John Gruber of Daring Fireball pens a barnburner that rails mightily against Anthropic’s plan to embed text watermarking in Claude-generated content, as mandated by European requirements:

One of my fundamental problems with this is that no two synonyms carry the exact same meaning. “He leaped at the chance” and “He jumped at the opportunity” are very similar sentences expressing the same general sentiment, but they are not the same. The exact words we choose when writing matter. I want any LLM I use to choose the very best, most precise words at every single decision point. […]

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

This isn’t just about text one might generate with the intention of passing it off as their own natural work. This isn’t even about LLM proofreading of work written by hand. Anthropic is saying that all new Claude models are going to adulterate every single bit of text longer than 200 tokens (~150 words) they generate, including everything it presents to its users to read. So even in a private conversation between a user and Claude, which will never be read by anyone other than the user, Claude will begin making word choices in the name of marking its output in statistically predictable ways rather than maximizing clarity and precision.

Even today’s so-called frontier models are already decidedly lacking in lucidity. Claude, ChatGPT, Grok, et al. are “better writers” than most humans and produce better prose than the median human. But: no shit. Most people are terrible writers. The “average person” is pretty stupid and half of all people are stupider than that. And there are many smart, interesting people who are miserable writers. So as impressive as LLMs are, the bar is low. The best writing I see come out of these models is worse than anything I would choose to read for pleasure. 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.

Gruber’s denunciation is forceful, persuasive, and wrong.

I don’t mean factually wrong. I mean his entire premise on which he builds his argument is misguided.

Gruber starts from the position that what Claude (and, by extension, any AI tool) generates is somewhat akin to human writing, and that, like a competent human author, it strives to select the best “right” words for its sentences. His argument is that by subtly changing which words are chosen for its generated text, Claude (etc.) creates lower-quality writing simply to satisfy the needs of watermarking.

His example sentences, “He leaped at the chance” and “He jumped at the opportunity,” do indeed carry different meaning to the reader, and the choice reflects both the writer and the writer’s preferred characterization of the person doing the leaping or jumping.

But Claude (etc.) isn’t making a literary decision. It’s using a statistical model to determine the next most likely token, and, for watermarking purposes, putting its statistical thumb on the scale in a predictable, recognizable way. The final sentence could just as plausibly be “He jumped at the chance” or “He leaped at the opportunity.”

Importantly, the final sentence can’t be judged as “worse” compared to something else, because we don’t know which alternate words were being considered—there is no something else. We can certainly opine as readers and writers on whether we ourselves would write “He jumped at the chance” rather than “He leaped at the opportunity,” but it’s hard to argue that one is inherently worse than the alternatives if we don’t know what the alternatives were.

Claude isn’t “writing” for clarity or creativity or precision. It’s not writing at all. It’s not pondering which word would have a given effect on a reader or best explain a concept. It’s selecting tokens from a list with a process that’s already biased by Anthropic’s models, and now that bias will be slightly more predictable and identifiable (to Anthropic, anyway). Arguing that one statistically generated alternative is “more” or “less” right ascribes human-like motivation—and capability—to an algorithm.