Remove the AI patterns from your text

What is a hidden AI watermark?

The term covers two very different things. A real AI watermark, like the one Anthropic added to Claude's output in August 2026, is a statistical pattern woven into the word choices themselves. Nothing is added to the text, there are no hidden characters, and no character-stripping tool can remove it. AI detectors work the same way: they judge wording and rhythm, not hidden symbols.

What people usually mean, though, is invisible characters: zero-width spaces, word joiners and unusual spaces that occupy no visual space but survive copy and paste. These are real and worth removing, but they almost never come from the AI itself. OpenAI's models briefly left narrow no-break spaces in text in early 2025; OpenAI called it a training quirk, and it was gone within days. The invisible characters in your text today come from web pages, rich-text editors and documents, or from someone deliberately hiding data. The Unicode tags block can smuggle an entire readable sentence inside what looks like a single emoji, and this tool strips all of it.

Which characters does this tool remove?

Zero-width spaces, zero-width joiners and non-joiners, word joiners, soft hyphens, byte-order marks, left-to-right and right-to-left marks, bidirectional overrides and isolates, variation selectors, and the full Unicode tags range. It also folds fifteen kinds of unusual space, such as the non-breaking space and the narrow no-break space, back to an ordinary space.

Optionally it also normalises visible punctuation that reads as machine-written: curly quotes back to straight quotes, em dashes and en dashes to hyphens, and the single-character ellipsis to three full stops. That part is a style choice rather than a hidden-data problem, so it is a checkbox rather than a default you cannot turn off.

Will this get my text past an AI detector?

We cannot promise that, and you should be sceptical of any tool that does. Removing invisible characters is exact and verifiable: they were in your text, and afterwards they are not. Whether a given detector flags a given piece of writing is a separate question that depends on the detector, and no tool controls that.

What we can say is that most AI writing is identifiable from its rhythm rather than its characters. Sentences that all run fifteen to twenty-five words. Lists that always contain exactly three items. The "it is not just X, it is Y" construction. Opening sentences with Moreover or Furthermore. The optional rewrite targets those patterns specifically, which is why it changes how a piece reads rather than just swapping words for synonyms.

Is my text uploaded anywhere?

Not for the hidden-character cleaning. That runs entirely in your browser using JavaScript on your own machine. You can disconnect from the internet and it still works, which is the simplest proof that nothing is being sent. There is no account, no sign-up and no log of what you pasted.

The optional rewrite is different, because rewriting requires a language model and that model does not run in your browser. If you tick that box, your text goes to our server and on to a third-party AI provider. We do not store it. If the text is confidential, leave that option off and use the free path, which does the part that actually removes hidden data.

Why did an automation agency build this?

Because it took an afternoon, and because the useful half of the problem turned out to be about forty lines of code rather than an AI system. That is a pattern we run into constantly: a job that sounds like it needs a model usually has a deterministic core that should never touch one, wrapped in a thin layer that genuinely benefits. Getting that split right is most of what good automation is.

We should also say the obvious thing: we are not sure the instinct to strip a watermark is always the right one. Lara wrote about exactly that in AI watermarking is not the real problem, and the argument is worth reading before you use this. Removing invisible characters someone attached to your writing without telling you is one thing. Passing off work as something it is not is another, and no tool settles that for you.

If you want to know where the automate-or-not split sits in your own business, the AI Bottleneck Finder is eight questions and takes about two minutes. We also write up what we learn on the FreshStack blog.