# Invisible Unicode vs AI text watermarks

Originally published at https://gifi.ai/guides/invisible-unicode-characters.

Invisible Unicode — zero-width spaces, tag characters, bidi overrides, a byte-order mark — is a real payload you can list and delete. A ChatGPT, Claude or Gemini watermark is not that. SynthID-Text never adds a character; it changes which ordinary word was chosen. Gifi shows the deletable marks for free on the home page. Disturbing the statistical trail is a rewrite.

Reviewed by [Can Balkaya](https://gifi.ai/authors/can-balkaya). Published 2026-08-17; updated 2026-08-17.

## How to verify the job

1. **Inspect the string.** Paste the text to name each invisible character, mixed-script confusable, and punctuation replacement.
2. **Remove named characters.** Clean the verifiable payload without pretending it was the statistical watermark.
3. **Decide on rewriting.** If the remaining concern is word choice, use a rewrite and keep the detector limitation in mind.

## Marks you can point at

A zero-width space occupies no column in your editor and still sits in the string. Unicode tags can smuggle a hidden payload after an apparent letter. A leading byte-order mark, leftover bidi controls, and some punctuation lookalikes do the same job: they change the bytes without changing what a casual reader sees.

Those are verifiable. Gifi names each code point and its offset. You can confirm the count in any hex editor. Cleaning them does not need a model.

## Marks you cannot point at

SynthID-Text derives a random number from a secret key and the words so far, then tilts the dice toward some next tokens and away from others. The sentence looks like ordinary English because it is ordinary English. Deleting characters does not touch that trail. Light edits usually keep it.

The only honest lever is to choose different words. That is the rewrite. It will change the prose. It will not come with a certificate from a vendor detector.

## Why Gifi keeps the jobs on different buttons

A visitor who pastes text should see the invisible characters immediately, without an account. Spending a credit on a rewrite when the only problem was a zero-width space would be a bad product. Spending a credit on a Unicode clean when the problem is SynthID-Text would be a lie.

If both are present — hidden characters in watermarked prose — clean the string, then decide whether the remaining words still need a rewrite.

## Questions

### Does deleting zero-width characters remove a text watermark?

No. A statistical mark is not stored as hidden characters. It is a pattern in the words that remain.

### Is the Unicode clean free?

Inspecting pasted text on the home page is free. The characters are shown in place. A plan is for the rewrite and for file cleaning.

### Will cleaning Unicode change the meaning?

Removing zero-width and tag characters should not change the visible sentence. Optional NFKC folding and homoglyph folding are lossy and off by default, because they can turn a micro sign into a Greek mu.

## Primary sources

- [Unicode security considerations](https://www.unicode.org/reports/tr36/)
- [Google DeepMind on SynthID](https://deepmind.google/models/synthid/)