Watermarking in Amazon Bedrock - Amazon Bedrock
Models that add a watermarkLimitations
Services or capabilities described in AWS documentation might vary by Region. To see the differences applicable to the AWS European Sovereign Cloud Region, see the AWS European Sovereign Cloud User Guide.

Watermarking in Amazon Bedrock

A watermark is a machine-readable signal embedded in AI-generated content. Some models in Amazon Bedrock add a watermark to the content that they generate. In text, the watermark is a statistical pattern in the model's token choices. In speech, it is a signal embedded in the media. Watermarks are not perceptible to people. Instead, detection tools, where the model provider makes them available, can identify the watermark.

Models that add a watermark

The following table lists the models available in Amazon Bedrock that add a watermark to their output. The Default column shows whether the watermark is on when you don't change any settings. The Detection method column shows how to check content for the watermark.

Model provider Model Output type Default Detection method
Amazon Amazon Nova 2.5 Sonic Speech On Contact your AWS account team
Amazon Amazon Nova 2 Sonic Speech On Contact your AWS account team
Anthropic Claude Fable 5.1 Text On Anthropic watermark detection
Anthropic Claude Mythos 5.1 Text On Anthropic watermark detection
Anthropic Claude Opus 5.5 Text On Anthropic watermark detection
Anthropic Claude Sonnet 5.5 Text On Anthropic watermark detection
Anthropic Claude Haiku 5.5 Text On Anthropic watermark detection
Anthropic Claude Opus 5 Text On Anthropic watermark detection
Anthropic Claude Sonnet 5 Text On Anthropic watermark detection
Anthropic Claude Fable 5 Text On Anthropic watermark detection
Anthropic Claude Opus 4.8 Text On Anthropic watermark detection

Amazon Nova Sonic models

Amazon Nova 2.5 Sonic and Amazon Nova 2 Sonic add an inaudible watermark to all speech that they generate. The watermark is enabled by default, and you can't disable it. The text transcripts in their responses don't carry a watermark.

To request watermark detection for an audio file, contact your AWS account team. For more information about these models, see Amazon Nova 2.5 Sonic and Amazon Nova 2 Sonic in AWS AI Service Cards.

Anthropic Claude models

The Claude models in the table add a text watermark to their output. The watermark is enabled by default, and you can't disable it. You don't need to change your requests to receive it, and it doesn't change the format of the API response.

Anthropic offers watermark detection for eligible organizations. For more information about how the watermark works and how to request detection access, see How Claude marks AI-generated content on the Anthropic website.

Limitations

A detected watermark indicates that content was likely generated by a model that applies that watermark. A watermark is one input to customers' own transparency measures. It has the following limits:

  • A watermark does not prove who created the content, whether someone edited it, or whether the content is accurate.

  • A missing watermark does not prove that a person created the content.

  • Not every model in Amazon Bedrock adds a watermark, and editing, paraphrasing, translating, compressing, or cropping content can weaken or remove it.

  • Detection is probabilistic. Detection tools can miss a watermark or incorrectly report one, and detection is less reliable for short text.

  • A watermark is not a visible label or notice.

You are responsible for the content that you generate with Amazon Bedrock and for how your application uses it. This includes evaluating and meeting any transparency obligations that might apply.