Does ChatGPT Watermark Text?

OpenAI's provenance work is described for files: C2PA Content Credentials as signed context, plus SynthID watermarking as a signal that survives when metadata does not, plus tooling to verify whether an image — and, later, supported audio — came from its systems. Verification is also exposed to developers.

Text is absent from that stack. Reporting has long noted that OpenAI researched text watermarking without shipping it to production models, and OpenAI retired its own AI text classifier in 2023 over accuracy. So for ordinary ChatGPT prose there is no documented mark, and no first-party checker.

Because copied text travels through software that inserts them. Rendered pages use zero-width characters for line-breaking control, rich-text editors and docs apps insert formatting characters, PDF exports scatter unusual spaces, and chat clients add joiners around emoji and mentions.

The reasoning error is treating an ambiguous artifact as a signature. These characters have ordinary explanations, appear in text no model touched, and are absent from plenty of model output. Meanwhile the providers that do watermark text state they add no characters at all — so hidden characters are, if anything, evidence against the watermark theory rather than for it.

The case for removing them has nothing to do with provenance. They break string equality, so search misses matches, diffs show phantom changes, validation rejects correct input, identifiers and URLs fail, and pasted code will not compile.

That is a formatting-hygiene job with an exactly measurable result: a count of what was removed. Our invisible character remover does it locally and free, and reports each code point by name.

A watermark is a keyed signal added at generation. A detector guesses from surface statistics, with false positives that fall hardest on plain and non-native English writing. Removing hidden characters changes neither, and cleaning text does not convert a detector's guess into proof of human authorship.

If you are subject to disclosure rules, comply with them. This tool exists for formatting hygiene, privacy, editorial independence and control over machine-readable signals — not for defeating an institution's policy.