Anthropic has introduced a watermarking system for its Claude AI models to identify text generated by the system, a move aimed at complying with the EU AI Act. The technology uses statistical patterns in word choice to tag AI-generated content without visible or hidden markers. According to the company, this method does not affect the content, creativity, or readability of the output. The approach is based on Google’s SynthID-Text technique, which adjusts the randomness source for word selection during text generation to create a detectable pattern. The system is designed to be imperceptible to users, yet it has drawn criticism from some experts and legal professionals. Source: thedecoder
Blogger John Gruber argues that the watermarking process compromises text quality by influencing word selection based on a secret key rather than semantic meaning. He claims that the system sometimes increases the probability of a less optimal word and reduces the likelihood of a better one, which he says makes the difference in output 'imperceptible' only in the company’s view. Gruber, who co-created Markdown, also questions the validity of using thumbs-up/down rates from the SynthID study as a measure of text quality. He suggests that the method may inadvertently affect the perceived performance of models like Gemini compared to Claude and ChatGPT. Source: thedecoder
Legal experts at Artificial Lawyer note that while the watermarking system is generally harmless, it could complicate situations where AI use is explicitly prohibited. The watermarks travel with the text, meaning AI-generated sections in contracts may carry detectable marks even if the rest of the document is human-written. This could affect fee negotiations or the outcome of legal proceedings if AI involvement is challenged. Anthropic claims that the watermarking is less dense in fact-heavy texts, but no empirical studies have been conducted to support this assertion. Source: thedecoder
