Anthropic defends Claude's watermark, John Gruber opposes it

Anthropic plans to add watermarking to texts produced by Claude and assures that readability will not suffer. The company describes a statistical imprint, without hidden characters, designed to meet the European requirement to label AI-generated content. John Gruber, on the other hand, denounces it as a "perversion of writing" and refuses to accept that word choice should serve detectability rather than the user. Subscription cancellations have occurred, although the company claims not to see a spike.

Anthropic ensures detection without loss of quality or hidden characters

On Friday, Anthropic explained that the watermarking method does not affect the quality of Claude's outputs and does not employ invisible characters. According to the company, the imprint relies on slight modifications of word choices to create a detectable statistical pattern. It claims to limit these adjustments to interchangeable terms within context. For example, after "The weather today was cold and...", Claude might choose "gray" or "cloudy" without materially altering the meaning of the sentence. Anthropic further asserts that concerns about writing clarity are unfounded.

Stated goal: compliance with European law and "imperceptible" watermarking

Anthropic presents this watermarking as a response to the requirements of the EU AI Act, which mandates indicating, in a machine-readable format, that a text has been generated by AI. The company initially described the feature as an "imperceptible watermark" inserted directly into the text produced by Claude. It adds that other popular models will also need to adopt comparable techniques, in its view.

John Gruber rejects writing oriented towards detectability

In a post published on Sunday, John Gruber referred to the watermark as a "perversion of writing." He argues that Claude should choose its words based on user interest, not to facilitate detection. For him, precise word choice matters, even if he accepts inference time and cost per token as constraints. He insists that he wants an LLM to select the best possible words at each decision point and finds it offensive that other considerations influence generation. In his view, as soon as the watermark affects the lexicon, something other than the best prose guides the outcome.

User reactions and the company's response to cancellations

The project has raised concerns among tech enthusiasts and led some Claude users to cancel their subscriptions. However, Anthropic indicates that it does not observe a spike in cancellations. The company did not immediately respond to a request for comment.