Anthropic switched on watermarking for Claude output this month. Its own documentation says a mark means the content may have been processed by Claude. Not that Claude wrote it. And the mark disappears if the text is paraphrased, translated, or mixed into other writing.

Which means the student who pastes an answer gets flagged and the student who runs it through a second tool does not.

I wrote up why the fix is a different request rather than a better detector: https://www.harvestkernel.com/claude-watermark-authorship-not-detection/

One question for you, and it takes one sentence to answer:

What is the single assignment in your course where you would have no idea whether the student actually did the thinking?

Name it. That is the one worth rebuilding first, and I will tell you exactly where I would start with it.