Anthropic’s bid to make AI text easier to spot has opened a sharper question: when a chatbot merely helps polish someone’s work, who gets branded as the author?

The company began embedding an “imperceptible watermark” in text from supported Claude models released after August 2, framing the rollout as part of its commitments under the EU AI Act. The system borrows Google DeepMind’s SynthID-Text method: Claude makes statistically patterned choices among low-stakes, equally plausible words, leaving a signal invisible to readers but detectable with a key.¹

Anthropic argues the mark does not alter the reading experience or settle ownership. It says watermarking “does not impact the quality of Claude’s output,” and that a watermark merely tests whether Claude may have generated or processed material—not who owns it or who is responsible for it.² For light proofreading, the company says, there may be little for the marker to attach to because most words remain the user’s.

That assurance has not calmed critics. A review of the EU rules found that Article 50(2) exempts standard editing that does not substantially change a user’s words or meaning, raising questions over whether Anthropic’s global approach goes further than Brussels requires. Caroline De Cock called it “a scalpel to catch deepfakes and mass synthetic generation” being replaced with “a sledgehammer that hits simple proofreading and translation.”³

The dispute is practical as well as philosophical. Supporters see a useful check on undisclosed AI in schoolwork, job applications and online posts. Detractors fear a positive result will be treated as proof that Claude wrote everything, potentially jeopardizing client work, academic standing or copyright claims. As Deepset chief Milos Rusic warned, watermarks can be evidence, but should not become “an automated judgment about authorship or misconduct.”³

Developers have already tested the system’s limits. One open-source remover rewrites text to disrupt the word-choice pattern; its creator, Guillaume Meyer, says he supports attribution but objects to technology that “treats authorship as a binary thing.”⁴ Researchers say wholesale rewriting can always defeat a watermark—leaving Anthropic caught between a transparency rule it must meet and a label users may neither trust nor be able to keep.