Free Mistral Watermark Remover & Checker for Text & Code
Paste Mistral-generated text or code, see every hidden watermark character exposed in an x-ray view, and copy a clean version in one click. Free, online, no sign-up, and nothing you paste ever leaves your browser.
More: what was found, x-ray view and options
X-ray view
Your text with every hidden character exposed as a labelled chip. Hover a chip for its Unicode name.
What was found
| Code point | Character | Count | Action |
|---|
Does Mistral watermark the text it generates?
No text watermark has been documented for Mistral's models or Le Chat at this page's verification date, and there's a structural reason it might stay that way longer than most: much of Mistral's catalogue ships as open weights, and a provider can't watermark text sampled on hardware it doesn't control. That doesn't mean copied Mistral output is clean, though. French-language typography and Le Chat's own formatting leave their own trail, which is what the checker below is for.
Last verified: 12 August 2026
Mistral occupies an unusual position among the providers this site tracks. It is the one major model lab headquartered inside the European Union, which puts it closest to the AI Act's transparency machinery, and at the same time it distributes many of its models as downloadable weights under permissive licences, which is precisely the distribution model that makes provider-side text watermarking hardest to deliver. Both halves of that tension are worth understanding before trusting any claim about marked Mistral output.
The record as it stands
| Question | Answer at the verification date |
|---|---|
| Has Mistral published a text watermarking scheme? | No |
| Is there a detector for Mistral-generated text? | No |
| Do Le Chat or the API insert hidden Unicode as a signature? | No evidence of it; paste a sample above and the tally will show every invisible code point present |
| Is Mistral subject to EU transparency rules? | Yes, as an EU-based provider of general-purpose models it sits squarely inside the AI Act's scope |
Why open weights and watermarks pull in opposite directions
A statistical text watermark is applied at sampling time: the software choosing the next token nudges its choices so that, over enough text, a keyed test can recognise the bias. That only works while the provider operates the sampler. Weights for models such as Mistral 7B and the Mixtral family are downloadable and run on anyone's machine with whatever sampling code the operator prefers. A mark implemented in the inference stack simply is not there when someone else's inference stack is used, and retraining or fine-tuning severs anything subtler.
The practical consequence: even if Mistral marked output from its own hosted endpoints tomorrow, text produced by the same weights running locally would be unmarked and indistinguishable in kind. Any detector promising to recognise "Mistral text" in general is promising something the distribution model rules out.
The EU angle is stronger here than for any other vendor
Article 50(2) of the EU AI Act requires providers of generative systems to mark synthetic output in a machine-readable form, and the general-purpose AI code of practice is the mechanism providers use to demonstrate compliance. For American and Chinese labs this is an export-market constraint. For a company incorporated in Paris it is home jurisdiction, so a deployed marking mechanism on Mistral's hosted services is a reasonable medium-term expectation. When one appears it will almost certainly be statistical or metadata-based, not hidden characters, and this page will record it.
French typography and the spaces that belong in the text
Mistral's assistant is widely used in French, and correctly typeset French places a narrow no-break space (U+202F) before question marks, exclamation marks, colons and semicolons, and inside « guillemets ». A plain no-break space (U+00A0) appears in the same roles in looser typesetting, and in numbers such as 12 500 €. These characters are not debris and they are certainly not a watermark; they are how the language is written.
Before cleaning French text: the checker's space normalisation converts every unusual space to an ordinary one, which is right for code and identifiers but flattens French punctuation spacing. Untick "Normalise unusual spaces" if the typography should survive; the tally will still show you exactly where each space sits.
What the checker can establish for Mistral text
Can
- List every invisible or ambiguous code point in text copied from Le Chat, the API, or a local Mistral deployment, with the reason each one matters.
- Distinguish typographic spaces that belong in French text from zero-width characters and bidirectional controls that never belong anywhere.
- Produce a cleaned copy whose visible content is unchanged.
Cannot
- Attribute a passage to Mistral. Nothing in the text identifies the model that produced it.
- Rule out a future statistical mark on hosted endpoints. That is a policy decision, not a property of today's text.
- Restore French punctuation spacing after it has been normalised away, clean a copy, not the original.
Questions about Mistral and watermarking
Is there any way to detect that text came from a Mistral model?
No. No watermark is documented, no detector exists, and open-weight distribution means identical text can be produced on infrastructure Mistral has never touched. Style-based AI detectors misfire on human writing and cannot name a model.
I found narrow no-break spaces in Le Chat output. Is that a signature?
Almost certainly not, U+202F is standard French punctuation spacing, and a French-first assistant produces it because the language requires it. A signature would need to appear where typography does not explain it, and no such pattern has been documented.
Will the EU AI Act force Mistral to watermark text?
It requires machine-readable marking of synthetic output from providers of generative systems, with the code of practice as the compliance route. How Mistral implements that on hosted services is its own announcement to make; weights already published cannot be retrofitted.
Does cleaning hidden characters remove a Mistral watermark?
There is no character-level watermark to remove: nothing in Mistral text identifies its origin. Cleaning fixes copy-paste damage, it neither hides authorship nor proves it.
Related reading
- Llama and Meta AI - the other open-weights case, with a research programme attached.
- SynthID and Gemini - what a deployed statistical text watermark actually looks like.
- What the EU AI Act requires of AI-generated content - the regulation this page keeps citing.
- The full character table - including both French spaces and their lookalikes.