Evidence first

What is a ChatGPT watermark?

Not every invisible character is a watermark, and not every watermark is a character.

Three different issues often get mixed together

Copy-and-paste artifacts are unexpected characters, such as a zero-width space or a nonstandard space, that can affect formatting. Some are accidental, and some are intentional in multilingual text. ClearCopy can identify or change selected characters.

File metadata and content credentials are another category. Image and document files can contain metadata about creation or editing. ClearCopy’s browser tool currently handles pasted text and small text files only, not image or PDF metadata.

Statistical text watermarks are patterns created by the sequence of words or tokens selected by a model. OpenAI describes its technology, textGrain, as influencing word choices rather than adding hidden spaces. A Unicode cleaner cannot remove this kind of signal simply by deleting characters.

What does OpenAI say about textGrain?

OpenAI’s public provenance documentation says ChatGPT text watermarking is used in the EU and API customers globally may opt in. Coverage depends on product and context. OpenAI also says these signals have technical limitations and may survive some editing. You can read the official provenance explanation.

Can rewriting remove a statistical mark?

Substantial rewriting may change word-choice patterns, but no general-purpose tool can honestly guarantee a given detector will not recognize text, or certify that text was written by a human. Rewrites can also change facts or lose your voice. For material you publish, keep necessary AI-use disclosures and review factual accuracy.

What our free tool promises

ClearCopy reports the exact deterministic text-hygiene changes it applies. It does not claim to detect a textGrain signal, calculate its strength, or certify its removal. This limit is important, even if some websites advertise stronger claims.

Further reading