LLM Watermark Detector

Why AI text watermarks matter now

Article 50 of the EU AI Act requires providers of generative AI to mark what their systems produce so that it can be detected as artificially generated, and those obligations have applied since August 2026. For images, a watermark can hide in noise no one notices. Text has no such room. As the engineer Sean Goedecke puts it in his write-up on text watermarks, you cannot make any change to a sentence that a human would not notice, with one exception: characters that look identical to the ones they replace.

That leaves providers two practical ways to mark text, and they behave very differently when you go looking for them.

Statistical watermarks: marks in the word choices

Google's SynthID Text, used for Gemini, is the best-known example. At each step a model has thousands of possible next words, most of them plausible. SynthID gives every candidate a score computed from the words that came before, then picks the highest-scoring word from among the few most likely ones. The text reads normally, but across a few hundred words the scores add up to more than chance allows, and a detector that knows the scoring can measure that.

OpenAI has said it built a similar method for ChatGPT without releasing it. The catch for anyone checking text is that a statistical watermark lives in the choice of words, not in the characters. Only a detector built for that provider's scheme can read it, so no general character scan, including this one, can see it.

Character watermarks: marks you can find

The other approach swaps ordinary characters for lookalikes. A normal space (U+0020) can become a three-per-em space (U+2004) or an ideographic space (U+3000); a Latin letter can become its Cyrillic twin. Placed in a pattern, such as every third space, the swap is invisible to a reader and very unlikely to occur in human writing. Goedecke reports that assistants do emit these lookalike characters, even if whether they are meant as a watermark is not public, and many writers have noticed code editors flagging unusual spaces in text pasted from ChatGPT.

Hidden characters also arrive for plain reasons: the chat interface, a word processor or a publishing tool can insert them. Whatever the source, they are real, they travel with copy and paste, and they are what the detector finds.

The hidden characters the detector looks for

Each check is deterministic: the same text always gives the same result, and every finding is tied to a position. The families it covers:

• Unusual spaces: the no-break space (U+00A0), the en, em, three-per-em, thin and hair spaces (U+2000 to U+200A), the narrow no-break space (U+202F) and the ideographic space (U+3000). They look like a normal space and behave differently.
• Zero-width characters: the zero-width space (U+200B), zero-width non-joiner and joiner (U+200C, U+200D), word joiner (U+2060) and a stray byte order mark (U+FEFF).
• Lookalike letters, such as a Cyrillic a or o in place of the Latin letter, or full-width letters from East Asian typography.
• Bidirectional controls, such as U+200E, U+200F, U+202A to U+202E and U+2066 to U+2069, which can reorder how a line displays.
• Unicode tag characters (U+E0000 to U+E007F), an invisible block that can carry a hidden message alongside the visible one.
• Variation selectors and unusual whitespace patterns that ride along with ordinary text.

How to read the report

The detector returns a count of the signals it found, grouped by kind, and marks each one in place. A short excerpt around each finding shows the invisible character as its code point, so a lookalike space between two words appears as a visible marker rather than nothing at all. A regular pattern, such as the same unusual space at every third word break, is worth more attention than a single stray character.

The results⟦U+2004⟧speak for⟦U+200B⟧themselves.

Found: 2 signals
  U+2004  three-per-em space   position 11
  U+200B  zero-width space     position 21

What the detector cannot tell you

A signal is evidence that a character is present, nothing more. Finding an unusual space does not mean an AI wrote the text, since word processors, publishing tools and some keyboards insert the same characters. Finding nothing does not mean a person wrote it: the text may carry a statistical watermark this scan cannot see, may have been cleaned, or may come from a model that emits plain characters.

Use the report to see what is actually in a piece of text before you publish or reuse it. Do not use it as an authorship verdict about a student, a writer or a colleague.

From clean text to being cited

Hidden characters do real damage to published content. A zero-width space inside a brand name stops a page matching searches for that name, and a lookalike letter makes a heading fail exact-match checks and internal search. The larger question for a brand is whether AI assistants mention it at all when buyers ask. AmICited asks ChatGPT, Perplexity and Google AI about your brand on a schedule, keeps every answer, and shows when your visibility moves.

If the detector found characters you want gone, the LLM Watermark Remover strips them and gives you a rewrite to review. To see where you stand with the assistants today, run the free AI Brand Visibility Report on your own brand.

Frequently asked questions