How do you check text for an AI watermark?
First identify the kind of watermark you are looking for. A literal marker can be read from the pasted characters. A statistical watermark woven into word choices needs the matching provider detector. A responsible checker reports those results separately.
Paste the text into the AI text watermark checker. It will expose supported invisible characters and decode exact C2PA 2.4 text credentials. Claude and SynthID remain marked “not testable” unless their matching detector is available.
What this checker tests
| Signal in the text | Result here | Evidence produced |
|---|---|---|
| C2PA 2.4 text credential | Detected and removable | Exact wrapper version, declared manifest size, offsets, and manifest SHA-256. |
| C2PA structured-text block | Detected; safely bounded forms removable | Form, reference kind, offsets, hashes, and an explicit no-fetch record. |
| Supported invisible Unicode | Detected | Code point, character name, count, offsets, and cleanup policy. |
| Claude model-level watermark | Not testable yet | Anthropic has not yet published the promised technical detector details. |
| Gemini SynthID Text | Not tested by generic paste scan | Detection needs a compatible tokenizer, watermark configuration, and detector. |
| Writing style | Not scored | No authorship probability or “human” score is invented. |
Step-by-step: check copied text
- Paste the original text. Copying the untouched output gives a literal marker its best chance of surviving.
- Read the separate scan states. “Found,” “not found,” and “not testable” mean different things.
- Inspect exact findings. A normal emoji variation selector is review-only; it is not automatically called a watermark.
- Remove only supported signals. Complete C2PA text wrappers and safe literal markers can be removed without rewriting the words.
- Download evidence if needed. The JSON report contains hashes, versions, findings, and limits—but not the raw text.
Why the C2PA text result is different
C2PA 2.4 defines two public text formats. Unstructured text uses a U+FEFF prefix and variation selectors that decode into C2PATXT\0, a version, a declared length, and a manifest store. Structured formats use fixed manifest delimiters around a URI or Base64 data in comments or front matter. That lets the checker identify the standard precisely instead of treating every unusual character as an AI fingerprint. See the invisible-wrapper rules and structured-text rules.
Why Claude and SynthID need separate detectors
Google documents SynthID Text as a logits processor that changes token probabilities using configured keys. Its Bayesian detector can return watermarked, not watermarked, or uncertain. Anthropic describes Claude's new embedded mark as model-level and says technical detection guidance is forthcoming. Neither description matches a generic search for zero-width characters.
How to interpret “0 found”
A zero is a scoped negative result: no supported literal Unicode marker and no complete C2PA text wrapper was found. It is not evidence that the prose is human-written, not generated by AI, or free of every provider watermark. That distinction is the central difference between this checker and tools that turn an empty Unicode scan into an authorship claim.
Primary sources and methodology
- C2PA 2.4: unstructured-text credentials
- Google AI: SynthID Text watermarking and detection
- Anthropic: how Claude marks AI-generated content
- Check AI Watermarks: exact scanner methodology
Frequently asked questions
Can I check text for an AI watermark online?
Yes, but the result is only authoritative for supported layers. Check AI Watermarks finds literal Unicode markers, invisible C2PA wrappers, and structured C2PA manifest blocks locally in your browser. It does not claim to detect provider-keyed systems such as SynthID Text or Claude's announced model-level watermark.
Does a zero result mean the text is human-written?
No. Zero means no supported literal marker or complete C2PA text wrapper was found. A provider-only watermark may still exist, and AI-generated text may have no detectable watermark at all.
Is an AI watermark checker the same as an AI detector?
No. A watermark checker searches for an intentionally embedded signal. An AI detector estimates authorship from learned patterns. Both can have limited coverage, but they answer different questions.
Does this checker send my text to an AI model?
No. Text inspection, visualization, cleanup, and evidence-report generation run in the browser without an LLM call or content upload.