Anthropic's watermarking system treats all AI-generated text as potentially dangerous by default, which assumes users cannot be trusted to evaluate or take responsibility for what they read.