Researcher publishes 48-minute video explaining Anthropic’s Claude text watermarking mechanism

Anthropic announced that its Claude models will watermark generated text to enable identification of AI output, and a researcher has published a detailed explainer of the mechanism. Sebastian Raschka, PhD, released a 48-minute video lecture on his Substack, accompanied by a transcript and 50+ slides, to clarify how the watermarking process works, its detection, and potential removal methods.

Raschka’s explainer emphasizes that the watermarking is applied during the token sampling phase of text generation, a process he describes as a minor modification to the standard LLM generation pipeline. The watermark is designed to be invisible to users and detectable only by Anthropic, which he notes is a core feature of the design.

The video begins with a primer on how LLM text generation works, using the example of generating the token “Berlin” in response to the prompt “the capital of Germany is.” Raschka walks through tokenization, embedding, and the sampling process to ground the audience in the mechanics before introducing the watermarking layer.

Raschka contrasts his detailed visual and technical breakdown with Anthropic’s initial announcement, which he describes as conceptual and lacking figures. He argues that understanding the implementation from scratch—even without coding—helps clarify the watermarking’s operation, failure modes, and potential circumvention techniques.

The explainer also addresses common concerns about watermarking, such as its impact on text quality and the trade-offs between detectability and user transparency. Raschka frames the discussion as a way to enable informed public debate about the pros and cons of watermarking in the context of widespread LLM usage.