Anthropic’s attempt to make AI writing more traceable has exposed a harder question: when does a helpful edit become evidence against the human who did the thinking?

The clash began as the EU’s transparency rules took effect on August 2. Anthropic said future Claude models would carry an imperceptible text watermark, using subtle choices among equally suitable words to create a pattern that can later be checked. The company argues that nothing visible or identifying is inserted, and that the system has “no practical impact on the quality or content” of Claude’s output.¹

Anthropic has applied the measure globally because it says it cannot yet reliably limit the feature by region. Its central caveat is also the source of much of the uproar: a detection result can indicate that Claude processed text, not that Claude authored it. Light proofreading may leave too little of a signal to detect, but longer edits, translations or summaries could retain one.²

That distinction landed badly with some paying customers. In the days after the announcement, users said they canceled subscriptions over fears that client code, academic work or original prose could be treated as AI-written merely because Claude had touched it. One consultant summed up the anxiety bluntly: “Even if you edit your own work that you wrote, they add the watermark anyway.”³ Anthropic says it has not seen a broader increase in cancellations.

Critics say the policy reaches beyond what Europe requires. The AI Act exempts ordinary assistive editing, yet a model-level system cannot cleanly tell a comma correction from a wholesale rewrite. Ars Technica warned that the marker’s practical message may be only that content “may have been processed by Claude,” a formulation easily misunderstood by employers, clients or teachers.⁴

Ben Thompson’s objection is more fundamental: transparency rules risk treating a tool as an author. Calling for a watermark on human-led work, he wrote, is like insisting “that a ballpoint pen advertise itself as the author.”⁵

Supporters see the same technology as overdue friction against synthetic-content confusion and AI systems training on their own output. Anthropic is promising a free detection API; whether that reassures users or turns Claude’s watermark into a scarlet letter will depend on how institutions interpret it.