Check some text yourself
This runs the same watermark test on whatever you paste. The calculation happens in this tab; your text is not sent or uploaded.
What the result can tell you
Most pasted text will fall below 1.80 because this checker looks for one published watermark. Text can also cross the cut-off by chance.
Whether the result is above or below the cut-off, it does not establish where the text came from, who wrote it or whether a model was involved.
This checker needs JavaScript, because the calculation runs in your browser rather than on a server. With it switched off there is nothing to send and nothing to show.
Why there is no remover
We have not added a general remover. The mechanism is published here so people can inspect it, but packaging it for arbitrary text would turn an explanation into an evasion tool.
Swapping letters for lookalikes changes how text is split into tokens. At low editing rates, blind changes weakened the stronger watermarks in the cited research without clearing them; at much higher rates they defeated every scheme tested. A trained stand-in for the detector made the attack roughly twice as successful at the same editing budget.
Ordinary AI-text detectors have different failure modes. A separate 2023 study found that seven detectors wrongly flagged a mean 61 per cent of the non-native English essays it tested. Its authors bypassed those detectors by rewriting prompts, not by changing characters.
The character mapping above is much narrower. It removes one zero-width character and maps five Cyrillic lookalikes to Latin. Either operation can damage legitimate writing, so inspect the result before using it.