Anthropic Adds Invisible Watermarks to Claude Output — Coders Bypass Them in Four Hours

Anthropic's move to watermark all Claude-generated text and files is a direct response to new EU transparency rules requiring AI outputs to be detectable. The watermarks are invisible, embedded in both text and file outputs, and intended to let downstream systems identify AI-generated content.

The rollout immediately ran into the fundamental problem with text watermarking: within four hours of the announcement being confirmed, coders were touting working overrides and workarounds, as Wired reported. Unlike image watermarks, textual signals can often be stripped by paraphrasing, re-formatting, or trivial transformations, making robust detection extremely hard.

Competitively, this is a compliance-driven feature rather than a capability play, and it positions Anthropic as trying to lead on EU regulatory alignment—consistent with its broader safety-forward branding this week ($65B run rate disclosure, risk raised to 'low'). But shipping a detection mechanism that's defeated within hours risks looking performative.

The community reaction was immediate and skeptical: threads argued the episode proves AI-text detection may be technically infeasible at all, and criticized the EU rule as unenforceable. For enterprises relying on watermark-based provenance for compliance, the takeaway is caution—don't treat the watermark as tamper-proof. Watch whether the EU accepts easily-defeated watermarking as compliant, and whether Anthropic strengthens the scheme.