Hidden Character Detector for AI Text Watermarks

Hidden characters are the ones you cannot see in rendered text: zero-width spaces and joiners, word joiners, byte-order marks, and unusual spaces such as the narrow no-break space at U+202F, which drew attention after readers spotted it in output from ChatGPT's o3 and o4-mini models. The scan also flags two visible-but-swapped families: smart punctuation such as curly quotes and em dashes, and Cyrillic or Greek letters shaped like Latin ones. None of them prove AI authorship on their own, because word processors, chat windows, and ordinary copy-paste produce them too.

Select the passage in Word, copy it, paste it into the checker above, then paste the cleaned version back over the original. Word will not show you what is there first. Its formatting-marks button reveals paragraph marks, tabs, and ordinary spaces, and stops well short of zero-width characters and lookalike letters. Find and Replace can pull out one code point at a time if you already know which to hunt, and the scan finds all of them in one pass. For a stack of papers, the iPhone and Mac apps run the same pass on each one offline, and on the Mac, Pro's slop cleanup command can script a whole folder.

Select the text in Google Docs, copy it, and paste it into the checker above. Docs has no built-in view for zero-width characters or lookalike letters, and its own autocorrect adds curly quotes and em dashes as you type, so part of what the scan reports will be Docs' work rather than an AI tool's. Paste the cleaned text back to replace the original.

Yes, as long as the text is selectable. Copy the passage out of the PDF and paste it into the checker above; the scan reads characters, so page layout and fonts do not matter. A scanned PDF is really an image and carries no character data until you run OCR on it, and OCR output carries its own transcription errors rather than the original's hidden characters.

It removes mechanical fingerprints: invisible Unicode, smart punctuation, and lookalike letters. Non-breaking spaces are swapped for ordinary ones rather than deleted, so 100 % keeps its separation; check the spacing before you paste the clean copy over the original. It does not touch statistical watermarks, the kind Google's SynthID-Text weaves into the word choices themselves. Anthropic says future Claude models will carry a version of that same SynthID-Text mark. A statistical mark survives a character swap, and reading one needs the provider's own detector. Cleaning also does not make AI writing read as human: the AI text detector reads the words, and it reads them the same way before and after.

That tells you the text carries no mechanical markers, and nothing more. Plenty of AI-generated writing has none, either because the tool never inserted any or because a trip through a plain-text editor stripped them, and plenty of human writing has some. For an answer about the writing itself, run the AI text check and read the verdict as supporting evidence rather than proof.

Yes. The hidden-character scan and the cleaned output are produced in memory on Numen's server and discarded with the response, so the passage is never stored. The AI text check is a separate step and travels to 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 entirely on-device.

No. The AI text detector reads sentence structure and word choice, not invisible characters, so removing them does not change the evidence it reads. 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. Treat any tool promising undetectable output after a Unicode strip as selling something it cannot deliver.

Yes, on Apple devices. A stack of student submissions or client drafts never has to touch a network: the iPhone and Mac apps run the scan and the cleanup on-device, and they add spelling conversion between US and UK English that the online scan does not do. On the Mac, Pro also ships a slop cleanup command and an MCP server, so a whole folder can go through it from a script, or from Claude, Codex, or Cursor.

Mostly by accident. Chat interfaces render answers as rich text, so quotes curl, hyphens become em dashes, and typographic spaces such as U+202F get baked into whatever you copy out. Some providers have experimented with deliberate provenance markers, and many documents pick these characters up from a word processor or a website long before any AI touches them. A hit is the start of a question, not the answer to one.