Story
August 15, 2026
Claude’s Invisible Watermark Could Turn AI Assistance Into a Scarlet Letter
Anthropic says its new Claude watermark will make AI use more transparent without changing what users read. Critics fear the global system could tag human work touched only by editing or translation—and still be easy to evade.
Claude’s coming invisible watermark is meant to expose machine-made writing. Its harder problem may be what it reveals—or appears to reveal—about people who merely used AI to polish their own work.
The shift follows the EU AI Act’s new transparency requirements, which took effect on August 2 and require providers to make generated or manipulated content identifiable. Anthropic says future Claude models will apply machine-readable marks worldwide, while older models are to be brought into the system later. Text will carry an embedded signal; supported files will receive C2PA provenance metadata.1
Anthropic’s account of the technology is deliberately reassuring. The company says the mark is not hidden text or an added character, but a pattern created through otherwise low-stakes word choices. It says readers will not be able to distinguish watermarked output from ordinary output, and that the system carries no information identifying a user, organization or chat.2
The company also concedes that the detector will answer a narrower question than its critics fear: “What is the likelihood this was partly written by Claude?”2 Detection becomes less reliable in short passages, factual material, code and heavily rewritten text. In Anthropic’s framing, a light grammar edit may leave too little model-written language for a mark to register; a full rewrite can remove it.
That distinction has not settled the backlash. Reporting on the rollout warned that communications teams using Claude to format, translate or clean up a human-drafted release could still find an AI signature attached to the final document.3 Ars Technica went further, arguing that a model-level system cannot reliably separate wholesale generation from “a comma fix,” potentially branding work the EU rules were designed to exempt.4
So the fight is not over whether audiences deserve clearer signals about AI. Both sides accept that goal. The dispute is whether Claude’s signal will clarify provenance—or flatten every degree of assistance into the same suspicious label.