Historia
agosto 15, 2026

Claude’s Invisible Watermark Promises Transparency but Blurs Who Really Wrote the Text

Anthropic will watermark future Claude output under EU transparency rules, but the system’s own limits have sharpened fears that a light AI edit could be mistaken for AI authorship.

Anthropic’s answer to the AI-authorship problem is an invisible mark that follows Claude’s words. Its complication is just as invisible: a detection result may reveal AI involvement without establishing who actually wrote the work.

The shift follows the EU AI Act’s August 2 transparency obligations. Anthropic said future Claude models would mark generated text and supported files globally, with text carrying embedded watermarks and files using signed provenance metadata; older models are still being brought into compliance.

The company says the text signal is not a hidden character or a visible tag. Instead, it uses low-stakes word choices to create a statistical pattern, based on a version of Google DeepMind’s SynthID-Text approach. Anthropic argues that the result has “no practical impact on the quality or content” of Claude’s responses and can only indicate the likelihood that Claude was involved.

That distinction has become the fault line. Reporting on the rollout noted that a document could carry a mark after Claude merely proofread, formatted or translated human-drafted copy—while heavily rewritten, mixed or short passages may lose a detectable signal altogether. Anthropic says light grammar-only edits will usually leave too little Claude-selected text for reliable detection, but critics fear recipients will not parse that nuance.

The concern is especially sharp for schools, publishers and workplaces, where a mark could be read as an authorship verdict. One critique called the policy a “nuke it from orbit” approach: a model-level system that may stamp editing work the EU rules were designed to exempt.

Users have split along a more practical line. Some Reddit commenters cast the watermark as a “digital tattoo” for students, writers and workers who use Claude for modest assistance; others replied that it is not claiming credit, but identifying AI-generated output where the risks warrant scrutiny.

Anthropic plans a detection API, but until outside parties can test it, the watermark’s central promise remains constrained: it may make Claude’s presence easier to spot, while leaving the harder question of human authorship unresolved.

Cobertura de la historia