Removing Anthropic Fable 5.1 Statistical Watermarks

The Mechanism of Statistical Watermarking

Anthropic's Fable 5.1 implements watermarking by adjusting the probability distribution of token selection during inference. Rather than embedding hidden metadata or code, the model biases its output toward specific, pre-determined token sequences that are statistically unlikely to occur by chance. This technique, based on the research paper "Scalable watermarking for identifying large language model outputs," allows entities possessing the secret key to calculate the probability that a given text was generated by the model. Because this watermark relies on statistical patterns rather than embedded artifacts, it is invisible to the end user and cannot be detected by third-party "AI detector" websites.

Bypassing the Watermark

To remove the watermark, the output must be rewritten by a model that does not employ the same statistical weighting. Light manual editing is insufficient, as the underlying statistical signature often remains. The most effective strategy involves passing the Claude-generated text through a local, open-source model that is not subject to the same regulatory compliance requirements.

• Install Ollama to manage and run local models on your hardware.
• Select an open-source model (e.g., Llama 3 or Mistral) that fits your local compute resources.
• Use a prompt to instruct the local model to rewrite the Claude output while preserving the original tone, voice, and factual content.
• Automate the workflow by creating a custom tool or script that pipes the Claude output directly into the local model for processing.

Context

This watermarking implementation is a compliance measure for the EU AI Act, which mandates that frontier AI labs provide a mechanism to identify machine-generated content. Anthropic claims the watermark does not impact factual accuracy or code performance, as it only influences token selection in cases where multiple words have similar probability weights. The author suggests that while the watermark is technically present, its impact is negligible for short-form content and only becomes statistically significant in long-form text.