Summary

Gruber argues that Anthropic's text watermarking scheme — which subtly biases word choices at inference time via 'green' and 'red' token lists — is a fundamental corruption of writing. He is doubly furious because Anthropic's original support document explicitly claimed the watermark would be 'imperceptible' and wouldn't 'change the meaning, quality, or readability,' which is provably false given how the technique actually works. He dissects the EU regulation driving the decision and finds it impractical nanny-state overreach that will only burden honest users while bad actors trivially circumvent it via paraphrasing tools. He also faults Anthropic for applying this globally rather than scoping it to the EU, and argues this reveals either alarming technical incompetence or a deliberate choice by a company on the eve of a $2 trillion IPO. The secrecy of the scheme — only Anthropic can detect its own watermarks — is what he calls the 'poison': users cannot know which words were chosen for meaning and which were chosen for fingerprinting.

Key Insight

Semantic watermarking is not a neutral technical feature but an act of writing adulteration — it structurally requires sacrificing optimal word choices for statistical fingerprinting, and Anthropic's claim that this is imperceptible and quality-neutral is either a lie or a confession that precise word choice doesn't matter, which is itself the deepest offense.

Spicy Quotes (click to share)

• 8

  The idea that anything other than *my* needs should factor into the generation of text *for me* is patently offensive.

• 9

  And now Anthropic is saying they're going to make it worse, on purpose, for purposes that do not benefit me in any way? Even if only slightly worse? Get fucked.

• 8

  By definition it *must* make text worse, unless the underlying LLM model's scoring is wrong, because the nature of the watermarking algorithm requires it to sometimes increase the probability of selecting a *worse* word choice and decrease the probability of selecting the model's *best* choice.

• 9

  Arguing that *grey* vs. *overcast* 'doesn't matter much to the reader' is the crux of my argument that this entire endeavor is a perverse adulteration of what it means to write — or to read. That it's subtle in some ways makes it *more* perverse, because it's sneaky.

• 8

  Translation: *We value precision in programming code; we do not in prose.*

• 9

  It is exceedingly rich to cite George Orwell's *Nineteen Eighty-Four*, approvingly, in the context of justifying a text adulteration scheme premised on the notion that specific words do not matter.

• 7

  Secrets are the poison here. When only Anthropic holds the secret keys that both produce the watermarking *and* perform the probabilistic detection of those marks, we're all left to wonder.

• 8

  What Google's thumb-counting data shows is only that it isn't so much worse as to make Gemini users click the thumbs-down button.

Tone

outraged, polemical, incisive