Key facts
- Tech blogger John Gruber has criticized Anthropic's plan to watermark AI-generated text from Claude.
- Gruber argues that watermarking compromises the AI's ability to select the most precise words.
- Anthropic states its watermarking method is imperceptible and does not affect output quality.
- The company intends the feature to comply with the European Union's AI Act requirements.
- Anthropic claims the watermarking involves subtle alterations in word choice among interchangeable options.
Tech blogger John Gruber has joined a chorus of criticism regarding Anthropic's intention to add watermarks to text generated by its AI model, Claude. Gruber, who writes the popular blog Daring Fireball, described the plan as a "perversion of writing" and "patently offensive."
Gruber's primary objection is that the watermarking process could influence the AI's word selection, prioritizing detectability over the most precise and effective language. He stated, "I want any LLM I use to choose the very best, most precise words at every single decision point." He believes that any factor other than the user's needs influencing text generation is unacceptable.
Anthropic has defended its approach, stating that the watermarking will be imperceptible and will not impact the quality of Claude's output. The company explained that the method involves subtly altering word choices among effectively interchangeable options to create a statistical pattern for identification. For example, when generating a sentence about the weather, Claude might choose between "gray" or "overcast" for a word, a choice Anthropic argues does not materially change the sentence's meaning.