Anthropic's 'Watermark' Text Adulteration in Claude Is a Perversion of Writing

Anthropic has confirmed that its Claude models will soon apply a semantic text watermark to all generated content, a move intended to comply with an EU regulation. The company’s initial announcement, titled “How Claude Marks AI-Generated Content,” offered no technical details and claimed the watermark would be imperceptible and would not change meaning, quality, or readability. Tech writer John Gruber, however, has sharply criticized the plan, calling it a “perversion of writing” after learning how the system actually works.

Instead of inserting invisible characters, Anthropic will use steganography based on word choice. At inference time, the model will be slightly biased toward picking words from designated “green” or “red” lists at each token-generation step. These choices leave probabilistic fingerprints that can later be detected, allowing AI-generated text to be identified without visible markings.

Gruber, writing on his blog, explained that he initially believed Anthropic’s promise of perfection and guessed the company would hide invisible Unicode characters. In a follow-up article, “How Claude’s Text Watermark Works,” Anthropic revealed the semantic technique in layman-friendly terms. Gruber argues that this approach necessarily adulterates the text by constraining token selection, contradicting the company’s earlier assertion that output quality would remain untouched.

The implications are significant for writers and businesses using Claude. If the watermark subtly alters phrasing to enable detection, it may degrade clarity, coherence, or stylistic nuance in AI-generated content. Gruber insists that any tool that sacrifices an iota of meaning for provenance tracking is unacceptable for personal and professional use.

Reactions from the AI community have been mixed, with some praising the transparency effort and others echoing Gruber’s concerns. Gruber points to an interactive essay by James Padolsey, “How AI Text Watermarking Works,” as the clearest explanation of the technique. Anthropic’s own description, he says, remains euphemistic and slightly misleading.

Looking ahead, regulatory pressure in the EU is likely to push more AI companies toward detectable watermarking. The challenge will be balancing compliance with preserving the unimpaired quality of generated text. Whether Anthropic can deliver on its promise of “imperceptible” and harmless watermarking remains to be seen.

Turn any website into clean JSON.

Ready-made scrapers & AI agents

Try Apify