Claude Reportedly Implements Steganographic Watermarking, Raising Concerns Over False Positives

A Reddit discussion highlights reports that Anthropic's Claude has officially begun steganographically watermarking its AI-generated text. Users have already flagged instances of false positives, where human-written text or unrelated outputs are incorrectly flagged or affected by these watermarking patterns. This development intensifies the debate between open-source and closed-source AI models, as users worry about hidden tracking and lack of control over output text. It also highlights the technical challenges of reliably identifying AI-generated content without disrupting normal user workflows or falsely accusing writers. Steganographic text watermarking typically works by subtly altering token probability distributions during generation, making the text identifiable to detectors but invisible to the casual reader. However, this method can be bypassed by paraphrasing and can lead to false positives if natural human writing patterns accidentally mimic the watermark's statistical signature.