How Claude’s Watermarking Works
Last week, Anthropic announced a new text watermarking system for Claude, designed to help readers identify content likely produced by the model. The system works by subtly favoring certain words during text generation, a preference not obvious to human readers.
Automated checking systems can then run statistical analysis to detect whether these specific words appear more often than expected in typical human writing. Anthropic maintains this will not change output quality.
Why Trust in Written Text Is Eroding
The need for this kind of transparency has become hard to ignore. AI-generated text now shows up everywhere, from LinkedIn posts to published books. LinkedIn itself has added options to report AI-generated content, and scandals involving authors passing off synthetic writing as their own continue to surface.
Accusations of using AI to generate text now carry real social and financial costs. Knowing where a piece of writing actually came from is becoming critical, especially as AI systems keep getting more sophisticated. Existing detection tools often produce false positives, since most rely on rough heuristics rather than anything definitive.
The EU AI Act Pushed This Forward
Anthropic’s move aligns with regulatory pressure already in motion. The company added this feature specifically to comply with the European Union’s AI Act, which mandates greater transparency in AI systems.
Not Everyone Is On Board
Reaction to the watermarking system has been mixed. Several users on Reddit pushed back, arguing the approach feels unfair, particularly for minor AI use like light editing. Others claimed the watermarks would be trivial to remove, catching only unsophisticated users while doing little to solve the actual problem.
Tech commentator John Gruber went further, calling the watermarking a perversion of writing.
He argued LLMs should be free to choose the best, most precise words
without external constraints, and that any intentional degradation of output quality, however slight, is patently offensive.
What Counts as Writing
Underneath the watermarking debate sits a harder question that predates Claude: what actually counts as writing? Some argue prompting a language model to produce text is fundamentally different from the human act of writing, which involves thinking, choosing, and revising in ways statistical text generation does not replicate. Others see that distinction as less clear-cut, especially as human editing and AI drafting increasingly blend together in real workflows.
Where each person lands on that question shapes how they feel about watermarks. If AI text and human writing are meaningfully different things, transparency about which is which starts to look necessary. If the line is blurrier than that, watermarking can feel more like a stigma applied to a legitimate tool.
Hashlytics Take
The most interesting part of this story is not the watermark itself but what it reveals about how unsettled the norms still are. Nobody has agreed on what disclosure should even look like: a full percentage breakdown, a simple badge, or nothing unless directly asked. Anthropic built for compliance with EU law, not for consensus on writing ethics, and that gap is exactly why both sides in this debate feel like they are talking past each other.
Follow Hashlytics on Bluesky, Facebook, LinkedIn , Telegram and X to Get Instant Updates
