, no site can honestly detect their text watermarks — so this one doesn’t pretend to.
- No binary verdicts. “No supported watermark signal was found by the checks performed” is the strongest negative statement we make.
Update policy
Provider pages show a last verified date that changes only when a human actually re-checks the primary sources — it is never bumped automatically. When a provider ships a public detection mechanism, it is integrated as a clearly labeled official check, and this page is updated to describe exactly how it works.