How to Check for AI Watermarks in Word and Google Docs

People search for a way to see AI watermarks in Word or Google Docs as if there were a hidden layer in the document that the right menu would reveal. There is not. What a document can carry, and what you can check today, is a short list, and only part of it lives inside the editor.

What "AI watermark" can mean in a document

Two different things get the same name.

A statistical watermark is a pattern in word choice. Claude's mark and Google's SynthID Text work this way: nothing is added to the text, so there is nothing for a formatting view to show. Reading one takes the provider's key. As of September 8, 2026, Anthropic's detector is in private preview for eligible organisations. Google opened its SynthID Detector at synthid.com to everyone on October 7, 2026, but it checks images, video and audio, not text. So there is no public paste-and-check page for either text mark. Any check you can run on a document reports that row as Unknown, and so does DetectMark.

Hidden characters are real code points that render as nothing or as an ordinary space: zero-width spaces, soft hyphens, byte-order marks, direction controls and the special spaces. They are not a watermark from any major provider, but they do turn up in pasted AI output, in web copy and in text that passed through other tools. These can be found, listed and removed.

So the useful question is not "where is the watermark in my document" but "what is in this text." That one has an answer.

What Word and Google Docs can show you

Both editors have a view for formatting marks. Neither is built for this job.

Word's Display options can show tab characters, spaces as dots, paragraph marks, hidden text, optional hyphens (Word's name for soft hyphens), object anchors and optional breaks, plus a "Show all formatting marks" switch. That reveals some of the list above. It does not name code points, and Microsoft's list of formatting marks does not mention zero-width spaces or joiners, so a view full of dots and pilcrows does not tell you whether they are there.

Google Docs has View > Show non-printing characters. Google's help page lists what it marks in blue: paragraph, line, section, page and column breaks, tabs and spaces. Zero-width characters are not on that list.

Both views are fine for layout problems. For a full answer you need to see the characters themselves.

How to check a Word or Google Docs file

1. Open the document and select the text you want to check, or all of it.
2. Copy it.
3. Paste it into the scanner. The scan runs in your browser and nothing is uploaded.
4. Read the result. Every hidden character is listed by code point and position, with an explanation of what it does.
5. If you want a clean copy, click Remove. The default cleanup deletes only the characters that are safe to delete, and you copy the result back.

The scan reads what survived the copy. A rich-text copy puts two versions on the clipboard, plain text and HTML, and the app reads both. If the copied markup names the program that produced it in a generator tag, the Clipboard formatting row lists that tag as "1 suspicious marker". If the text has no hidden characters, the headline still reads "No removable marks". The tag names the editor that made the document, not an AI tool.

What the result can and cannot tell you

A Found result means the text contains hidden characters, and the report says which ones and where. It does not say who or what put them there. Editors upstream, web pages and chat interfaces all leave them.

A clean hidden-character result says nothing about a statistical watermark. That row stays Unknown, with the reason attached, because no browser can read Claude's mark or SynthID Text without the provider's key. Does Claude watermark its text? and what SynthID is cover those marks.

Images pasted into a document are a separate check. Run the original image file through the image tool, since a copy taken out of a document may not keep the file's metadata.

Why not upload the .docx?

Because nothing needs to leave your device. Copying the text out gives the scanner everything a hidden-character check can use, and it keeps the promise in the header: local processing, 0 B uploaded. The only requests the app ever makes with your content are two optional relays, each behind a consent dialog, and a document check uses neither.