Frequently asked questions

It removes markings from text and files that are not visible in them: invisible Unicode characters, metadata in images and documents, C2PA signatures, provenance entries in HTML and Markdown. You upload a file or paste text, get the cleaned result back, and with it a report of exactly what was removed.

No and no. There is no sign-up, no paywall and no per-person quota. The only thing that needs a key of your own is the optional LLM rewrite — more on that below.

Layer A is deterministic. The marking consists of characters, or of fields inside the file, so it gets deleted, counted, and the file is checked again afterwards. What the report shows can be verified.

Layer B is an attempt. Some watermarks sit in the word choice itself and cannot be touched mechanically; the only thing that helps is having a language model rewrite the text. It works, but it comes with no promise.

Images, Office and OpenDocument files, PDF, SVG, HTML, Markdown, source code and plain text. The exact current list is on the home page under “What gets detected and removed” — it is not maintained by hand but comes from the same patterns the cleaner searches for.

For invisible characters and metadata: yes, that can be shown. The file is read back after cleaning, and the report names every character removed.

For statistical watermarks in the word choice: no. The key belongs to the vendor. Without it nobody can check whether anything is left — us included. Anyone promising otherwise cannot know.

No, and that matters enough to be its own question. A removed mark means: this mark is gone. It does not mean no AI was ever involved. Presenting the report as proof of human authorship — in an exam, to a client — presents something that does not carry that claim.

No. For images raimove removes metadata and C2PA segments, not the mark in the pixels. That one only goes away by having an AI recompute the image, which needs substantial computing power and changes the picture. It exists as an optional add-on in the core project, not in this web app.

C2PA provenance comes in two designs. One sits inside the file and can be deleted. The other sits on the vendor's server, and the file only carries an unobtrusive mark pointing at it. After cleaning, the enclosure is gone — but that mark could be used to reattach it. Hence a warning rather than a success message.

No. An uploaded file is processed in memory or in a temporary folder and deleted immediately after the response. That folder lives in RAM, not on the server's disk. There is no database for any of it to end up in.

Two exceptions worth knowing. Messages sent through the contact form sit in the server log until it is overwritten — that is stated there too. And if you request a token for the MCP server, your address is stored alongside it, together with counters for calls and data volume. Without that a key could neither be attributed nor switched off if it were abused. What gets cleaned through the key is not recorded.

Both are deleted on request: one message through the contact form is enough.

Because every rewrite costs money at an LLM provider. A service with no accounts that picked up the bill would be drained on day one. Your key is passed through for that single call, never stored and never logged.

Yes. The cleaning itself comes from the watermarks-remover project (MIT licence), linked in the footer of every page. The footer also names the version running here.

In metadata: yes — the vendor list on the home page comes straight from the patterns being searched for. In the word choice: that is Layer B, and the limit above applies. The question “is anything still there?” cannot be answered without the vendor's key.

For content you own: your own drafts, your own research, your own files that got stamped along the way without being asked. That provenance labelling is broadly useful does not contradict that — being pro-privacy is not the same as being against transparency. What follows in a particular case, towards a university or a client, is a question of their rules rather than of the technology.

There is an MCP server for that, letting an agent — Claude Code, for instance — clean text and files directly without going through this page. Access is open: enter an address in the section at the end of the home page and the token arrives by email straight away.

Tell us through the contact form — ideally with the file type and, if you know it, the application the file came from. That is exactly how the detection list grows.