Reading Anthropic's Explainer on Claude's Text Watermark: Good for Spotting AI Writing, but I Worry About Quality
Anthropic published an explainer on how Claude's text watermark works. Future Claude models will generate text with a watermark, and the company says this is to comply with the EU AI Act. This post follows two parts of it: telling AI writing apart, and what happens to the quality of writing and code.
Before we start — key terms
Random numbers and the key. A model picks each next word from a list of candidates. When the candidates mean nearly the same thing, a random number settles it. With a watermark, the key and a few preceding words settle it instead. Checking the word sequence against the key later gives a probability that Claude wrote the text.
The EU Code of Practice. Its full name is the Code of Practice on Transparency of AI-Generated Content. It requires AI system providers to use methods of marking AI-generated text. Anthropic signed it in July 2026, among around 190 signatories.
SynthID-Text. A watermarking method Google DeepMind published in a 2024 Nature paper. Claude's watermark is a version of it.
I'll start with the part I think is good.
Watermarking uses low-stakes choices like these—which occur many times over a piece of generated text—to leave a pattern in Claude’s responses. That pattern is undetectable to the reader, but is detectable to anyone who has a key that encodes it.
Being able to pick out AI-written text seems good to me. The key only goes as far as the likelihood that Claude wrote part of a text, and detection doesn't work well on short samples. According to the article, the detection API is also a private preview, open to groups such as regulators, media and researchers.
Quality is the part I worry about. The article says there is no impact.
In internal testing, we’ve seen no impact of watermarking on the content, level of creativity, or readability of Claude’s text. In the SynthID-Text paper, which introduced the technique we use, Google DeepMind tested this impact by serving a model that used watermarking to a portion of their Gemini traffic and comparing thumbs-up and thumbs-down ratings. They found no statistically significant differences from the unwatermarked model.
That's the article's evidence, and I still worry about performance dropping, especially the quality of writing and code. Something I notice when I use Claude is that it often picks Korean words people rarely use.
The article also addresses word choice.
And it’s not the case that the watermarking method pushes Claude to choose a word it wouldn’t have considered anyway
Going by that, the word habit I noticed and the watermark may be separate things, and the article says nothing specific about Korean. Still, reading that the watermark touches how words get picked, I worry about quality, because Claude's Korean word choices already stand out to me.
The explanation for code is similar.
Having said that, in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code. But by definition, it will have a negligible effect on the actual code produced.
The article says the watermark isn't applied where an exact output is required. I still worry that code quality could drop too.
According to the article, Claude models launched before August 2, 2026 have a transition period, and watermarking will be added to them over the coming months.
