About this site

"AI watermark" names at least four unrelated technologies, and almost every argument about watermark removal is really an argument between two people talking about different ones.

Keeping those five apart is the whole editorial spine. A metadata cleaner does nothing to a pixel watermark. Rewriting text does not prove a statistical watermark was ever present. A detector score is a guess about style, not a decoded signal.

The commercial temptation on a domain like this one is to imply the tool defeats every watermark ever shipped. These are the lines we hold instead, and they are testable. If you find us over any of them, that is a correction and we will publish it.

• Our text checker finds and removes invisible Unicode characters. It does not remove Anthropic's statistical text watermark, and we do not claim it does. No public method is known to.
• We do not present community speculation as established fact. Reddit, Hacker News, and X threads are evidence of what people believe and search for, not evidence of how a watermark works.
• We do not publish a number without a source, and we do not manufacture data to fill a chart.
• When two primary sources contradict each other, we report the contradiction rather than quietly picking the more convenient one.
• We label a claim by its weakest link. A provider's marketing page and a peer-reviewed evaluation do not get the same badge.

No AI provider sponsors, reviews, or approves anything published here. Nothing on this site is paid placement.

As of the update date above, the site carries no advertising, no affiliate links, and no sponsored content. The tools are free and run in your browser. If that changes (sponsorship, affiliate relationships, a paid tier), the change gets disclosed on this page before it appears anywhere else on the site.

The one commercial interest worth naming plainly: this site ranks for queries about removing AI watermarks, and it benefits when people arrive looking for that. That is precisely why the limits of the tool are stated on the tool page itself rather than buried here.

Research starts from primary sources: provider documentation and support articles, official announcements, legislative text, academic papers, and patents. Secondary reporting is used to find those primaries, not to substitute for them.

Drafting and research are AI-assisted, which on a site about AI provenance would be absurd not to disclose. Every factual claim is checked against the cited source before publication, and the evidence label on each page reflects the weakest source supporting its central claim.

The full method is written out on the methodology page rather than summarised here: the evidence ladder, how volatile provider claims are re-verified, and how experiments are run.

• Research notes and source dossiers are kept per provider and dated.
• Every provider claim carries the date it was last verified, because these facts change monthly.
• Corrections are logged publicly rather than edited away.

Use the contact form. There is no published email address on this site, and that is deliberate: an unmonitored mailbox is worse than no mailbox, and a scraped address on a page about AI content invites exactly the traffic nobody wants.

For a factual error, include the URL and the sentence you are disputing, plus a source. Corrections get published in the corrections log with the date and what changed. See the editorial policy for how that works.

Research collaboration, dataset requests, and press questions go through the same form. If you are citing a page, the published and updated dates at the top of it are the ones to use.