Anthropic has introduced a text watermarking method that can detect AI-generated content by analyzing statistical patterns in word choices. Unlike visible watermarks in images or videos, this approach relies on subtle biases in the selection of words during text generation by Claude, Anthropic’s AI model, according to medianama.com.
The watermarking works by temporarily favoring a specific group of words when Claude predicts the next word in a sentence. For example, when completing a phrase like “The weather was ___,” Claude might prefer less probable words within a set, creating a detectable statistical pattern. This pattern can be identified by Anthropic’s system when it analyzes the text, even though there are no visible marks or hidden characters embedded in the output.
This technique addresses the difficulty of detecting AI-generated text, which can be copied and pasted without alteration, unlike images or videos where watermarks can be embedded or removed. However, the watermarking has limitations: it does not identify the user or organization, cannot confirm that Claude wrote the entire text, and performs poorly on short, factual, or code-heavy content. It may also detect text heavily edited or translated by Claude.
Anthropic’s text watermarking represents a novel approach to AI content detection, focusing on statistical biases in word choice rather than explicit markers. The method was detailed in a report published on medianama.com on August 18, 2026.