AI text watermarking and quality loss

Major AI labs are implementing text watermarking to comply with EU AI Act transparency requirements, but there's debate about whether this affects output quality. While companies like Anthropic claim watermarking has no impact on quality based on single-response testing, critics argue this doesn't reflect real usage patterns where users have multi-turn conversations, and Anthropic's own documentation reveals a "diversity/detectability trade-off" suggesting quality may suffer across multiple responses to the same prompt.