Anthropic's no-quality-loss claim for Claude watermarking challenged over response diversity

EU transparency rules are pushing labs to watermark AI text

The EU AI Act's transparency guidelines have taken effect, and most major AI labs have announced they will soon watermark their text output to comply, according to a blog post on blog.keyvan.net that reached the front page of Hacker News. In this context, watermarking means steering a model's word choices so a hidden, statistically detectable pattern marks the text as machine-generated. Anthropic is applying the technique to Claude, a move technology writer John Gruber attacked in a piece calling it a perversion of writing.

The argument over quality

As the post recounts, Gruber's position is that a writing tool should not give up even a sliver of clarity, coherence or meaning in order to embed hidden provenance clues. He received considerable pushback. Some respondents asked how quality could even be defined or measured in this setting. Others argued watermarking costs nothing, because it merely changes which random numbers guide generation. Developer Daniel Jalkut, whom Gruber later linked, offered an analogy: a random number generator that routinely reversed the digits of its output would remain exactly as random, and another commenter likened watermarking to swapping two faces of a die. Gruber was unmoved, saying he hoped the no-impact claim was true but thought it highly unlikely.

Anthropic states plainly that watermarking does not affect the quality of Claude's output. Its design builds on Google's SynthID-Text, described in a Nature paper in 2024, which similarly says the method preserves response quality.

The trade-off buried in the supplement

The blog post's core argument is that these guarantees were tested on single responses, while millions of people use these systems as ongoing conversations involving retries and repeated prompts.

A 56-page supplementary document linked from the research, the post notes, describes a diversity-detectability trade-off: the quality of each individual response is protected, but diversity across multiple responses is not. When several answers are sampled for the same prompt, watermarked answers resemble one another more than unwatermarked ones do. The researchers themselves concede this could hurt systems that generate many candidates and select the best. The post reads that as an admission that overall quality can fall, and as evidence that something more than a random-number swap is going on, since a pure swap would produce no such difference.

A small experiment

Technical consultant Natzir Turrado tested the claim directly by reproducing SynthID-Text on an open model with the paper's settings and a watermark key, writing in Spanish. Asked five times to describe a memorable trip to a coastal town, the watermarked model produced four answers that opened in almost exactly the same way, invoking a heart that sings or dances when recalling the town. The unwatermarked model opened three of its answers differently.

The post adds a more speculative concern: the watermark consults only a short sliding window of context when picking the next token, so the same convergence could surface across different prompts whenever two passages pass through an identical short phrase, and writing is full of those.

Where users would notice

The practical effects the post identifies include:

• Regeneration: pressing retry on an unsatisfying answer now tends to return similar alternatives rather than genuinely fresh ones.
• Creative writing: tasks that benefit from varied output lose range.
• Consensus checks: a common technique for gauging confidence and reducing hallucinations is to repeat a prompt and look for agreement, but watermarked answers are steered by the same key and are no longer independent, making the apparent consensus less trustworthy.

The post also points out that the paper's human evaluations appear to have covered single responses only, meaning the aspect most affected by the watermark may never have been tested with people. The author adds a personal aside: he already believes model writing has deteriorated since labs prioritized code generation, and watermarking risks narrowing it further.

Why it matters

Watermarking is on track to become the default compliance mechanism for AI-generated text under EU rules, so its side effects will be felt by essentially everyone who uses these tools, not just compliance teams. If the diversity loss described in SynthID-Text's own supplement holds at scale, everyday behaviours like regenerating an answer or sampling several responses become measurably worse, even while vendor marketing says quality is untouched. The episode also shows why single-response benchmarks are a weak proxy for real usage, and why claims of zero impact deserve scrutiny against the fine print. The debate between Gruber and his critics may feel abstract, but it now has regulatory force behind it.