Anthropic says it plans to apply invisible watermarking to text generated by Claude worldwide at launch. The move is connected to EU requirements for AI companies to mark AI-generated content, and Anthropic says it does not yet have a durable way to limit the feature by region. The watermark will not be visible to ordinary users and does not add hidden characters or visible markers to a response. Based on Google DeepMind's SynthID-Text approach, it works while Claude generates text: when several next-word choices are reasonable, a secret key and preceding words influence the randomness used to make the choice. Those choices should appear normal to readers, but can leave a statistical pattern across a sufficiently long passage. A detector with Anthropic's key can estimate how likely it is that Claude was involved in producing the text; it cannot prove who wrote it. Anthropic says internal testing found no practical effect on creativity, readability, or the content of Claude's responses. It also says the system requires no extra tokens and has a negligible effect on generation speed. Future Claude models will generate watermarked text, while models released before August 2, 2026 are expected to gain the feature over coming months. Watermarking is not applied where an exact output is required, such as a factual statement with only one correct answer. Code generally receives less watermarking because changing a term can break the result. Detection is more reliable for longer text and where the model has more reasonable word choices. Light proofreading may leave too little Claude-generated material for reliable detection, while Claude-generated translations carry a watermark because Claude chooses every word. A complete rewrite can remove it. Anthropic plans to offer a detection API. It will not distinguish between text Claude wrote and text Claude heavily edited, and it cannot identify text made by other AI models. For PNG, JPG, and SVG files, Claude will instead attach cryptographically signed C2PA provenance metadata stating that the file was created or processed with Claude. For Somali users, the useful distinction is between detection and proof: the planned system may indicate Claude's involvement, but it does not settle authorship on its own.