Tech writer John Gruber has denounced Anthropic's decision to watermark every Claude response, a change driven by an EU code with roughly 190 signatories. Key Points: Gruber says the watermar
Tech writer John Gruber has denounced Anthropic's decision to watermark every Claude response, a change driven by an EU code with roughly 190 signatories.
- Gruber says the watermark trades precise word choice for detectability, and calls the result a corruption of writing itself.
- Anthropic says its method copies Google DeepMind's SynthID-Text approach and has no practical effect on output quality.
- Some paying subscribers are canceling, and a promised detection service still has no release date.
Gruber argued in a post published Sunday that the system corrupts the craft of writing itself. He objects to the mechanism, not the goal. The watermark steers Claude toward particular words at moments when two or three candidates would each fit the sentence, and no two synonyms carry precisely the same meaning. Private chats get the same treatment as published work. He also calls the European rule futile, since anyone determined to hide their tracks can run the output through a rewriting tool.
His verdict was blunt, and he described the plan as "a perversion of writing."
Also Read:iPhone Ultra Launch Faces Severe Shortages From 10M Unit Order
Anthropic explained on Aug. 14 that its method borrows from Google DeepMind's SynthID-Text, altering only the source of the randomness that settles low-stakes word choices.
The company says the technique "does not have any practical impact on the quality or content of Claude's outputs." It withholds the nudge wherever a single answer is correct, including arithmetic and working code. Comments buried inside code remain fair game for the mark.
The company also stresses that the mark carries no identifying information and cannot be traced to a person, an organization or a chat.
Verification depends on a key that only Anthropic holds. Outside parties can therefore check nothing until a promised detection service arrives, and no release date has been set. Detection weakens on short passages, where too few word choices exist to support a confident reading. Anthropic concedes that proofreading, summarizing and translation can all leave a mark on work that began in someone else's hands.
Dozens of paying subscribers began canceling their plans this week, describing the mark as a "scarlet letter" that light rewriting fails to erase.
The fight has built over roughly two weeks. Transparency duties under Article 50 of the EU AI Act took effect on Aug. 2, carrying penalties of up to 15 million euros or 3 percent of global annual turnover.
Anthropic first updated its support pages, then published a fuller technical explainer on Aug. 14 after readers complained that the earlier document described nothing at all.
Gruber's essay landed two days later. Reports of cancellations followed within a day, and older Claude models still carry no watermark, though the company says it will extend marking support to them over the coming months.
Read Next:XRP Bulls Bet $2.16B On Rebound As Institutional Demand Stays Weak
