TL;DR

Anthropic will alter the random element in how Claude chooses words, embedding a signal only it can read, to meet EU rules requiring AI text to be watermarked from December. A prominent critic argues this constrains the model into worse choices. A computer scientist at UCL says the objection misunderstands what those choices are.

What is being changed

When a model writes, small decisions carry no meaning — whether running water is a “stream” or a “brook”, whether a day is “grey” or “overcast”. Those picks are already stochastic. Anthropic’s watermark biases them into a pattern detectable by anyone holding the key, and the company says the effect is invisible to an ordinary reader.

Tech writer John Gruber disagrees sharply, calling the whole exercise “a perverse adulteration of what it means to write”. His objection is that constraining word selection must, on average, mean less precise selection.

Why the objection may not hold

UCL’s Steven Murdoch, who holds a chair in computer science there, expects no noticeable impact, and his reasoning is the interesting part. The randomness is not a flaw the watermark degrades — it is load-bearing. Strip it out and models “get stuck in loops and start repeating the same thing over and over again”.

Nor do these systems weigh register the way a writer does. A model does not select “brook” for its slightly antique feel; it lands there by chance. Murdoch’s summary is that the numbers were completely random before and are now statistically predictable but random still.

The reason to doubt the quality of AI prose lies elsewhere. A chemistry journal recently retracted a paper in which an AI tool had swapped in “mass killing of an ethnic group” as a synonym for “final solution” — in a passage about a zinc nanogel. Watermarking does not touch that failure mode.

Looking forward

Murdoch raises a second motive that has had less attention: so much synthetic text is already circulating that models trained on it risk collapse, confusing concepts as the pollution compounds. Detection is partly self-defence for the labs. The rules bind any company operating in the EU, so UK firms buying these tools inherit the change regardless of where the EU AI Act formally applies.