Anthropic’s Claude AI Watermark Likely MirrorMark, Analysis Suggests

Image credit: Search Engine Journal

Anthropic’s recently announced text watermarking technology for its Claude AI likely utilizes MirrorMark, a system developed at George Mason University, according to a recent analysis of publicly available information.

The strong correlation stems from the six key qualities Anthropic described for its new watermarking, which closely match the characteristics of MirrorMark, a technology designed to embed imperceptible signals within generated text.

Anthropic, a U.S.-based artificial intelligence company, stated its watermarking is directly embedded into generated text, is imperceptible to users, preserves text quality and is generated at the model level. The company also said the watermark remains detectable even after users or third parties edit the text.

A July 23 update to Anthropic’s transparency page indicated the technology was developed in collaboration with academia, further supporting the link to a university-developed solution.

MirrorMark, developed by researchers at George Mason University, specifically aims to survive editing and paraphrasing, a critical feature highlighted by Anthropic.

Another potential candidate, MCmark, is an unbiased watermarking method that embeds a hidden statistical signal during token generation. However, MCmark has shown lower true-positive detection rates after paraphrasing compared to MirrorMark, according to research.

InvisibleID, an entity associated with George Mason University, commercializes MirrorMark, suggesting its availability for licensing to companies like Anthropic.

The integration of such watermarking technologies is seen as a step toward addressing concerns about the provenance of AI-generated content and combating the spread of misinformation.

Source: Search Engine Journal