Home / Technology / AI Writing Shifts: Opus 5.5 Mimics Humans Better
AI Writing Shifts: Opus 5.5 Mimics Humans Better
30 Sep
Summary
- Opus 5.5 shows a 19% reduction in word-choice divergence from human text.
- Em-dash usage in Opus 5.5 plummeted by 99% compared to its predecessor.
- Newer AI models are changing writing habits, complicating AI detection.
A recent analysis by Graphite indicates that Anthropic's new flagship large language model, Claude Opus 5.5, exhibits writing that more closely resembles human text compared to its predecessor. The study found a 19% reduction in word-distribution divergence, moving Opus 5.5's score from 0.064 to 0.052 against a human text sample.
This update to Graphite's AI Tells study also highlights significant shifts in specific writing features. Notably, Opus 5.5's use of em-dashes decreased by approximately 99%, contrasting with earlier models and popular associations of em-dashes with AI-generated content.
Furthermore, the model scored 37% lower for "mannered prose," indicating less reliance on metaphorical or flowery language. Despite these improvements, Opus 5.5 retains thousands of recurring words and phrases that appear more frequently than in human writing, with some, like "can help you," appearing eight times more often.
As AI models evolve, they present distinct characteristics. Opus 5.5 shows greater use of superlatives, while OpenAI's GPT-6 Astra uses more qualifiers. This divergence means that evaluations of AI writing quality must go beyond identifying a few familiar expressions, as AI habits continue to change with each release.
