ChatGPT Watermark Detector for Text

Not in the cryptographic sense most people mean. OpenAI built a statistical text watermark and has discussed it publicly, but never shipped it or a detector for it, so there is no ChatGPT mark anyone outside OpenAI can decode. What ChatGPT output does carry is mechanical: typographic characters the chat interface bakes into whatever you copy out. Those are what the scan finds.

U+202F is the narrow no-break space, an ordinary typographic character used before punctuation in French and inside some number formats. It picked up its AI reputation when readers noticed it turning up in output from ChatGPT's o3 and o4-mini models, often before a colon or a percent sign. Treat a hit as a hint worth following, not a verdict: any passage that has been through a typesetting tool can contain one.

No. OpenAI withdrew its own AI text classifier in 2023 for low accuracy, and it has not published a checker for the statistical watermark it developed. The provider-side tooling that does exist covers images instead: C2PA Content Credentials on OpenAI image output, and SynthID on the supported OpenAI and Google image variants.

Yes, and they are a different check. OpenAI image output can carry C2PA Content Credentials and a SynthID watermark, which Slop or Not reads directly rather than guessing at. Use the ChatGPT image detector for files; the check here is for pasted writing.

The AI text detector asks one question about any writing: does it read as machine-generated? The check here starts narrower, with the characters ChatGPT specifically tends to leave behind, and then hands you the same text verdict for the writing itself. Character traces are fast and concrete but easy to lose; the text verdict survives a copy through a plain-text editor.

What you can remove are the mechanical traces, and one click here does it: invisible Unicode, smart punctuation, and lookalike letters all go, and no word is rewritten. Non-breaking spaces are deleted rather than swapped for ordinary ones, so French spacing closes up and 100 % comes back as 100%; check that before you paste the clean copy over the original. What you cannot remove is the writing style, which is the evidence a detector actually reads. Anything advertised as a ChatGPT watermark remover is doing the first job and charging for the second.

No. The text model reads sentence structure and word choice, so stripping characters does not change what it sees. That model caught 99.5% of AI text at a 5% false-positive rate on RAID, a public benchmark we do not run, with 11 adversarial attacks in the test set. Character cleanup is for tidy text, not for hiding its origin.

It means the text carries no mechanical traces, and nothing beyond that. ChatGPT output routed through a plain-text editor, an API call, or a code block often arrives clean, and plenty of human writing arrives with curly quotes and em dashes from a word processor. Run the AI text check when you want an answer about the writing rather than the characters.

Select the passage, copy it, and paste it into the checker above, then paste the cleaned version back over the original. Word can show paragraph marks and tabs but not zero-width characters or lookalike letters, and Google Docs has no view for them at all. Both also add curly quotes and em dashes through autocorrect, so some of what the scan reports will be the word processor rather than ChatGPT.

No. The character scan and the cleaned output are produced on your own device, so the passage is never uploaded and never stored. The AI text check is a separate step and does travel: it runs on Numen's private Mac server, away from third-party AI providers, and the text is deleted after processing. The iPhone and Mac apps run both checks on-device.

Yes, on Apple devices. Copy an answer out of the ChatGPT app and the iPhone and Mac apps run both the character scan and the text check on-device, with no network involved. They also add spelling conversion between US and UK English that the browser scan does not do. On the Mac, Pro ships a slop command and an MCP server, so piping the clipboard through slop cleanup handles it in one line, and Claude, Codex, or Cursor can call the same check.