Anthropic brings hidden markers to Claude outputs

Anthropic has added invisible, machine-readable markers to Claude-generated text and compatible files, enhancing provenance as AI companies face increased transparency requirements in Europe. These markings are intended to track generated material beyond the Claude interface, though extensive editing can potentially remove the detection.

The system is being implemented from the launch of Claude models on or after August 2, 2026, while older models will need additional support. Markings are applied separately to text and files. Text watermarks are imperceptible and woven into generated text, while file metadata uses the C2PA standard to indicate Claude processing.

While the presence of a watermark indicates potential Claude involvement, it does not confirm that Claude authored every sentence or idea. Users should understand that editing, paraphrasing, translation, or combining Claude output with other content can weaken or remove the detectable signal. Detection tools are still being developed, leaving independent researchers unable to assess the system's false-positive rates, robustness, or reliability across languages.