How Anthropic describes Claude text watermarking
Anthropic documents a statistical text watermark and a private-preview detection API for eligible organizations. The documentation also explains limits on what a watermark result can establish. Check current access and scope directly with the provider if your task needs that method.
GPTClean-up does not use that API. Our inspector works on supported Unicode characters in the input. It cannot take the place of a provider verification process, and it does not label a passage as Claude-generated from the presence of ordinary formatting artifacts.
Claude-generated text vs Claude-edited text
A passage can be drafted by a person and lightly edited, translated, substantially rewritten, or generated from a prompt. Those workflows are not equivalent. Before using a detection result, decide which form of involvement matters to the policy or question at hand.
Keep the evidence proportionate. A result about likely provider involvement does not automatically identify a user, establish a complete drafting history, or determine responsibility for every claim. Be careful with labels that imply more than the method measures.
Check sources and drafting records
Check whether the reasoning follows, whether quoted sources support the claims, and whether numbers can be reproduced. Ask for appropriate drafting records when the review process allows it. A concrete unsupported citation is a clearer issue to discuss than a general impression that the prose sounds like a model.
If you are the writer, preserve notes about the assistance used and the revisions you made. That record can explain your process without relying on a later classifier guess. Follow the destination’s requirements for disclosure and attribution.
What a Claude character scan detects
A character scanner can identify a covered invisible character in a supplied sample. If that character causes a field mismatch, you can document the mismatch, edit the field, and repeat the operation. That is a useful and bounded result.
It cannot establish that Claude inserted the character, that a statistical watermark was removed, or that the resulting draft will receive a different score elsewhere. A report with no findings leaves those questions open too.
Evaluate Claude detection claims and limitations
If the outcome affects a person, document the method used, the evidence available, and its limitations. Allow a response to specific concerns and distinguish uncertainty from a verified fact. Avoid presenting a polished dashboard or a precise percentage as a substitute for validated evidence.
For a straightforward copy-paste task, the goal is simpler: preserve the original, clean only the unwanted artifacts, and verify the result. Keeping that workflow separate from authorship assessment makes both the tool and the final conclusion easier to trust.
Sources & further reading
Primary references for the technical points in this guide.