Anthropic adds invisible machine-readable watermarks to all Claude outputs globally

Anthropic announced it will embed machine-readable, invisible watermarks in content produced by all its AI models worldwide. All new models will carry these watermarks from launch; older models have a grace period until December 2026. Text outputs carry embedded invisible watermarks, while other generated files include digitally signed provenance metadata where supported. The scope goes beyond what the EU AI Act strictly requires. A detection tool and technical guidance are planned for release so third parties can identify the watermarks. One limitation is that the model-level approach cannot distinguish wholesale AI generation from minor edits such as grammar fixes, so it may watermark content the EU law explicitly exempts. OpenAI also intends to comply with watermarking requirements but appears likely to miss the deadline.