AI text watermarks:
what is actually in your text?
A hidden character, a statistical watermark and an AI detector score are three different things. Choosing a tool starts with knowing which one you are dealing with.
Three different questions
| Question | What it measures | What AIWatermark.dev offers |
|---|---|---|
| Are there hidden characters? | Specific Unicode code points in the string. | Local inspection and selected character cleanup. |
| Is a model watermark present? | A model-specific statistical pattern over a passage. | No official signal detection or removal certification. |
| Does this sound AI-written? | A classifier’s estimate, often based on writing patterns. | No AI probability or detector score. |
Hidden Unicode characters
U+200B is a zero-width space. U+FEFF can be a byte order mark. Other characters affect hyphenation, direction, letter joining and emoji presentation. Editors and webpages can introduce these characters during copying, so they do not establish AI authorship.
A character cleaner can show its work: which code point occurred, how many times, and whether it was removed, normalized or preserved. AIWatermark.dev’s invisible character remover and zero-width space remover do that locally.
Statistical model watermarks
OpenAI describes textGrain as a pattern in word choices, without adding hidden characters. Anthropic likewise describes Claude’s text watermark as statistical. Google DeepMind’s SynthID includes text watermarking. These systems need their relevant detection method; looking for Unicode artifacts is a different operation.
Whether a passage carries a watermark depends on its source, model and rollout. A generic cleaner should not tell you a statistical signal is absent just because it found no invisible characters.
What happens when you rewrite text?
Rewriting changes the sequence and structure of words. That may weaken a statistical pattern, but results depend on the scheme and passage. Small edits can leave signals intact. A rewrite can also introduce errors or a new model’s signal.
AIWatermark.dev’s optional rewrite is therefore a writing transformation, with no detector certificate. Use it only when you want new wording. Compare facts, numbers, names and citations before using the result.
Choose the workflow that fits
- For odd spacing, copying errors or mismatched strings, inspect the characters first.
- For a better draft, clean artifacts and revise the wording yourself or use the optional rewrite.
- For provenance verification, use the relevant provider’s authorized detector when access is available.
- For work with disclosure requirements, follow those requirements regardless of a cleaner or classifier result.
Provider-specific text guides
Read the Claude watermark remover guide or the ChatGPT text watermark remover guide for product context. For the full local tool, return to the AI text watermark remover homepage.