Google found no feedback difference across 20 million outputs

Zvi Mowshowitz published a post on August 21 arguing watermarking AI text costs nothing in quality. The claim matters because providers have used quality as their excuse to slow-walk watermarking, while the EU's Code of Practice is forcing the question.

Every token a model generates already comes from a cloud of near-equivalent options, per Mowshowitz. Swap the random sampler for a keyed pseudorandom one, and the choice encodes a detectable signal in word choices that were arbitrary anyway. No quality cost, because there was never a meaningfully better option being discarded.

Google isn't running a hypothetical. It's shipped watermarking in Gemini for two years and reported no difference in user feedback across 20 million outputs, per Mowshowitz's post.

Anthropic rolled out its own watermarking quietly, to comply with the EU's Code of Practice, per the post. OpenAI intends to follow but looks likely to miss the deadline, per the same post.

The method isn't bulletproof. Paraphrasing strips the mark, in proportion to how much of the model's original phrasing survives the rewrite. Heavy paraphrasing erases it; light editing leaves most of it intact. The post doesn't say whether any provider is building paraphrase-resistant watermarking to close that gap.