Anthropic engineer Jackson Kernion addressed the decline in the writing quality of Claude models, noting that recent optimizations prioritized mathematical and coding capabilities over natural language generation. This shift in focus has led to the emergence of a specific style often referred to as Claudeish.
The issue stems from training models on vast amounts of technical explanations designed for other AI systems. Consequently, these models have adapted to the psychological patterns of large language models, prioritizing machine-readable density over human-readable clarity, which makes the output difficult for people to process.
Kernion compared this phenomenon to a group that develops an internal communication style unintelligible to outsiders. Because LLMs possess superior working memory and nuance detection, they naturally gravitate toward formats that are highly efficient for AI comprehension but appear cluttered to human readers.
The root cause is identified as the reinforcement learning reward mechanism. Current training protocols often favor AI understanding, which necessitates active intervention to reward simplicity and human-centric explanations. Anthropic is now working to recalibrate these rewards to improve readability.
Regarding the latest developments, Kernion indicated that the Opus 5.5 model achieves a better balance in writing performance compared to its predecessors. While he expressed satisfaction with this progress, he emphasized that the challenge remains complex and the company continues to refine its approach.
