Stealthy Watermarks: Anthropics New Technique for Marking AI-Generated Text
In this article (6)
Anthropic, a pioneering company in the field of Artificial Intelligence (AI), has now started including unique, invisible watermarks on AI-generated text by their most recent models. This comes as a response to the European Union’s AI Act that came into effect on August 2, mandating transparency for AI-produced content.
Invisible Watermarks in AI-Generated Text: The New Norm
As outlined in an update on Anthropic’s official support page, all new Claude models introduced post August 2 will incorporate a system known as “machine-readable marking” from their inception. These watermarks, while imperceptible to human users, would be incorporated directly into the generated text by any supported Claude model.
These markers don’t affect the meaning, quality, or readability of the AI’s response, and yet remain part of the text, enabling them to move along with the text even when it’s copied, pasted, or edited elsewhere. It’s important to note that these watermarks will be applied at the model level, which implies that they will be present irrespective of the Claude product or platform the text originates from.
Universally Applicable Watermarks
The invisible watermark will be seamlessly embedded into any text generated, remaining undetectable by users. The company maintains that this marking system will have no impact on the nature or quality of the AI’s response. This system is not only applicable to Claude but also extends to the Claude API, Claude Code, Claude Cowork, and Claude Tag, as well as all platforms utilizing these AI tools.
Questions & Answers
What is the purpose of the invisible watermark in AI-generated text?
The invisible watermark is intended to provide transparency for AI-generated content, as required by the European Union’s AI Act.
Will the invisible watermark change the quality or readability of the AI’s response?
According to Anthropic, these watermarks won’t affect the meaning, quality, or readability of the AI’s response.
Are the watermarks included only in Claude’s AI-generated text?
No, the watermarks will be applied at the model level, implying that they will be included in any text produced by Claude, the Claude API, Claude Code, Claude Cowork, and Claude Tag, as well as all platforms that use these AI tools.