Search for "AI watermark detector" and you will find tools that promise three very different things. It helps to know which one you actually need before trusting a result. Three separate jobs hide under that phrase, and no single tool does all of them well.

Three different ideas hiding under one name

1. Invisible characters in text

Some generated text contains zero-width spaces, special non-breaking spaces, or control characters inserted between visible letters. These are real Unicode code points. A character scanner can find them, count them, and remove them. This is what this tool does.

Because they are real code points, they have assigned numeric values, they follow defined encoding rules, and they can be copied, counted, and deleted like any other character. Detection is therefore deterministic: the character is either present in the bytes or it is not, and there is no guesswork in the answer.

2. Statistical watermarking

Methods like Google's SynthID Text change the probability of which token the model chooses next. No hidden character is inserted. Detection requires the same watermark key and configuration used during generation. A generic scanner cannot detect this.

The output is ordinary text; the difference is that the model's token choices were biased while it wrote. With nothing extra inserted into the text, a tool that reads only the finished text has no signal to look for.

3. Style-based AI detection

Some services try to guess whether a human or an AI wrote a paragraph based on writing style. These are heuristic estimates, not verifiable signals. They can be wrong and should not be treated as proof.

Sentence length, punctuation habits, repetition, and vocabulary are indirect clues. A style score describes a guess about a pattern, not a measurement of a mark, and it can be wrong in both directions.

Why the name is confusing

Because "AI watermark" is used for all three, a tool that finds invisible characters can be mistaken for a tool that detects AI output. The honest thing is to separate them. This tool finds the first kind and tells you clearly when it cannot help with the others.

A visitor who expects a human-versus-AI verdict can run a character scan, see a clean result, and conclude the text is unmarked, even though the scan never examined the other two categories at all.

How invisible characters behave in text

Zero-width characters are the classic example. Code points such as the zero-width space (U+200B), the zero-width non-joiner (U+200C), and the zero-width joiner (U+200D) occupy no width on screen, so they render as nothing between two visible letters. Because they are invisible, they survive copy-paste without anyone noticing, and a string that contains them can look identical to one that does not.

Other candidates work the same way. The word joiner prevents a line break without adding a visible mark. A byte-order mark sometimes sits at the very start of a text file. Special spaces such as the non-breaking space (U+00A0), the narrow no-break space, and the ideographic space look like ordinary whitespace but use a different code point, so a comparison between them and a plain space fails even though they render similarly.

Because these are real, deterministic characters, a scanner can report the exact code point and position it found. There is no inference involved: the tool compares each character against a list of known invisible or look-alike code points and lists the matches it sees.

How statistical watermarking differs

Statistical watermarking works at a different layer of the pipeline. Rather than inserting characters, it nudges which tokens the model prefers while it generates. The visible text stays ordinary, and a character scanner sees nothing unusual, because there is nothing unusual to see in the bytes.

Verifying such a mark means re-running the model's own detection process with the correct key. That is why a tool that only reads the finished text cannot confirm or deny the mark. If a provider offers an official detector for its own watermark, that detector is the only reliable way to check it.

How to tell which kind you are dealing with

  1. Ask where the text came from. Text pasted out of a chat interface is a candidate for stray invisible characters. Text produced through an API that documents a statistical mark is a candidate for that mark only.
  2. Inspect the raw characters, not the rendered text. A character scanner does this for you and lists any zero-width or look-alike characters it finds.
  3. Check what a detector actually claims to do. If it says it finds a named provider's watermark, it usually needs that provider's key or API. If it does not have them, treat the claim with care.
  4. Do not treat a style score as evidence. A style number describes a guess about writing patterns, not a verified mark.

Common misconceptions

  • "No hidden characters found" means the text is unwatermarked. It only means the scanner did not match the characters it knows about.
  • "A watermark is always a hidden character." Many watermarks never insert a character at all.
  • "Every invisible character is malicious." Some are intentional, such as a joining control that a language genuinely needs.
  • "One detector covers everything." Removing characters, verifying a provider mark, and guessing authorship are three different tasks.

What this scanner covers and what it leaves to others

This tool focuses on the part that can be verified from the text itself: the presence of invisible, look-alike, or control characters, including bidirectional controls and mixed-script tricks. Everything it reports is something you can open, inspect, and confirm in the raw text. It does not attempt a statistical or stylistic verdict, and it says so instead of inventing one.

Bottom line: if you need to detect a specific vendor's statistical watermark, use that vendor's official detector or wait until they offer a public API. If you need to clean hidden characters from text, a character scanner is the right tool.

Related: Read the full limits in What a Character Scanner Cannot Prove, or see how hidden characters become dangerous in code in Trojan Source: When Text Order Lies.

← Back to AI Watermark Detector

Related Tools