Anthropic Engineer Explains Why Claude's Writing Quality Degraded in Recent Models
An Anthropic fine-tuning engineer explained that Claude's writing quality declined after Opus 4.6 because post-training reinforcement learning prioritized math and coding capabilities. Furthermore, training models on technical explanations designed for other AI models led to an overly dense, unnatural writing style known as 'Claudeish.' This reveals how optimizing large language models for complex technical tasks and AI-to-AI communication can unintentionally degrade human readability. It highlights the delicate challenge AI labs face in tuning reinforcement learning rewards to advance reasoning capabilities without sacrificing clear, natural prose. Because LLMs possess vast working memory and process nuanced information, AI-targeted training led Claude to produce info-dense responses that feel overwhelming to human readers. Anthropic claims to have achieved a better balance in Opus 5.5 by introducing active counter-rewards for concise, human-friendly explanations.
## BACKGROUND
Large language models undergo post-training using Reinforcement Learning from Human Feedback (RLHF) to refine their behaviors and output formatting. When training incentives focus heavily on verifiable domains like code and mathematics, models often experience style drift, adopting complex phrasing patterns optimized for machine evaluation rather than everyday human reading.