Anthropic embeds invisible watermarks in Claude to meet EU AI law

Anthropic launched this week an automatic watermarking mechanism on every text generated by Claude. The invisible code is embedded in the chatbot’s responses and flags them as AI-generated so automated detection systems can identify them. The move is intended to satisfy the “transparency clause” of the European Union’s AI Act, which took effect and obliges technology companies to label content created or edited by AI in a mechanically detectable way. Regulators in Europe may be pleased, but some users are not.

On Reddit a heated discussion is underway. User visionode, whose account is three weeks old, posted an angry thread calling the system a “dragon-scale conspiracy” designed to incriminate innocent users. He argues that sophisticated users could hide Claude usage by paraphrasing or routing the text through other AI services, but the average user will be caught. He gave examples: a student who asked to reorganize a paragraph, a journalist who asked to summarize a 200-page transcript, a writer stuck in a creative block who asked for synonyms. “They will emerge from the process with a digital tattoo on their forehead,” he wrote.

Those claims do not hold water. A journalist who uses Claude to summarize a long transcript will not be harmed by the watermark unless he copies the summary word-for-word into an article, which is already unethical and unlikely. The same applies to a student who pastes Claude’s output into a paper after asking to “reorganize a paragraph.” Other Redditors were unimpressed by the outcry: “Look at this guy,” one replied. Another suggested the OP “take a deep breath.” A third commenter noted sarcastically that the complainant probably used Claude to write the complaint itself.

A slightly more reasoned criticism came from another angle: hypocrisy. One user claimed that marking edited output that “steals” others’ work is a “very evil” direction. “I don’t use Claude to write anything, but an AI that marks your work is terrifyingly ironic, given how much training data these models have siphoned,” he wrote. On the other hand, most support the measure: “There is no good argument against this,” a commenter summed up in another thread. “The only reason to oppose is to lie to people.”