Future versions of Claude will embed a watermark in their generated text, allowing for the determination of the likelihood that Claude was involved in its creation. This implementation aligns with the EU AI Act requirements and is being adopted by several major AI providers. The watermarking method, based on Google Deep Mind's Synth ID-Text approach, subtly alters the random number generation process used when selecting words.

This ensures that the quality, content, and readability of Claude's outputs remain unaffected and indistinguishable to human readers. No characters are added, and the process does not incur extra costs or require additional tokens. The watermark itself carries no identifying information and cannot be traced to a specific user or chat.

It is designed to be a pattern detectable only with a specific key, indicating the probability of Claude's involvement. Limitations exist, as watermarking is less effective on short texts or factual passages where word choices are constrained, and it cannot definitively confirm human authorship or identify specific AI models beyond indicating a likelihood of Claude's generation. The watermark is also not applied in situations demanding exact outputs, such as in code generation or strict grammatical edits.