AI 安全與開發平台
New Claude Models Will Embed Text Watermarks During Generation, Including via the API and Claude Code
Anthropic will add model-level text watermarks to Claude models released after August 2, 2026, while embedding signed C2PA provenance data in supported files. The watermark can only indicate that content may have been processed by Claude; it cannot prove that a work was originally created by AI. No public detection tool or empirical accuracy results have yet been released.
Anthropic has announced a new scheme for labeling generated content. Supported models released from August 2, 2026 onward will weave human-imperceptible watermarks into text during generation. Because the marker is embedded in the text itself, it may persist after copying and pasting. The measure will apply globally across interfaces including the Claude app, API, and Claude Code, rather than being limited to EU traffic. Supported formats such as PNG, JPEG, and SVG will follow a different approach, incorporating digitally signed C2PA Content Credentials that record that the file was processed by Claude.
The implications for engineering teams extend beyond content publishing. API responses may carry the same signal when written into documentation, code comments, knowledge bases, or training data. Data pipelines will therefore need to preserve model versions, histories of human edits, and original provenance, rather than treating watermark detection results as direct evidence of authorship or plagiarism. Anthropic explicitly acknowledges that proofreading, translation, summarization, and format conversion may also leave a marker. Conversely, short text, fixed answers, excerpts, rewrites, or cross-language transformations may leave an insufficient signal. A detection result therefore means only that Claude may have been involved in processing the content, while a failure to detect the watermark does not rule out AI involvement.
Anthropic has not disclosed the watermarking algorithm, its false-positive rate, its reliability across different languages or in code, or a detector that third parties can use for independent verification. The related SynthID method adjusts the sampling probabilities of candidate tokens, but that does not establish that Claude uses the same design. The deployment primarily responds to the machine-readable marking requirements of Article 50 of the EU AI Act. Key issues to watch next include the migration timeline for older models, access conditions for a detection API, and the watermark’s practical effects on low-temperature decoding, structured outputs, and code generation.