“Perversion of writing:” Anthropic’s Claude Watermark Plan Draws Pushback From Daring Fireball’s John Gruber

Anthropic’s plan to introduce an invisible watermark into text generated by its Claude AI models is facing growing criticism, with tech blogger John Gruber saying that the technology could compromise one of the most fundamental functions of a language model: choosing the best possible words for a user.

Gruber, who writes the influential Daring Fireball blog, criticized the proposal on Sunday, describing it as a “perversion of writing.” His objection goes beyond whether users can detect the watermark. He argues that any system that changes how Claude selects words to make its output identifiable is allowing an objective other than the user’s needs to influence the writing.

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

The move by Anthropic was necessitated by growing tension facing AI developers as governments seek reliable ways to identify machine-generated content. Developers are being pushed to make AI outputs traceable while also trying to ensure that those mechanisms do not degrade the quality of generated text.

Anthropic has sought to address those concerns by arguing that its proposed watermark would be effectively invisible and would not alter the quality of Claude’s writing.

The company initially described the feature as an “imperceptible watermark” embedded directly into Claude’s output. Anthropic has linked the technology to compliance with the European Union’s AI Act, which requires certain AI-generated content to be marked in a machine-readable manner.

The company said other major AI developers will face similar requirements and will need mechanisms capable of identifying AI-generated material.

Anthropic provided more detail about its approach in a blog post published Friday, saying the watermark would not involve hidden characters, unusual formatting or other visible alterations that could affect how text appears to readers. Instead, the system would make subtle changes to Claude’s word-selection process to produce a statistical signature that can later be detected.

Anthropic says that the changes would occur among words that are essentially interchangeable in a particular context.

For example, the company said that after generating a sentence such as “The weather today was cold and,” Claude might choose “gray” or “overcast.” Both words convey essentially the same meaning, allowing the model to alter its statistical pattern without materially changing the sentence.

That explanation, however, does not resolve Gruber’s central objection.

From his perspective, the issue is not whether readers can notice the difference or whether Anthropic can demonstrate that the resulting sentence remains understandable. The issue is whether watermarking introduces a competing objective into the model’s generation process.

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

That criticism goes to the heart of how AI-generated writing should be evaluated. A language model normally optimizes its output according to a combination of factors such as relevance, coherence, accuracy, and the user’s instructions. Watermarking adds another consideration: whether a sequence of word choices contributes to a detectable statistical pattern.

Anthropic maintains that the trade-off is negligible because the watermarking system operates primarily when multiple words would produce effectively equivalent prose. Critics such as Gruber question whether an algorithm can reliably determine that two words are interchangeable in every context, particularly in writing where tone, rhythm, precision and connotation matter.

The distinction could become more important in professional writing. A journalist, lawyer, researcher or author may care about subtle differences between words that appear interchangeable to a statistical model. Even when two words communicate roughly the same idea, one may be more precise, natural, or appropriate for a particular audience.

The controversy has already generated concern among some Claude users, with reports of users cancelling subscriptions over the watermarking plan. Anthropic, however, told Business Insider that it had not seen a spike in cancellations.

The debate also exposes a difficult problem for the AI industry: machine-readable identification is easier to mandate than it is to implement without affecting the underlying technology.

Watermarking generated text could help publishers, educators and other institutions distinguish AI-assisted material from human writing. It could also make it easier to investigate the provenance of content circulating online, particularly as AI-generated material becomes more difficult to distinguish from human-produced work.

But the technology must contend with an important limitation. A watermark that depends on subtle word-selection patterns may become less reliable after text is edited, translated, paraphrased, or rewritten by another AI system. That means the practical value of watermarking may depend as much on how robust the detection mechanism is after content leaves Claude as on how accurately it works on untouched output.

For Anthropic, the challenge is therefore to satisfy emerging regulatory requirements without undermining the quality that users expect from Claude.

Gruber’s criticism is based on a fundamental concern: if watermarking influences even marginal word choices, users may reasonably ask whether the text is being optimized entirely for them or partly for the needs of the platform and regulators.