Claude's watermark didn't go a day without complaints. The cancellations are here, along with the fear we already wrote about

Yesterday we covered the announcement and put the caveat in writing: Claude's invisible signature doesn't stop whoever mass-produces junk, it points at whoever uses AI well and in the open. We didn't have to wait long for proof. Within 24 hours, the conversation on X and Reddit has swung from transparency to revolt, with users calling the move 'hugely problematic' and the first public subscription cancellations rolling in.

The cancellations are here, and the reasons sound very familiar

The profile of the angry user isn't the one churning out automated articles by the shovelful, it's the one using Claude as a tool. Forbes tells the story of US radio host Erick Erickson, who had replaced Grammarly with Claude to proofread his own writing and now fears his prose, his from start to finish, will be flagged as AI-assisted. TechCrunch sums it up: users are furious the mark 'will catch them using it at their jobs and classes'.

In Spanish, the case that travelled furthest is Arturo Villa's, who announced the cancellation of his Max subscription with a five-reason list headed by the watermark: 'nobody in their right mind should accept being marked and singled out for using a tool that is essential nowadays and that everyone uses'. The impossibility of switching it off, even on the most expensive plan, is the detail drawing the most blood.

And Reddit delivered the day's most ironic scene. The angriest thread was titled 'Claude watermarking our work is unethical and disgusting', with a reasonable argument inside: 'I gave the instructions, context and decisions, Claude was the tool'. The punchline came from the commenters, who smelled that the complaint itself was written with Claude. 'Bro couldn't even complain about Claude without using Claude to write it', reads the top reply, with 1,100 upvotes.

The thread ended up deleted.

The most serious critique isn't about spies, it's about labels

There is also analysis with more substance, and the one that travelled furthest separates the wheat from the chaff: the problem is not that Claude watches you, because the mark can't trace your account or your conversations, it only shows a text passed through the model. The problem is interpretation. That 'processed by Claude' becomes 'written by AI' in the eyes of a boss, a client or a university panel with no appetite for nuance.

There are two more technical snags. Quality: the mark works by nudging the model's word choices, and Anthropic insists it's imperceptible, but it hasn't published technical data or independent testing to back that up. And effectiveness: anyone determined to erase it can do so with a deep rewrite, so the signature will end up marking, above all, the honest user asking to polish an email. EU rules, in fact, exempt standard editing that doesn't change a text's meaning; Anthropic's implementation appears to cover considerably more than that minimum.

Anthropic stays quiet and the ball sits in the detector's court

For now, the company hasn't publicly answered the anger beyond its support page, which already acknowledged the limitations. We'll know more in the coming weeks: whether they ship the promised detector and with what guarantees, whether they open an exception for proofreading use, and whether OpenAI and Google switch on their own watermarks or let Anthropic pay the price of going first alone. Yesterday's bet stands, and after day one of the reaction, firmer than ever: the cat and the mouse have only just started running.