AI text watermarking and quality loss

After the EU AI Act's transparency guidelines took effect and major AI labs announced plans to watermark text output for compliance, a debate has erupted over whether watermarking degrades AI text quality, sparked by tech writer John Gruber's post criticizing Anthropic's watermarking in Claude as "a perversion of writing." Anthropic claims "watermarking does not impact the quality of Claude's output," citing Google's SynthID-Text method published in Nature in 2024, which also asserts it "preserves response quality." Defenders like Daniel Jalkut argue watermarking merely alters random number selection without reducing randomness, but the author contends the quality-impact claim rests on single-response testing rather than the multi-turn conversations most users actually have. Anthropic's own 56-page supplementary materials reportedly acknowledge a "diversity/detectability trade-off," admitting that watermarked responses to the same prompt show greater similarity to each other than unwatermarked ones. The stakes are high given the EU mandate affecting most major AI labs, and the key question to watch is whether multi-turn quality degradation—reduced response diversity over extended conversations—draws regulatory scrutiny or user pushback as watermarking rolls out industry-wide.