DeepSeek zero-width space remover

DeepSeek's tokeniser shares a single vocabulary across Chinese and English, and its reasoning models can display a visible chain-of-thought section ahead of the final answer. Both properties create seams: a boundary where two scripts meet, and a boundary where a reasoning block ends and the answer begins, and either can leave a character that occupies space in the string while producing no visible output.

This tool detects and removes the zero-width space along with the related family of invisible codepoints, without altering any character that actually renders. Detection and removal both run client-side, in JavaScript, with no upload and therefore no run limit.

There's nothing to preview beforehand, since the whole category of problem is invisible until the scan runs.

Step 1

Paste the reply copied from DeepSeek, reasoning trace or final answer, into the input field.

Step 2

Run the removal. It executes instantly, locally.

Step 3

Read the result. Every rendered character is unchanged.

A tokeniser trained across two scripts as different as Chinese and English represents each in its own region of the vocabulary, and the point where a model transitions from one to the other during generation is a natural place for a small artefact to appear: a hair space or non-breaking space substituting for an ordinary one, or a piece of full-width punctuation surviving into an English sentence where it looks slightly out of place but rarely wrong enough to notice.

A reasoning model that displays its intermediate reasoning ahead of the final answer introduces a second seam, structurally distinct from the language boundary, at the point where the visible reasoning ends and the delivered answer begins. Copying text spanning that boundary can carry an invisible character sitting exactly at the join.

This tool removes the zero-width space, U+200B, along with the zero-width joiner and non-joiner, the word joiner, the byte-order mark, the soft hyphen, and the non-breaking, narrow and hair spaces, the fuller family that travels with copied AI text under a name that only mentions one of its members.

A joiner detected inside a genuine emoji sequence is preserved rather than removed, since stripping it would visually split the composed emoji apart. The space-like characters in the set are replaced with an ordinary space rather than deleted outright, so word boundaries and line wrapping remain correct.

Inside a heading or a URL slug, a hidden character splits the string for any process comparing it byte by byte: full-text search, a regular expression, a string equality check. Two values that render identically fail to match, because one carries a codepoint the other does not.

In bilingual workflows specifically, a hidden character sitting at a script boundary can silently fail a translation memory match or a term-base lookup, since the comparison depends on byte-for-byte equality that a visually identical string does not guarantee.

These characters are overwhelmingly incidental: byproducts of bilingual tokenisation and rendering rather than markers deliberately embedded to identify a passage later. Their presence indicates that text passed through a tokeniser working across two scripts, or across a reasoning-trace boundary; it does not indicate who wrote the underlying words.

A genuinely different research area, statistical watermarking, biases which tokens a model is more likely to select across an entire passage, leaving a pattern distributed through ordinary word choice rather than a discrete character. No character-level scanner, including this one, can detect or affect that kind of watermark, and no tool on this site claims otherwise.

Translation and localisation teams working with DeepSeek output run it through here before the text enters a translation memory or a term base, since a script-boundary artefact can silently break a match that should have succeeded. Engineers pulling a DeepSeek answer into a config file or a commit use it for the same class of reason: a hidden character produces a failure with no visible trace of its cause.

Researchers evaluating a DeepSeek reasoning trace alongside the final answer use it to clean either section independently, since both can carry the same kind of boundary artefact.