Would you fail a student based on a Claude watermark alone?
A student is accused of using AI to write an assignment. The university’s evidence is a Claude watermark. Is that enough to prove cheating?
Claude’s new models mark generated text worldwide. The method is a variant of DeepMind’s SynthID-Text. The watermark sits in patterns of word choice, not in hidden characters. It survives copy-paste, and supported files carry C2PA metadata. Anthropic plans a detector, but none is public yet.
The mark proves less than it seems. Detection gives a probability, not a verdict, and the mark applies only to words Claude chose. A student who asks Claude to translate their own draft carries a full watermark. A student who paraphrases AI output sentence by sentence may carry none. Anthropic itself states that a mark cannot separate “Claude wrote this” from “Claude edited this.”
At the same time, an open-source project called “watermarks-remover” gained 10,000 stars within days. It removes invisible Unicode and file metadata, and both steps can be verified. Neither step touches the text mark itself. For that, it offers a rewrite it calls “best effort,” which no one can test without Anthropic’s detector.
A detected mark shows contact with the model. It does not show authorship, intent, or cheating. A missing mark does not rule out AI use. Before watermarks enter misconduct cases, universities need independent false-positive and false-negative rates.
Would you fail a student based on a watermark alone?