Claude Opus 5.5 Drastically Reduces AI Writing Markers Like Em-Dashes and Semicolons
According to stylometric testing by Arena, Anthropic's Claude Opus 5.5 reduced em-dash usage by approximately 95% and semicolon usage by 73% compared to Opus 5. The new model dropped em-dash frequency from 15.2 to just 0.8 per 1,000 words, significantly paring back recognizable punctuation patterns. Punctuation quirks like excessive em-dashes and semicolons are key stylistic markers used by readers and automated detectors to spot LLM-generated content. Minimizing these artifacts makes AI output feel more natural, concise, and harder to distinguish from human writing. Opus 5.5 features a shorter average sentence length of 10.03 words compared to 12.14 words in Opus 5, though its overall response length slightly increased from 453 to 481 words. Overall, 10 out of 12 evaluated writing metrics shifted in what Arena considers positive directions.
## BACKGROUND
Large language models (LLMs) often exhibit predictable writing habits, known as 'AI artifacts,' caused by patterns in their training data and alignment tuning. Frontier models frequently over-index on punctuation like em-dashes and semicolons to structure complex thoughts or mimic authoritative prose, making AI-generated text easily recognizable.