Does Copying and Pasting AI Text Remove the Watermark?

The only thing copy and paste can affect is hidden characters that were riding along as styling. Those do get dropped when you paste into a plain-text field. Statistical watermarks are not those characters, so nothing about the transfer touches them.

Why the belief exists

People assume "watermark" means "something added to the file," the way a logo is added to a stock photo. That intuition is correct for a visible overlay on a generated image. It is wrong for text.

The April 2025 ChatGPT story reinforced the intuition at exactly the wrong moment. OpenAI's o3 and o4-mini were briefly emitting U+202F narrow no-break spaces in long responses, and a plain-text paste really could strip them. It was a real hidden-character quirk, and OpenAI said explicitly that it was not a watermark.

Because there really is a category of hidden characters in AI text, and because pasting really does affect that category, the fact gets over-generalized into "paste strips the watermark." It does not.

The mechanism, plainly

Claude's watermark is a small, keyed bias in which word Claude picks when several would work equally well. Google's SynthID Text works the same way, and the technique was published in Nature in October 2024. Google's own detector, in the version integrated into open-source tooling, reads token sequences. It does not read characters, and it does not read metadata.

Anthropic's statement is the shortest version of the argument:

"Nothing is added to the text and there are no hidden characters."

If nothing is added, there is nothing for a paste to strip.

What copy and paste actually does

When you copy from a chat window, the clipboard usually receives two versions of the selection: a styled HTML flavour and a plain-text flavour. What happens next depends only on which flavour the destination accepts.

Paste into a rich-text field, such as a word processor or most email composers, and the HTML flavour is used. The words are identical. A watermark, if present, is intact.

Paste into a plain-text field, such as a text editor or a code editor, and the plain flavour is used. Styling is dropped. The words are identical. The watermark is intact.

Paste from that plain-text field into a third app, and the result is the same again. The mark travels with the text because the mark is the text.

What does get changed by a paste

• Rich-text styling: fonts, colours, sizes, links as formatting.
• Sometimes a byte-order mark (U+FEFF) at the start of the pasted run.
• Sometimes non-breaking or narrow no-break spaces, depending on how the destination app handles a paste.
• Clipboard HTML class names and attributes, when the styled flavour is discarded.

Every one of those is worth cleaning for ordinary reasons. None of them is a statistical watermark. All of them are visible in the text tool, with the code point and position for each.

What does weaken a statistical watermark

The short list, all from Anthropic's own August 14 post and the independent commentary that followed:

• A full rewrite through a different model.
• A full manual rewrite.
• Translation through a non-Claude route.
• Heavy editing across a long passage.

Light editing does not. Copy and paste does not. There is a dedicated article on where the line falls: does editing or paraphrasing remove an AI watermark.

Images are a different question

The paste question gets asked about pictures too, and the answer splits.

Pixel watermarks, including SynthID Image, live in the pixel values themselves. Screenshotting, re-saving as JPEG, resizing and cropping do not remove them. A screenshot is a copy of the pixels, and the mark is in the pixels.

File metadata, including C2PA Content Credentials, EXIF and XMP, does get dropped by some transfers, and can be removed deliberately. The image tool strips it and re-scans the export to confirm.

So for images the honest answer is: a re-save may remove the label, and it will not remove the mark. The two live in different parts of the same file.

Where the confusion causes harm

Two failure modes come out of this belief, and they point in opposite directions.

The first is false safety. Someone pastes generated text through a plain-text editor, watches the formatting disappear, and concludes the text is now unmarked. If the source model was marking, nothing changed. The confidence is the only thing that moved.

The second is false accusation. Someone finds a zero-width space in a submitted document and treats it as proof of AI use. Rich-text pastes leave those characters constantly, in documents written entirely by hand, and the mark that would actually indicate a model was involved is not stored that way at all.

Both mistakes come from the same missing distinction: characters are a transport artifact, and word choice is the signal. Only one of them survives a paste, and it is not the one people are looking at.

The short version

Copy and paste is a transport operation. It moves words from one place to another, and sometimes it drops formatting on the way. A statistical watermark is a property of which words were chosen, so it survives every transport that preserves the words, which is all of them.