Overview
- Anthropic announced that Claude instances released after early August will embed an imperceptible watermark in generated text and that the company is working to add the feature to older models and cloud deployments.
- The watermark is designed not to change meaning or readability and will survive copy and paste and some light edits, but heavy rewriting, translation, or mixing with other text can erase or obscure it.
- Anthropic will provide third-party detection tools so institutions can scan for the mark, but the company cautions that finding a watermark does not by itself prove Claude wrote the original work.
- The change responds to recent publishing disputes over alleged AI use in books and follows Google DeepMind’s earlier rollout of text watermarking, placing Anthropic alongside other labs building provenance tools.
- Publishers, schools and investigators may gain a new practical way to flag likely AI output, yet courts and editors will still need traditional evidence because the mark can persist from minor uses such as proofreading or be removed by edits.