AI Tells: Opus 5.5 Update (graphite.io)

The latest update of the AI Tells study introduces Claude Opus 5.5, revealing significant advancements in AI's ability to mirror human writing. Notably, Opus 5.5 shows a 19% reduction in word-distribution divergence when compared to Opus 5, indicating a closer alignment with human linguistic patterns. While it utilizes less mannered prose—reducing its score by 37% from the previous version—it maintains a score still 1.6 times higher than human writing. Additionally, both Opus 5.5 and GPT-6 Astra have substantially reduced their use of stylistic features such as em dashes, with Opus 5.5 employing them 99% less frequently, reflecting a shift towards simpler language structures. Despite these improvements, Opus 5.5 still exhibits 2,548 "tells," a mere 4% decrease from its predecessor, showcasing that while it may become more human-like in certain respects, many artificial markers remain. The model is noted for incorporating more language emphasizing helpfulness and tends to use superlatives more frequently than Astra. This highlights the ongoing evolution in AI language models, suggesting while strides have been made towards naturalistic writing, the presence of identifiable traits in AI-generated text persists, raising important considerations for communication and content creation in the AI/ML landscape.