Sebastian Raschka'swatermarking lecture explains how choosing a model's next word can also leave a detectable pattern. Anthropic says Claude's method uses a secret key and preceding words to guide the randomness involved in those choices. Nothing needs to be appended as a hidden character or visible label.
A detector checks whether the resulting sequence fits that keyed pattern. Anthropic's technical explanation, also covered in the earlier FAQ item, says this estimates Claude's involvement. It does not identify the user or distinguish original generation from substantial editing. Short passages and lightly proofread text provide fewer choices to inspect, so detection can be weaker.
