Anthropic's Claude Watermark Explained: What It Does and Does Not Prove

Anthropic signed the EU AI Act's Code of Practice on Transparency and began embedding an invisible statistical watermark in text generated by Claude. The watermark has documented limitations: it can persist even in human-written text that Claude merely edited or translated, and its absence does not confirm a text was written entirely by a human. Commentary online has conflated this watermark with unrelated third-party AI detectors like ZeroGPT and with individual platform moderation decisions, treating them as a single automatic chain. Experts and careful readers note that detecting model involvement in a text is fundamentally different from concluding the text was 'written by AI.' The broader debate sidesteps a key distinction: whether a human retained editorial control over the idea, structure, and final output throughout the process.