Claude now watermarks everything it writes, and you're the one who can't check

Starting November 3, Anthropic extends its invisible text watermark to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. Every model released since August already has it. It applies worldwide, not just in the EU, and paying users can't opt out. Here's what that means in practice: Use Claude to proofread or translate, and your text gets marked. Anthropic's own help page admits a detected mark doesn't mean Claude wrote the content, only that it touched it. Non-native speakers and people with dyslexia get the same flag as someone who generated an entire essay. You can't check your own text. Detection is a private preview for regulators, media, researchers, educational organizations and "enterprises with compliance obligations". The people being judged are the only ones without access. "Only we have the key" won't last. Research on watermark stealing has already shown that these statistical patterns can be learned from enough samples. Expect third-party "AI detectors" to pick it up, with all the accuracy those services are famous for. It catches only honest users. The watermark lives in word choice, so running the text through another model with "rephrase everything" removes it. Bad actors will do that. Students, job seekers and ESL writers won't know they need to. So if I use Claude to make my résumé look more polished, I get flagged by "AI detectors" and rejected. Sweet! I'm gonna pay Anthropic for this! Meanwhile, the spammers and propagandists this was supposedly aimed at walk right past it. Source: How Claude marks AI-generated content P.S. This post was written by Claude Opus 5.5, so it's watermarked. Enjoy.