John Gruber calls Claude’s AI watermarking ‘patently offensive’

• Anthropic is adding a watermark to its LLM outputs to comply with new EU AI rules.
• Tech blogger John Gruber isn’t sold on their approach.
• In a Sunday blog post, he called the new watermark a “perversion of writing.”

Tech blogger John Gruber is joining the chorus of dissent over Anthropic’s plan to watermark Claude’s writing.

On Sunday, Gruber, the popular blogger who writes Daring Fireball, called watermarking plans a “perversion of writing.”

His main criticism: Claude should choose words because they’re best for the user, not because they help make the output detectable.

“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. An obvious constraint that I accept is time and computation. Within the constraint of executing inference quickly, and at a certain cost per token, I want the best words. The idea that anything other than my needs should factor into the generation of text for me is patently offensive.”

In blog posts, Anthropic says that concerns around writing clarity are unfounded.

The disagreement highlights the confusion surrounding Anthropic’s watermark plan, which has sparked concern among techies and prompted some Claude users to cancel their subscriptions (though the company told Business Insider it hasn’t seen a spike in cancellations).

Anthropic initially described the feature as an “imperceptible watermark” woven directly into Claude’s text. The plan was hatched to comply with the European Union’s AI Act, which mandates that all AI-generated text must be marked in a machine-readable format. Anthropic wrote that other popular models will have to use similar watermarking techniques.

On Friday, in a post titled “How Claude’s text watermark works,” the company wrote that its watermarking method does “not impact the quality of Claude’s output,” and doesn’t use hidden characters.

Instead, Claude’s word-selection process is subtly altered to create a statistical pattern that can later help identify its involvement. Anthropic argues that those changes occur among words that are effectively interchangeable in context.

If the LLM writes a sentence like “The weather today was cold and…”, Claude is likely to choose “gray” or “overcast” for the next word, the company wrote. The sentence’s meaning doesn’t materially change with either word, Anthropic argues.

Gruber isn’t buying it: In his view, once watermarking influences Claude’s word choice at all, something besides producing the best possible prose is shaping the result.

“The exact words we choose when writing matter,” he wrote. “I want any LLM I use to choose the very best, most precise words at every single decision point.”

Anthropic didn’t immediately respond to a request for comment.