AI Watermark Check

Free Copilot Watermark Remover & Checker for Text & Code

Paste Copilot-generated text or code, see every hidden watermark character exposed in an x-ray view, and copy a clean version in one click. Free, online, no sign-up, and nothing you paste ever leaves your browser.

More: what was found, x-ray view and options

X-ray view

Your text with every hidden character exposed as a labelled chip. Hover a chip for its Unicode name.

Is Microsoft Copilot output watermarked?

Copilot adds no watermark of its own — whether one shows up depends entirely on which underlying model answered, since Copilot routes across several, and Content Credentials are applied only to the images it generates, not the text. In practice, the Office documents Copilot writes into contribute far more hidden characters than the model ever does, which is the more common problem worth checking for.

Last verified: 11 August 2026

Copilot is not a model. It is a product surface, in Windows, Edge, Microsoft 365 and GitHub, sitting on top of models from OpenAI and Microsoft's own family. That architecture is the whole story here: the watermarking question has no single answer for Copilot, because Copilot is not one generator.

Why "does Copilot watermark text" has no single answer

When you ask Copilot something, the request may be served by a different model depending on the surface, the subscription tier, the task and the routing decisions Microsoft makes at the time. The watermarking status of the answer is inherited from whichever model produced it, not decided by Copilot as a brand.

In practice, at the verification date above, that inheritance mostly resolves to no text watermark, because the model families Copilot draws on for text do not deploy one. But it is an inherited property that can change without any announcement about Copilot itself, which is a materially different situation from a single-model product like Claude.

Images are different and clearer. Microsoft applies Content Credentials, its implementation of the C2PA standard, to images generated through Copilot and Designer. That is signed metadata in the image file, and it is checkable, until a platform re-encodes the image and strips it.

The Office problem, which is the real one

For most people using Copilot, the practical text-integrity issue has nothing to do with watermarking. It is that Copilot output usually lands in Word, Outlook, Excel or Teams, and those applications are prolific sources of hidden characters.

None of that is model output. All of it breaks the same things: lookups, imports, comparisons and code. The checker above is aimed squarely at this, and the hidden characters guide covers each source in detail.

GitHub Copilot and code

Code suggestions raise a separate concern that is worth stating explicitly, because it is the one case where invisible characters are dangerous rather than merely annoying. Any code you accept from a suggestion, a documentation page, an issue comment or a Stack Overflow answer can carry bidirectional control characters that change how the code appears in your editor without changing what the compiler reads.

That is the code-display problem documented by Nicholas Boucher and Ross Anderson in 2021. It is not specific to Copilot, and it is not hypothetical: it is why several compilers now warn on unterminated bidirectional controls in source files. Running unfamiliar code through the checker before committing it takes a second and rules the whole class out.

What this tool can and cannot do for Copilot output

Can

  • Clean the hidden characters that Word, Excel, Outlook and Teams put into text on its way to and from Copilot. This is the common case and the tool handles it completely.
  • Flag bidirectional controls in code suggestions before you commit them, with high-risk chips in the x-ray view.
  • Replace no-break spaces in numbers with ordinary spaces so a spreadsheet paste imports as numeric data.

Cannot

  • Tell you which model produced a given Copilot answer, or what its watermarking status is.
  • Detect or remove a statistical watermark, if one is inherited from an underlying model in future.
  • Read Content Credentials from an image. This tool works on pasted text.

Questions about Copilot output

Does Microsoft Copilot add hidden characters to its answers?

Not as a marker. Hidden characters in Copilot output almost always come from the application the text passed through, Word, Outlook, Excel or Teams, rather than from the model. The checker above shows you exactly which ones are present.

Why does Copilot output in Excel break my formulas?

Usually because a no-break space, a narrow no-break space or a figure space is sitting inside a value. It looks like an ordinary space, so the cell looks numeric, but the value is text and arithmetic fails. Cleaning the text replaces those with ordinary spaces, after which the value converts normally.

Should I check code that GitHub Copilot suggests?

Checking unfamiliar code for bidirectional control characters is a sensible habit regardless of where it came from. Those characters can make a line render differently from what the compiler reads. The checker flags every one of them as high risk.

Are Content Credentials the same as a watermark?

No. Content Credentials are Microsoft's implementation of C2PA: signed metadata attached to an image file. They are removed by re-encoding or by any platform that strips metadata. A watermark is embedded in the content and survives that kind of handling.

If the underlying model starts watermarking, will Copilot output be watermarked?

Yes, for answers that model produces. Because Copilot routes across several models, the property is inherited per answer rather than being a fixed attribute of the product, which is why this page tracks the underlying providers rather than a single Copilot policy.

Related reading